Page Summary

outlined_flag

• This glossary defines a wide range of machine learning terms, including those specific to TensorFlow and large language models.

• It provides clear explanations, examples, and applications of each term for a comprehensive understanding.

• The content covers various aspects of machine learning, such as model types, training processes, fairness, and evaluation metrics.

• It serves as a valuable resource for anyone seeking to grasp the terminology and fundamental concepts in this field.

• The glossary also addresses advanced topics like deep learning, reinforcement learning, and natural language processing.

This glossary defines artificial intelligence terms.

See ourFAQ.

A

ablation

A technique for evaluating the importance of afeatureor component by temporarilyremoving it from amodel. You thenretrain the model without that feature or component, and if the retrained modelperforms significantly worse, then the removed feature or component waslikely important.

For example, suppose you train aclassification modelon 10 features and achieve 88%precision on thetest set. To check theimportanceof the first feature, you can retrain the model using only the nine otherfeatures. If the retrained model performs significantly worse (for instance,55% precision), then the removed feature was probably important. Conversely,if the retrained model performs equally well, then that feature was probablynot that important.

Ablation can also help determine the importance of:

• Larger components, such as an entire subsystem of a larger ML system
• Processes or techniques, such as a data preprocessing step

In both cases, you would observe how the system's performance changes (ordoesn't change) after you've removed the component.

A/B testing

A statistical way of comparing two (or more) techniques—theAand theB. Typically, theA is an existing technique, and theB is a new technique.A/B testing not only determines which technique performs betterbut also whether the difference is statistically significant.

A/B testing usually compares a singlemetric on two techniques;for example, how does modelaccuracy compare for twotechniques? However, A/B testing can also compare any finite number ofmetrics.

accelerator chip

A category of specialized hardware components designed to perform keycomputations needed for deep learning algorithms.

Accelerator chips (or justaccelerators, for short) can significantlyincrease the speed and efficiency of training and inference taskscompared to a general-purpose CPU. They are ideal for trainingneural networks and similar computationally intensive tasks.

Examples of accelerator chips include:

• Google's Tensor Processing Units (TPUs) with dedicated hardwarefor deep learning.
• NVIDIA's GPUs which, though initially designed for graphics processing,are designed to enable parallel processing, which can significantlyincrease processing speed.

accuracy

The number of correct classificationpredictions dividedby the total number of predictions. That is:

For example, a model that made 40 correct predictions and 10 incorrectpredictions would have an accuracy of:

Binary classification provides specific namesfor the different categories ofcorrect predictions andincorrect predictions. So, the accuracy formula for binary classificationis as follows:

where:

• TP is the number oftrue positives (correct predictions).
• TN is the number oftrue negatives (correct predictions).
• FP is the number offalse positives (incorrect predictions).
• FN is the number offalse negatives (incorrect predictions).

Compare and contrast accuracy withprecision andrecall.

accuracy = (TP + TN) / (TP + TN + FP + FN)accuracy = (0 + 36499) / (0 + 36499 + 0 + 25) = 0.9993 = 99.93%

SeeClassification: Accuracy, recall, precision and relatedmetricsin Machine Learning Crash Course for more information.

action

Inreinforcement learning,the mechanism by which theagenttransitions betweenstates of theenvironment. The agent chooses the action by using apolicy.

activation function

A function that enablesneural networks to learnnonlinear (complex) relationships between featuresand the label.

Popular activation functions include:

• ReLU
• Sigmoid

The plots of activation functions are never single straight lines.For example, the plot of the ReLU activation function consists oftwo straight lines:

A plot of the sigmoid activation function looks as follows:

weighted sum = (2)(-1.3) + (-1)(0.6) + (3)(0.4) = -2.0

SeeNeural networks: Activationfunctionsin Machine Learning Crash Course for more information.

active learning

Atraining approach in which thealgorithmchooses some of the data it learns from. Active learningis particularly valuable whenlabeled examplesare scarce or expensive to obtain. Instead of blindly seeking a diverserange of labeled examples, an active learning algorithm selectively seeksthe particular range of examples it needs for learning.

AdaGrad

A sophisticated gradient descent algorithm that rescales thegradients of eachparameter, effectively giving each parameteran independentlearning rate. For a full explanation, seeAdaptive Subgradient Methods for Online Learning and StochasticOptimization.

adaptation

Synonym for tuning orfine-tuning.

agent

Software that can reason about multimodal user inputs in order to plan andexecute actions on behalf of the user.

Inreinforcement learning,an agent is the entity that uses apolicy to maximize the expectedreturn gained fromtransitioning betweenstates of theenvironment.

agentic

The adjective form ofagent. Agentic refers to the qualitiesthat agents possess (such as autonomy).

agentic workflow

A dynamic process in which anagent autonomously plans andexecutes actions to achieve a goal. The process may involve reasoning,invoking external tools, and self-correcting its plan.

agglomerative clustering

Seehierarchical clustering.

AI slop

Output from agenerative AI system that favorsquantity over quality. For example, a web page with AI slop is filledwith cheaply produced, AI-generated, low-quality content.

anomaly detection

The process of identifyingoutliers. For example, if the meanfor a certainfeature is 100 with a standard deviation of 10,then anomaly detection should flag a value of 200 as suspicious.

AR

Abbreviation foraugmented reality.

area under the PR curve

SeePR AUC (Area under the PR Curve).

area under the ROC curve

SeeAUC (Area under the ROC curve).

artificial general intelligence

A non-human mechanism that demonstrates abroad range of problem solving,creativity, and adaptability. For example, a program demonstrating artificialgeneral intelligence could translate text, compose symphonies,and excel atgames that have not yet been invented.

artificial intelligence

A non-human program ormodel that can solve sophisticated tasks.For example, a program or model that translates text or a program or model thatidentifies diseases from radiologic images both exhibit artificial intelligence.

Formally,machine learning is a sub-field of artificialintelligence. However, in recent years, some organizations have begun using thetermsartificial intelligence andmachine learning interchangeably.

attention

A mechanism used in aneural network that indicatesthe importance of a particular word or part of a word. Attention compressesthe amount of information a model needs to predict the next token/word.A typical attention mechanism might consist of aweighted sum over a set of inputs, where theweight for each input is computed by another part of theneural network.

Refer also toself-attention andmulti-head self-attention, which are thebuilding blocks ofTransformers.

SeeLLMs: What's a large languagemodel?in Machine Learning Crash Course for more information about self-attention.

attribute

Synonym forfeature.

In machine learning fairness, attributes often refer tocharacteristics pertaining to individuals.

attribute sampling

A tactic for training adecision forest in which eachdecision tree considers only a random subset of possiblefeatures when learning thecondition.Generally, a different subset of features is sampled for eachnode. In contrast, when training a decision treewithout attribute sampling, all possible features are considered for each node.

AUC (Area under the ROC curve)

A number between 0.0 and 1.0 representing abinary classification model'sability to separatepositive classes fromnegative classes.The closer the AUC is to 1.0, the better the model's ability to separateclasses from each other.

For example, the following illustration shows aclassification model that separates positiveclasses (green ovals) from negative classes (purple rectangles) perfectly.This unrealistically perfect model has an AUC of 1.0:

Conversely, the following illustration shows the results for aclassification model that generated randomresults. This model has an AUC of 0.5:

Yes, the preceding model has an AUC of 0.5, not 0.0.

Most models are somewhere between the two extremes. For instance, thefollowing model separates positives from negatives somewhat, and thereforehas an AUC somewhere between 0.5 and 1.0:

AUC ignores any value you set forclassification threshold. Instead, AUCconsidersall possible classification thresholds.

SeeClassification: ROC andAUCin Machine Learning Crash Course for more information.

augmented reality

A technology that superimposes a computer-generated image on a user's view ofthe real world, thus providing a composite view.

autoencoder

A system that learns to extract the most important information from theinput. Autoencoders are a combination of anencoder anddecoder. Autoencoders rely on the following two-step process:

1. The encoder maps the input to a (typically) lossy lower-dimensional(intermediate) format.
2. The decoder builds a lossy version of the original input by mappingthe lower-dimensional format to the original higher-dimensionalinput format.

Autoencoders are trained end-to-end by having the decoder attempt toreconstruct the original input from the encoder's intermediate formatas closely as possible. Because the intermediate format is smaller(lower-dimensional) than the original format, the autoencoder is forcedto learn what information in the input is essential, and the output won'tbe perfectly identical to the input.

For example:

• If the input data is a graphic, the non-exact copy would be similar tothe original graphic, but somewhat modified. Perhaps thenon-exact copy removes noise from the original graphic or fills insome missing pixels.
• If the input data is text, an autoencoder would generate new text thatmimics (but is not identical to) the original text.

See alsovariational autoencoders.

automatic evaluation

Using software to judge the quality of a model's output.

When model output is relatively straightforward, a script or program cancompare the model's output to agolden response.This type of automatic evaluation is sometimes calledprogrammatic evaluation. Metrics such asROUGE orBLEU are often useful for programmatic evaluation.

When model output is complex or hasno one right answer, a separate ML program called anautorater sometimes performs the automaticevaluation.

Contrast withhuman evaluation.

automation bias

When a human decision maker favors recommendations made by an automateddecision-making system over information made without automation, evenwhen the automated decision-making system makes errors.

SeeFairness: Types ofbiasin Machine Learning Crash Course for more information.

AutoML

Any automated process for buildingmachine learningmodels. AutoML can automatically do tasks such as the following:

• Search for the most appropriate model.
• Tunehyperparameters.
• Prepare data (including performingfeature engineering).
• Deploy the resulting model.

AutoML is useful for data scientists because it can save them time andeffort in developing machine learning pipelines and improve predictionaccuracy. It is also useful to non-experts, by making complicatedmachine learning tasks more accessible to them.

SeeAutomated MachineLearning (AutoML)in Machine Learning Crash Course for more information.

autorater evaluation

Prebuilt autoraters are available, but the best autoraters arefine-tuned specifically to the task you are evaluating.

auto-regressive model

Amodel that infers a prediction based on its own previouspredictions. For example, auto-regressive language models predict the nexttoken based on the previously predicted tokens.AllTransformer-basedlarge language models are auto-regressive.

In contrast,GAN-based image models are usually not auto-regressivesince they generate an image in a single forward-pass and not iteratively insteps. However, certain image generation modelsare auto-regressive becausethey generate an image in steps.

auxiliary loss

Aloss function—used in conjunction with aneural networkmodel's mainloss function—that helps acceleratetraining during theearly iterations when weights are randomly initialized.

Auxiliary loss functions push effectivegradientsto the earlierlayers. This facilitatesconvergence duringtrainingby combating thevanishing gradient problem.

average precision at k

A metric for summarizing a model's performance on a single prompt thatgenerates ranked results, such as a numbered list of book recommendations.Average precision atk is, well, the average of theprecision at k values for eachrelevant result.The formula for average precision atk is therefore:

\[{\text{average precision at k}} = \frac{1}{n} \sum_{i=1}^n {\text{precision at k for each relevant item} } \]

where:

• \(n\) is the number of relevant items in the list.

Contrast withrecall at k.

List the 6 funniest movies of all time in order.

axis-aligned condition

In adecision tree, aconditionthat involves only a singlefeature. For example, ifareais a feature, then the following is an axis-aligned condition:

area > 200

Contrast withoblique condition.

B

backpropagation

The algorithm that implementsgradient descent inneural networks.

Training a neural network involves manyiterationsof the following two-pass cycle:

1. During theforward pass, the system processes abatch ofexamples to yield prediction(s). The system compares eachprediction to eachlabel value. The difference betweenthe prediction and the label value is theloss for that example.The system aggregates the losses for all the examples to compute the totalloss for the current batch.
2. During thebackward pass (backpropagation), the system reduces loss byadjusting the weights of all theneurons in all thehidden layer(s).

Neural networks often contain many neurons across many hidden layers.Each of those neurons contribute to the overall loss in different ways.Backpropagation determines whether to increase or decrease the weightsapplied to particular neurons.

Thelearning rate is a multiplier that controls thedegree to which each backward pass increases or decreases each weight.A large learning rate will increase or decrease each weight more than asmall learning rate.

In calculus terms, backpropagation implements thechain rule.from calculus. That is, backpropagation calculates thepartial derivative of the error withrespect to each parameter.

Years ago, ML practitioners had to write code to implement backpropagation.Modern ML APIs like Keras now implement backpropagation for you. Phew!

SeeNeural networksin Machine Learning Crash Course for more information.

bagging

A method totrain anensemble where eachconstituentmodel trains on a random subset of trainingexamplessampled with replacement.For example, arandom forest is a collection ofdecision trees trained with bagging.

The termbagging is short forbootstrapaggregating.

SeeRandom forestsin the Decision Forests course for more information.

bag of words

A representation of the words in a phrase or passage,irrespective of order. For example, bag of words represents thefollowing three phrases identically:

• the dog jumps
• jumps the dog
• dog jumps the

Each word is mapped to an index in asparse vector, wherethe vector has an index for every word in the vocabulary. For example,the phrasethe dog jumps is mapped into a feature vector with non-zerovalues at the three indexes corresponding to the wordsthe,dog, andjumps. The non-zero value can be any of the following:

• A 1 to indicate the presence of a word.
• A count of the number of times a word appears in the bag. For example,if the phrase werethe maroon dog is a dog with maroon fur, then bothmaroon anddog would be represented as 2, while the other words wouldbe represented as 1.
• Some other value, such as the logarithm of the count of the number oftimes a word appears in the bag.

baseline

Amodel used as a reference point for comparing how well anothermodel (typically, a more complex one) is performing. For example, alogistic regression model might serve as agood baseline for adeep model.

For a particular problem, the baseline helps model developers quantifythe minimal expected performance that a new model must achieve for the newmodel to be useful.

base model

Apre-trained model that can serve as the startingpoint forfine-tuning to address specific tasks orapplications.

See alsopre-trained modelandfoundation model.

batch

The set ofexamples used in one trainingiteration.Thebatch size determines the number of examples in abatch.

Seeepoch for an explanation of how a batch relates toan epoch.

SeeLinear regression:Hyperparametersin Machine Learning Crash Course for more information.

batch inference

The process ofinferring predictions on multipleunlabeled examples divided into smallersubsets ("batches").

Batch inference can take advantage of the parallelization features ofaccelerator chips. That is, multiple acceleratorscan simultaneously infer predictions on different batches of unlabeledexamples, dramatically increasing the number of inferences per second.

SeeProduction ML systems: Static versus dynamicinferencein Machine Learning Crash Course for more information.

batch normalization

Normalizing the input or output of theactivation functions in ahidden layer. Batch normalization canprovide the following benefits:

• Makeneural networks more stable by protectingagainstoutlier weights.
• Enable higherlearning rates, which canspeed training.
• Reduceoverfitting.

batch size

The number ofexamples in abatch.For instance, if the batch size is 100, then the model processes100 examples periteration.

The following are popular batch size strategies:

• Stochastic Gradient Descent (SGD), in which the batch size is 1.
• Full batch, in which the batch size is the number of examples in the entiretraining set. For instance, if the training setcontains a million examples, then the batch size would be a millionexamples. Full batch is usually an inefficient strategy.
• mini-batch in which the batch size is usually between10 and 1000. Mini-batch is usually the most efficient strategy.

See the following for more information:

• Production ML systems: Static versus dynamicinferencein Machine Learning Crash Course.
• Deep Learning TuningPlaybook.

Bayesian neural network

A probabilisticneural network that accounts foruncertainty inweights and outputs. A standard neural networkregression model typicallypredicts a scalar value;for example, a standard model predicts a house priceof 853,000. In contrast, a Bayesian neural network predicts a distribution ofvalues; for example, a Bayesian model predicts a house price of 853,000 witha standard deviation of 67,200.

A Bayesian neural network relies onBayes' Theoremto calculate uncertainties in weights and predictions. A Bayesian neuralnetwork can be useful when it is important to quantify uncertainty, such as inmodels related to pharmaceuticals. Bayesian neural networks can also helppreventoverfitting.

Bayesian optimization

Aprobabilistic regression modeltechnique for optimizing computationally expensiveobjective functions by instead optimizing a surrogatethat quantifies the uncertainty using a Bayesian learning technique. SinceBayesian optimization is itself very expensive, it is usually used to optimizeexpensive-to-evaluate tasks that have a small number of parameters, such asselectinghyperparameters.

Bellman equation

In reinforcement learning, the following identity satisfied by the optimalQ-function:

\[Q(s, a) = r(s, a) + \gamma \mathbb{E}_{s'|s,a} \max_{a'} Q(s', a')\]

Reinforcement learning algorithms apply thisidentity to createQ-learning using the following updaterule:

\[Q(s,a) \gets Q(s,a) + \alpha \left[r(s,a) + \gamma \displaystyle\max_{\substack{a_1}} Q(s',a') - Q(s,a) \right]\]

Beyond reinforcement learning, the Bellman equation has applications todynamic programming. See theWikipedia entry for Bellman equation.

BERT (Bidirectional EncoderRepresentations from Transformers)

A model architecture for textrepresentation. A trainedBERT model can act as part of a larger model for text classification orother ML tasks.

BERT has the following characteristics:

• Uses theTransformer architecture, and therefore reliesonself-attention.
• Uses theencoder part of the Transformer. The encoder's jobis to produce good text representations, rather than to perform a specifictask like classification.
• Isbidirectional.
• Usesmasking forunsupervised training.

BERT's variants include:

• ALBERT,which is an acronym forALightBERT.
• LaBSE.

SeeOpen Sourcing BERT: State-of-the-Art Pre-training for Natural LanguageProcessingfor an overview of BERT.

bias (ethics/fairness)

1. Stereotyping, prejudice or favoritism towards some things, people,or groups over others. These biases can affect collection andinterpretation of data, the design of a system, and how users interactwith a system. Forms of this type of bias include:

• automation bias
• confirmation bias
• experimenter's bias
• group attribution bias
• implicit bias
• in-group bias
• out-group homogeneity bias

2. Systematic error introduced by a sampling or reporting procedure.Forms of this type of bias include:

• coverage bias
• non-response bias
• participation bias
• reporting bias
• sampling bias
• selection bias

Not to be confused with thebias term in machine learning modelsorprediction bias.

SeeFairness: Types ofbias inMachine Learning Crash Course for more information.

bias (math) or bias term

An intercept or offset from an origin. Bias is a parameter inmachine learning models, which is symbolized by either of thefollowing:

• b
• w₀

For example, bias is theb in the following formula:

In a simple two-dimensional line, bias just means "y-intercept."For example, the bias of the line in the following illustration is 2.

Bias exists because not all models start from the origin (0,0). For example,suppose an amusement park costs 2 Euros to enter and an additional0.5 Euro for every hour a customer stays. Therefore, a model mapping thetotal cost has a bias of 2 because the lowest cost is 2 Euros.

Bias is not to be confused withbias in ethics and fairnessorprediction bias.

SeeLinear Regressionin Machine Learning Crash Course for more information.

bidirectional

A term used to describe a system that evaluates the text that bothprecedesandfollows a target section of text. In contrast, aunidirectional system onlyevaluates the text thatprecedes a target section of text.

For example, consider amasked language model thatmust determine probabilities for the word or words representing the underline inthe following question:

What is the _____ with you?

A unidirectional language model would have to base its probabilities onlyon the context provided by the words "What", "is", and "the". In contrast,a bidirectional language model could also gain context from "with" and "you",which might help the model generate better predictions.

bidirectional language model

Alanguage model that determines the probability that agiven token is present at a given location in an excerpt of text based onthepreceding andfollowing text.

bigram

AnN-gram in which N=2.

binary classification

A type ofclassification task thatpredicts one of two mutually exclusive classes:

• thepositive class
• thenegative class

For example, the following two machine learning models each performbinary classification:

• A model that determines whether email messages arespam (the positive class) ornot spam (the negative class).
• A model that evaluates medical symptoms to determine whether a personhas a particular disease (the positive class) or doesn't have thatdisease (the negative class).

Contrast withmulti-class classification.

See alsologistic regression andclassification threshold.

SeeClassificationin Machine Learning Crash Course for more information.

binary condition

In adecision tree, aconditionthat has only two possible outcomes, typicallyyes orno.For example, the following is a binary condition:

temperature >= 100

Contrast withnon-binary condition.

SeeTypes of conditionsin the Decision Forests course for more information.

binning

Synonym forbucketing.

black box model

Amodel whose "reasoning" is impossible or difficult for humansto understand. That is, although humans can see howpromptsaffectresponses, humans can't determine exactly how a blackbox model determines the response. In other words, a black box model is lackinginterpretability.

Mostdeep models andlarge language models are black boxes.

BLEU (Bilingual Evaluation Understudy)

A metric between 0.0 and 1.0 for evaluatingmachine translations, for example, fromSpanish to Japanese.

To calculate a score, BLEU typically compares an ML model's translation(generated text) to a human expert's translation(reference text).The degree to whichN-grams in the generated text andreference text match determines the BLEU score.

The original paper on this metric isBLEU: a Method for Automatic Evaluation of Machine Translation.

See alsoBLEURT.

BLEURT (Bilingual Evaluation Understudy from Transformers)

A metric for evaluatingmachine translationsfrom one language to another, particularly to and from English.

For translations to and from English, BLEURT aligns more closely to humanratings thanBLEU. Unlike BLEU, BLEURT emphasizes semantic(meaning) similarities and can accommodate paraphrasing.

BLEURT relies on apre-trained large language model(BERT to be exact) that is thenfine-tunedon text from human translators.

The original paper on this metric isBLEURT: Learning Robust Metrics for Text Generation.

Boolean Questions (BoolQ)

A dataset for evaluating an LLM's proficiency in answering yes-or-no questions.Each of the challenges in the dataset has three components:

• A query
• A passage implying the answer to the query.
• The correct answer, which is eitheryes orno.

For example:

• Query: Are there any nuclear power plants in Michigan?
• Passage: ...three nuclear power plants supply Michigan withabout 30% of its electricity.
• Correct answer: Yes

Researchers gathered the questions from anonymized, aggregated Google Searchqueries and then used Wikipedia pages to ground the information.

For more information, seeBoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions.

BoolQ is a component of theSuperGLUE ensemble.

BoolQ

Abbreviation forBoolean Questions.

boosting

A machine learning technique that iteratively combines a set of simple and notvery accurateclassification models (referred toas "weak classifiers") into a classification model with high accuracy(a "strong classifier") byupweighting the examples thatthe model is currently misclassifying.

SeeGradient Boosted DecisionTrees?in the Decision Forests course for more information.

bounding box

In an image, the (x,y) coordinates of a rectangle around an area ofinterest, such as the dog in the image below.

broadcasting

Expanding the shape of an operand in a matrix math operation todimensions compatible for that operation. For example,linear algebra requires that the two operands in a matrix addition operationmust have the same dimensions. Consequently, you can't add a matrix of shape(m, n) to a vector of length n. Broadcasting enables this operation byvirtually expanding the vector of length n to a matrix of shape (m, n) byreplicating the same values down each column.

A = [[7, 10, 4],     [13, 5, 9]]B = [2]

 [[2, 2, 2],  [2, 2, 2]]

[[7, 10, 4],  +  [[2, 2, 2],  =  [[ 9, 12, 6], [13, 5, 9]]      [2, 2, 2]]      [15, 7, 11]]

See the following description ofbroadcasting in NumPy for more details.

bucketing

Converting a singlefeature into multiple binary featurescalledbuckets orbins,typically based on a value range. The chopped feature is typically acontinuous feature.

For example, instead of representing temperature as a singlecontinuous floating-point feature, you could chop ranges of temperaturesinto discrete buckets, such as:

• <= 10 degrees Celsius would be the "cold" bucket.
• 11 - 24 degrees Celsius would be the "temperate" bucket.
• >= 25 degrees Celsius would be the "warm" bucket.

The model will treat every value in the same bucket identically. Forexample, the values13 and22 are both in the temperate bucket, so themodel treats the two values identically.

SeeNumerical data:Binningin Machine Learning Crash Course for more information.

C

calibration layer

A post-prediction adjustment, typically to account forprediction bias. The adjusted predictions andprobabilities should match the distribution of an observed set of labels.

candidate generation

The initial set of recommendations chosen by arecommendation system. For example, consider abookstore that offers 100,000 titles. The candidate generation phase createsa much smaller list of suitable books for a particular user, say 500. But even500 books is way too many to recommend to a user. Subsequent, more expensive,phases of a recommendation system (such asscoring andre-ranking) reduce those 500 to a much smaller,more useful set of recommendations.

SeeCandidate generationoverviewin the Recommendation Systems course for more information.

candidate sampling

A training-time optimization that calculates a probability for all thepositive labels, using, for example,softmax, but only for a randomsample of negative labels. For instance, given an example labeledbeagle anddog, candidate sampling computes the predicted probabilitiesand corresponding loss terms for:

• beagle
• dog
• a random subset of the remaining negative classes (for example,cat,lollipop,fence).

The idea is that thenegative classes can learn from less frequentnegative reinforcement as long aspositive classes always get proper positivereinforcement, and this is indeed observed empirically.

Candidate sampling is more computationally efficient than training algorithmsthat compute predictions forall negative classes, particularly when thenumber of negative classes is very large.

categorical data

Features having a specific set of possible values. For example,consider a categorical feature namedtraffic-light-state, which can onlyhave one of the following three possible values:

• red
• yellow
• green

By representingtraffic-light-state as a categorical feature,a model can learn thediffering impacts ofred,green, andyellow on driver behavior.

Categorical features are sometimes calleddiscrete features.

Contrast withnumerical data.

SeeWorking with categoricaldatain Machine Learning Crash Course for more information.

causal language model

Synonym forunidirectional language model.

Seebidirectional language model tocontrast different directional approaches in language modeling.

CB

Abbreviation forCommitmentBank.

centroid

The center of a cluster as determined by ak-means ork-median algorithm. For example, if k is 3,then the k-means or k-median algorithm finds 3 centroids.

SeeClustering algorithmsin the Clustering course for more information.

centroid-based clustering

A category ofclustering algorithms that organizes datainto nonhierarchical clusters.k-means is the most widelyused centroid-based clustering algorithm.

Contrast withhierarchical clusteringalgorithms.

SeeClustering algorithmsin the Clustering course for more information.

chain-of-thought prompting

Aprompt engineering technique that encouragesalarge language model (LLM) to explain itsreasoning, step by step. For example, consider the following prompt, payingparticular attention to the second sentence:

How many g forces would a driver experience in a car that goes from 0 to 60miles per hour in 7 seconds? In the answer, show all relevant calculations.

The LLM'sresponse would likely:

• Show a sequence of physics formulas, plugging in the values 0, 60, and 7in appropriate places.
• Explain why it chose those formulas and what the various variables mean.

Chain-of-thought prompting forces the LLM to perform all the calculations,which might lead to a more correct answer. In addition, chain-of-thoughtprompting enables the user to examine the LLM's steps to determine whetheror not the answer makes sense.

Character N-gram F-score (ChrF)

A metric to evaluatemachine translation models.Character N-gram F-score determines the degree to whichN-grams inreference text overlap theN-grams in an ML model'sgenerated text.

Character N-gram F-score is similar to metrics in theROUGEandBLEU families, except that:

• Character N-gram F-score operates oncharacter N-grams.
• ROUGE and BLEU operate onword N-grams ortokens.

chat

The contents of a back-and-forth dialogue with an ML system, typically alarge language model.The previous interaction in a chat(what you typed and how the large language model responded) becomes thecontext for subsequent parts of the chat.

Achatbot is an application of a large language model.

checkpoint

Data that captures the state of a model'sparameters eitherduring training or after training is completed. For example, during training,you can:

1. Stop training, perhaps intentionally or perhaps as the result ofcertain errors.
2. Capture the checkpoint.
3. Later, reload the checkpoint, possibly on different hardware.
4. Restart training.

Choice of Plausible Alternatives (COPA)

A dataset for evaluating how well an LLM can identify the better of twoalternative answers to a premise. Each of the challenges in the datasetconsists of three components:

• A premise, which is typically a statement followed by a question
• Two possible answers to the question posed in the premise, one of whichis correct and the other incorrect
• The correct answer

For example:

• Premise: The man broke his toe. What was the CAUSE of this?
• Possible answers:
  1. He got a hole in his sock.
  2. He dropped a hammer on his foot.
• Correct answer: 2

COPA is a component of theSuperGLUE ensemble.

class

A category that alabel can belong to.For example:

• In abinary classification model that detectsspam, the two classes might bespam andnot spam.
• In amulti-class classification modelthat identifies dog breeds, the classes might bepoodle,beagle,pug,and so on.

Aclassification model predicts a class.In contrast, aregression model predicts a numberrather than a class.

SeeClassificationin Machine Learning Crash Course for more information.

class-balanced dataset

Adataset containingcategoricallabels in which the number of instances of each category isapproximately equal. For example, consider a botanical dataset whose binarylabel can be eithernative plant ornonnative plant:

• A dataset with 515native plants and 485nonnative plants is aclass-balanced dataset.
• A dataset with 875native plants and 125nonnative plants is aclass-imbalanced dataset.

A formal dividing line between class-balanced datasets andclass-imbalanced datasets doesn't exist.The distinction only becomes important when a model trained on a highlyclass-imbalanced dataset can't converge. SeeDatasets: imbalanced datasetsin Machine Learning Crash Course for details.

classification model

Amodel whose prediction is aclass.For example, the following are all classification models:

• A model that predicts an input sentence's language (French? Spanish?Italian?).
• A model that predicts tree species (Maple? Oak? Baobab?).
• A model that predicts the positive or negative class for a particularmedical condition.

In contrast,regression models predict numbersrather than classes.

Two common types of classification models are:

• binary classification
• multi-class classification

classification threshold

In abinary classification, anumber between 0 and 1 that converts the raw output of alogistic regression modelinto a prediction of either thepositive classor thenegative class.Note that the classification threshold is a value that a human chooses,not a value chosen by model training.

A logistic regression model outputs a raw value between 0 and 1. Then:

• If this raw value isgreater than the classification threshold, thenthe positive class is predicted.
• If this raw value isless than the classification threshold, thenthe negative class is predicted.

For example, suppose the classification threshold is 0.8. If the raw valueis 0.9, then the model predicts the positive class. If the raw value is0.7, then the model predicts the negative class.

The choice of classification threshold strongly influences the number offalse positives andfalse negatives.

SeeThresholds and the confusionmatrixin Machine Learning Crash Course for more information.

classifier

A casual term for aclassification model.

class-imbalanced dataset

Adataset for aclassificationin which the total number oflabels of eachclassdiffers significantly. For example, consider abinary classification dataset whose two labelsare divided as follows:

• 1,000,000 negative labels
• 10 positive labels

The ratio of negative to positive labels is 100,000 to 1, so thisis a class-imbalanced dataset.

In contrast, the following dataset isclass-balanced because the ratio of negativelabels to positive labels is relatively close to 1:

• 517 negative labels
• 483 positive labels

Multi-class datasets can also be class-imbalanced. For example, the followingmulti-class classification dataset is also class-imbalanced because one labelhas far more examples than the other two:

• 1,000,000 labels with class "green"
• 200 labels with class "purple"
• 350 labels with class "orange"

Training class-imbalanced datasets can present special challenges. SeeImbalanced datasetsin Machine Learning Crash Course for details.

See alsoentropy,majority class,andminority class.

clipping

A technique for handlingoutliers by doingeither or both of the following:

• Reducingfeature values that are greater than a maximumthreshold down to that maximum threshold.
• Increasing feature values that are less than a minimum threshold up to thatminimum threshold.

For example, suppose that <0.5% of values for a particular feature falloutside the range 40–60. In this case, you could do the following:

• Clip all values over 60 (the maximum threshold) to be exactly 60.
• Clip all values under 40 (the minimum threshold) to be exactly 40.

Outliers can damage models, sometimes causingweightsto overflow during training. Some outliers can also dramatically spoilmetrics likeaccuracy. Clipping is a common technique to limitthe damage.

Gradient clipping forcesgradient values within a designated range during training.

SeeNumerical data:Normalizationin Machine Learning Crash Course for more information.

Cloud TPU

A specialized hardware accelerator designed to speed up machinelearning workloads on Google Cloud.

clustering

Grouping relatedexamples, particularly duringunsupervised learning. Once all theexamples are grouped, a human can optionally supply meaning to each cluster.

Many clustering algorithms exist. For example, thek-meansalgorithm clusters examples based on their proximity to acentroid, as in the following diagram:

A human researcher could then review the clusters and, for example,label cluster 1 as "dwarf trees" and cluster 2 as "full-size trees."

As another example, consider a clustering algorithm based on anexample's distance from a center point, illustrated as follows:

See theClustering coursefor more information.

co-adaptation

An undesirable behavior in whichneurons predict patterns intraining data by relying almost exclusively on outputs of specific other neuronsinstead of relying on the network's behavior as a whole. When the patterns thatcause co-adaptation are not present in validation data, then co-adaptationcausesoverfitting.Dropout regularization reduces co-adaptationbecause dropout ensures neurons cannot rely solely on specific other neurons.

collaborative filtering

Makingpredictions about the interests of one userbased on the interests of many other users. Collaborative filteringis often used inrecommendation systems.

SeeCollaborativefilteringin the Recommendation Systems course for more information.

CommitmentBank (CB)

A dataset for evaluating an LLM's proficiency in determining whetherthe author of a passage believes a target clause within that passage.Each entry in the dataset contains:

• A passage
• A target clause within that passage
• A Boolean value indicating whether the passage's author believes the targetclause

For example:

• Passage: What fun to hear Artemis laugh. She's such a serious child.I didn't know she had a sense of humor.
• Target clause: she had a sense of humor
• Boolean: True, which means the author believes the target clause

CommitmentBank is a component of theSuperGLUE ensemble.

compact model

Any small model designed to run on small devices with limited computationalresources. For example, compact models can run on mobile phones, tablets, orembedded systems.

compute

(Noun) The computational resources used by a model or system, such asprocessing power, memory, and storage.

Seeaccelerator chips.

concept drift

A shift in the relationship between features and the label.Over time, concept drift reduces a model's quality.

During training, the model learns the relationship between the features andtheir labels in the training set. If the labels in the training set aregood proxies for the real-world, then the modelshould make goodreal world predictions. However, due to concept drift, the model'spredictions tend to degrade over time.

For example, consider abinary classificationmodel that predicts whether or not a certain car model is "fuel efficient."That is, the features could be:

• car weight
• engine compression
• transmission type

while the label is either:

• fuel efficient
• not fuel efficient

However, the concept of "fuel efficient car" keepschanging. A car model labeledfuel efficient in 1994 would almost certainlybe labelednot fuel efficient in 2024. A model suffering from concept drifttends to make less and less useful predictions over time.

Compare and contrast withnonstationarity.

condition

A condition is also called a split or a test.

Contrast condition withleaf.

See also:

• binary condition
• non-binary condition.
• axis-aligned-condition
• oblique-condition

SeeTypes of conditionsin the Decision Forests course for more information.

confabulation

Synonym forhallucination.

Confabulation is probably a more technically accurate term than hallucination.However, hallucination became popular first.

configuration

The process of assigning the initial property values used to train a model,including:

• the model's composinglayers
• the location of the data
• hyperparameters such as:
  • learning rate
  • iterations
  • optimizer
  • loss function

In machine learning projects, configuration can be done through a specialconfiguration file or using configuration libraries such as the following:

• HParam
• Gin
• Fiddle

confirmation bias

The tendency to search for, interpret, favor, and recall information in away that confirms one's pre-existing beliefs or hypotheses.Machine learning developers may inadvertently collect or labeldata in ways that influence an outcome supporting their existingbeliefs. Confirmation bias is a form ofimplicit bias.

Experimenter's bias is a form of confirmation bias in whichan experimenter continues training models until a pre-existinghypothesis is confirmed.

confusion matrix

An NxN table that summarizes the number of correct and incorrect predictionsthat aclassification model made.For example, consider the following confusion matrix for abinary classification model:

The preceding confusion matrix shows the following:

• Of the 19 predictions in whichground truth was Tumor,the model correctly classified 18 and incorrectly classified 1.
• Of the 458 predictions in which ground truth was Non-Tumor, the modelcorrectly classified 452 and incorrectly classified 6.

The confusion matrix for amulti-class classificationproblem can help you identify patterns of mistakes.For example, consider the following confusion matrix for a 3-classmulti-class classification model that categorizes three different iris types(Virginica, Versicolor, and Setosa). When the ground truth was Virginica, theconfusion matrix shows that the model was far more likely to mistakenlypredict Versicolor than Setosa:

As yet another example, a confusion matrix could reveal that a model trainedto recognize handwritten digits tends to mistakenly predict 9 instead of 4,or mistakenly predict 1 instead of 7.

Confusion matrixes contain sufficient information to calculate avariety of performance metrics, includingprecisionandrecall.

constituency parsing

Dividing a sentence into smaller grammatical structures ("constituents").A later part of the ML system, such as anatural language understanding model,can parse the constituents more easily than the original sentence. For example,consider the following sentence:

My friend adopted two cats.

A constituency parser can divide this sentence into the followingtwo constituents:

• My friend is a noun phrase.
• adopted two cats is a verb phrase.

These constituents can be further subdivided into smaller constituents.For example, the verb phrase

adopted two cats

could be further subdivided into:

• adopted is a verb.
• two cats is another noun phrase.

contextualized language embedding

Anembedding that comes close to "understanding" wordsand phrases in ways that fluent human speakers can. Contextualized languageembeddings can understand complex syntax, semantics, and context.

For example, consider embeddings of the English wordcow. Older embeddingssuch asword2vec can represent Englishwords such that the distance in theembedding spacefromcow tobull is similar to the distance fromewe (female sheep) toram (male sheep) or fromfemale tomale. Contextualized languageembeddings can go a step further by recognizing that English speakers sometimescasually use the wordcow to mean either cow or bull.

context window

The number oftokens a model can process in a givenprompt. The larger the context window, the more informationthe model can use to provide coherent and consistentresponsesto the prompt.

continuous feature

A floating-pointfeature with an infinite range of possiblevalues, such as temperature or weight.

Contrast withdiscrete feature.

convenience sampling

Using a dataset not gathered scientifically in order to run quickexperiments. Later on, it's essential to switch to a scientifically gathereddataset.

convergence

A state reached whenloss values change very little ornot at all with eachiteration. For example, the followingloss curve suggests convergence at around 700 iterations:

A modelconverges when additional training won'timprove the model.

Indeep learning, loss values sometimes stay constant ornearly so for many iterations before finally descending. During a long periodof constant loss values, you may temporarily get a false sense of convergence.

See alsoearly stopping.

SeeModel convergence and losscurvesin Machine Learning Crash Course for more information.

conversational coding

An iterative dialog between you and a generative AI model for the purposeof creating software. You issue a prompt describing some software. Then,the model uses that description to generate code. Then, you issue a newprompt to address the flaws in the previous prompt or in the generatedcode, and the model generates updated code. You two keep going back andforth until the generated software is good enough.

Conversation coding is essentially the original meaning ofvibe coding.

Contrast withspecificational coding.

convex function

A function in which the region above the graph of the function is aconvex set. The prototypical convex function isshaped something like the letterU. For example, the followingare all convex functions:

In contrast, the following function is not convex. Notice how theregion above the graph is not a convex set:

Astrictly convex function has exactly one local minimum point, whichis also the global minimum point. The classic U-shaped functions arestrictly convex functions. However, some convex functions(for example, straight lines) are not U-shaped.

SeeConvergence and convexfunctionsin Machine Learning Crash Course for more information.

convex optimization

The process of using mathematical techniques such asgradient descent to findthe minimum of aconvex function.A great deal of research in machine learning has focused on formulating variousproblems as convex optimization problems and in solving those problems moreefficiently.

For complete details, see Boyd and Vandenberghe,ConvexOptimization.

convex set

A subset of Euclidean space such that a line drawn between any two points in thesubset remains completely within the subset. For instance, the following twoshapes are convex sets:

In contrast, the following two shapes are not convex sets:

convolution

In mathematics, casually speaking, a mixture of two functions. In machinelearning, a convolution mixes theconvolutionalfilter and the input matrixin order to trainweights.

The term "convolution" in machine learning is often a shorthand way ofreferring to eitherconvolutional operationorconvolutional layer.

Without convolutions, a machine learning algorithm would have to learna separate weight for every cell in a largetensor. For example,a machine learning algorithm training on 2K x 2K images would be forced tofind 4M separate weights. Thanks to convolutions, a machine learningalgorithm only has to find weights for every cell in theconvolutional filter, dramatically reducingthe memory needed to train the model. When the convolutional filter isapplied, it is simply replicated across cells such that each is multipliedby the filter.

convolutional filter

One of the two actors in aconvolutional operation. (The other actoris a slice of an input matrix.) A convolutional filter is a matrix havingthe samerank as the input matrix, but a smaller shape.For example, given a 28x28 input matrix, the filter could be any 2D matrixsmaller than 28x28.

In photographic manipulation, all the cells in a convolutional filter aretypically set to a constant pattern of ones and zeroes. In machine learning,convolutional filters are typically seeded with random numbers and then thenetworktrains the ideal values.

convolutional layer

A layer of adeep neural network in which aconvolutional filter passes along an inputmatrix. For example, consider the following 3x3convolutional filter:

The following animation shows a convolutional layer consisting of 9convolutional operations involving the 5x5 input matrix. Notice that eachconvolutional operation works on a different 3x3 slice of the input matrix.The resulting 3x3 matrix (on the right) consists of the results of the 9convolutional operations:

convolutional neural network

Aneural network in which at least one layer is aconvolutional layer. A typical convolutionalneural network consists of some combination of the following layers:

• convolutional layers
• pooling layers
• dense layers

Convolutional neural networks have had great success in certain kindsof problems, such as image recognition.

convolutional operation

The following two-step mathematical operation:

1. Element-wise multiplication of theconvolutional filter and a slice of aninput matrix. (The slice of the input matrix has the same rank andsize as the convolutional filter.)
2. Summation of all the values in the resulting product matrix.

For example, consider the following 5x5 input matrix:

Now imagine the following 2x2 convolutional filter:

Each convolutional operation involves a single 2x2 slice of theinput matrix. For example, suppose we use the 2x2 slice at thetop-left of the input matrix. So, the convolution operation onthis slice looks as follows:

Aconvolutional layer consists of aseries of convolutional operations, each acting on a different sliceof the input matrix.

COPA

Abbreviation forChoice of Plausible Alternatives.

cost

Synonym forloss.

co-training

Asemi-supervised learning approachparticularly useful when all of the following conditions are true:

• The ratio ofunlabeled examples tolabeled examples in the dataset is high.
• This is a classification problem (binary ormulti-class).
• Thedataset contains two different sets ofpredictive features that are independent of each other and complementary.

Co-training essentially amplifies independent signals into a stronger signal.For example, consider aclassification model thatcategorizes individual used cars as eitherGood orBad. One set ofpredictive features might focus on aggregate characteristics such as the year,make, and model of the car; another set of predictive features might focus onthe previous owner's driving record and the car's maintenance history.

The seminal paper on co-training isCombining Labeled and Unlabeled Data withCo-Training byBlum and Mitchell.

counterfactual fairness

Afairness metric that checks whether aclassification model produces the same result forone individual as it does for another individual who is identical to the first,except with respect to one or moresensitive attributes. Evaluating aclassification model for counterfactual fairness isone method for surfacing potential sources of bias in a model.

See either of the following for more information:

• Fairness: Counterfactualfairnessin Machine Learning Crash Course.
• When Worlds Collide: Integrating Different Counterfactual Assumptionsin Fairness

coverage bias

Seeselection bias.

crash blossom

A sentence or phrase with an ambiguous meaning.Crash blossoms present a significant problem innaturallanguage understanding.For example, the headlineRed Tape Holds Up Skyscraper is acrash blossom because an NLU model could interpret the headline literally orfiguratively.

critic

Synonym forDeep Q-Network.

cross-entropy

A generalization ofLog Loss tomulti-class classification problems. Cross-entropyquantifies the difference between two probability distributions. See alsoperplexity.

cross-validation

A mechanism for estimating how well amodel would generalize tonew data by testing the model against one or more non-overlapping data subsetswithheld from thetraining set.

cumulative distribution function (CDF)

A function that defines the frequency of samples less than or equal to atarget value. For example, consider a normal distribution of continuous values.A CDF tells you that approximately 50% of samples should be less than or equalto the mean and that approximately 84% of samples should be less than or equalto one standard deviation above the mean.

D

data analysis

Obtaining an understanding of data by considering samples, measurement,and visualization. Data analysis can be particularly useful when adataset is first received, before one builds the firstmodel.It is also crucial in understanding experiments and debugging problems withthe system.

data augmentation

Artificially boosting the range and number oftraining examplesby transforming existingexamples to create additional examples. For example,suppose images are one of yourfeatures, but your dataset doesn'tcontain enough image examples for the model to learn useful associations.Ideally, you'd add enoughlabeled images to your dataset toenable your model to train properly. If that's not possible, data augmentationcan rotate, stretch, and reflect each image to produce many variants of theoriginal picture, possibly yielding enough labeled data to enable excellenttraining.

DataFrame

A popularpandas data type for representingdatasets in memory.

A DataFrame is analogous to a table or a spreadsheet. Each column ofa DataFrame has a name (a header), and each row is identified by aunique number.

Each column in a DataFrame is structured like a 2D array, except thateach column can be assigned its own data type.

See also the officialpandas.DataFrame referencepage.

data parallelism

A way of scalingtraining orinferencethat replicates an entiremodel ontomultiple devices and then passes a subset of the input data to each device.Data parallelism can enable training and inference on very largebatch sizes; however, data parallelism requires that themodel be small enough to fit on all devices.

Data parallelism typically speeds training and inference.

See alsomodel parallelism.

Dataset API (tf.data)

A high-levelTensorFlow API for reading data andtransforming it into a form that a machine learning algorithm requires.Atf.data.Dataset object represents a sequence of elements, in whicheach element contains one or moreTensors. Atf.data.Iteratorobject provides access to the elements of aDataset.

data set or dataset

A collection of raw data, commonly (but not exclusively) organized in oneof the following formats:

• a spreadsheet
• a file in CSV (comma-separated values) format

decision boundary

The separator betweenclasses learned by amodel in abinary class ormulti-class classification problems. For example,in the following image representing a binary classification problem,the decision boundary is the frontier between the orange class andthe blue class:

decision forest

A model created from multipledecision trees.A decision forest makes a prediction by aggregating the predictions ofits decision trees. Popular types of decision forests includerandom forests andgradient boosted trees.

See theDecisionForestssection in the Decision Forests course for more information.

decision threshold

Synonym forclassification threshold.

decision tree

A supervised learning model composed of a set ofconditions andleaves organized hierarchically.For example, the following is a decision tree:

decoder

In general, any ML system that converts from a processed, dense, orinternal representation to a more raw, sparse, or external representation.

Decoders are often a component of a larger model, where they are frequentlypaired with anencoder.

Insequence-to-sequence tasks, a decoderstarts with the internal state generated by the encoder to predict the nextsequence.

Refer toTransformer for the definition of a decoder withinthe Transformer architecture.

SeeLarge language modelsin Machine Learning Crash Course for more information.

deep model

Aneural network containing more than onehidden layer.

A deep model is also called adeep neural network.

Contrast withwide model.

deep neural network

Synonym fordeep model.

Deep Q-Network (DQN)

InQ-learning, a deepneural networkthat predictsQ-functions.

Critic is a synonym for Deep Q-Network.

demographic parity

Afairness metric that is satisfied ifthe results of a model's classification are not dependent on agivensensitive attribute.

For example, if both Lilliputians and Brobdingnagians apply toGlubbdubdrib University, demographic parity is achieved if the percentageof Lilliputians admitted is the same as the percentage of Brobdingnagiansadmitted, irrespective of whether one group is on average more qualifiedthan the other.

Contrast withequalized odds andequality of opportunity, which permitclassification results in aggregate to depend on sensitive attributes,but don't permit classification results for certain specifiedground truth labels to depend on sensitive attributes. See"Attackingdiscrimination with smarter machine learning" for a visualizationexploring the tradeoffs when optimizing for demographic parity.

SeeFairness: demographicparityin Machine Learning Crash Course for more information.

denoising

A common approach toself-supervised learningin which:

1. Noise is artificially added to the dataset.
2. Themodel tries to remove the noise.

Denoising enables learning fromunlabeled examples.The originaldataset serves as the target orlabel andthe noisy data as the input.

Somemasked language models use denoisingas follows:

1. Noise is artificially added to an unlabeled sentence by masking some of the tokens.
2. The model tries to predict the original tokens.

dense feature

Afeature in which most or all values are nonzero, typicallyaTensor of floating-point values. For example, the following10-element Tensor is dense because 9 of its values are nonzero:

Contrast withsparse feature.

dense layer

Synonym forfully connected layer.

depth

The sum of the following in aneural network:

• the number ofhidden layers
• the number ofoutput layers, which is typically 1
• the number of anyembedding layers

For example, a neural network with five hidden layers and one output layerhas a depth of 6.

Notice that theinput layer doesn'tinfluence depth.

depthwise separable convolutional neural network (sepCNN)

Aconvolutional neural networkarchitecture based onInception,but where Inception modules are replaced with depthwise separableconvolutions. Also known as Xception.

A depthwise separable convolution (also abbreviated as separable convolution)factors a standard 3D convolution into two separate convolution operationsthat are more computationally efficient: first, a depthwise convolution,with a depth of 1 (n ✕ n ✕ 1), and then second, a pointwise convolution,with length and width of 1 (1 ✕ 1 ✕ n).

To learn more, seeXception: Deep Learning with Depthwise SeparableConvolutions.

derived label

Synonym forproxy label.

device

An overloaded term with the following two possible definitions:

1. A category of hardware that can run a TensorFlow session, includingCPUs, GPUs, andTPUs.
2. When training an ML model onaccelerator chips(GPUs or TPUs), the part of the system that actually manipulatestensors andembeddings.The device runs on accelerator chips. In contrast, thehosttypically runs on a CPU.

differential privacy

In machine learning, an anonymization approach to protect any sensitive data(for example, an individual's personal information) included in a model'straining set from being exposed. This approach ensuresthat themodel doesn't learn or remember much about a specificindividual. This is accomplished by sampling and adding noise during modeltraining to obscure individual data points, mitigating the risk of exposingsensitive training data.

Differential privacy is also used outside of machine learning. For example,data scientists sometimes use differential privacy to protect individualprivacy when computing product usage statistics for different demographics.

dimension reduction

Decreasing the number of dimensions used to represent a particular featurein a feature vector, typically byconverting to anembedding vector.

dimensions

Overloaded term having any of the following definitions:

• The number of levels of coordinates in aTensor. Forexample:

  • A scalar has zero dimensions; for example,["Hello"].
  • A vector has one dimension; for example,[3, 5, 7, 11].
  • A matrix has two dimensions; for example,[[2, 4, 18], [5, 7, 14]].You can uniquely specify a particular cell in a one-dimensional vectorwith one coordinate; you need two coordinates to uniquely specify aparticular cell in a two-dimensional matrix.

• The number of entries in afeature vector.

• The number of elements in anembedding layer.

direct prompting

Synonym forzero-shot prompting.

discrete feature

Afeature with a finite set of possible values. For example,a feature whose values may only beanimal,vegetable, ormineral is adiscrete (or categorical) feature.

Contrast withcontinuous feature.

discriminative model

Amodel that predictslabels from a set of one ormorefeatures. More formally, discriminative models define theconditional probability of an output given the features andweights; that is:

p(output | features, weights)

For example, a model that predicts whether an email is spam from featuresand weights is a discriminative model.

The vast majority of supervised learning models, including classificationand regression models, are discriminative models.

Contrast withgenerative model.

discriminator

A system that determines whetherexamples are real or fake.

Alternatively, the subsystem within agenerative adversarialnetwork that determines whetherthe examples created by thegenerator are real or fake.

SeeThe discriminatorin the GAN course for more information.

disparate impact

Making decisions about people that impact different populationsubgroups disproportionately. This usually refers to situationswhere an algorithmic decision-making process harms or benefitssome subgroups more than others.

For example, suppose an algorithm that determines a Lilliputian'seligibility for a miniature-home loan is more likely to classifythem as "ineligible" if their mailing address contains a certainpostal code. If Big-Endian Lilliputians are more likely to havemailing addresses with this postal code than Little-Endian Lilliputians,then this algorithm may result in disparate impact.

Contrast withdisparate treatment,which focuses on disparities that result when subgroup characteristicsare explicit inputs to an algorithmic decision-making process.

disparate treatment

Factoring subjects'sensitive attributesinto an algorithmic decision-making process such that different subgroupsof people are treated differently.

For example, consider an algorithm thatdetermines Lilliputians' eligibility for a miniature-home loan based on thedata they provide in their loan application. If the algorithm uses aLilliputian's affiliation as Big-Endian or Little-Endian as an input, itis enacting disparate treatment along that dimension.

Contrast withdisparate impact, which focuseson disparities in the societal impacts of algorithmic decisions on subgroups,irrespective of whether those subgroups are inputs to the model.

distillation

The process of reducing the size of onemodel (known as theteacher) into a smaller model (known as thestudent) that emulatesthe original model's predictions as faithfully as possible. Distillationis useful because the smaller model has two key benefits over the largermodel (the teacher):

• Faster inference time
• Reduced memory and energy usage

However, the student's predictions are typically not as good asthe teacher's predictions.

Distillation trains the student model to minimize aloss function based on the difference between the outputsof the predictions of the student and teacher models.

Compare and contrast distillation with the following terms:

• fine-tuning
• prompt-based learning

SeeLLMs: Fine-tuning, distillation, and promptengineeringin Machine Learning Crash Course for more information.

distribution

The frequency and range of different values for a givenfeature orlabel.A distribution captures how likely a particular value is.

The following image shows histograms of two different distributions:

• On the left, a power law distribution of wealth versus the number of peoplepossessing that wealth.
• On the right, a normal distribution of height versus the number of peoplepossessing that height.

Understanding each feature and label's distribution can help you determine howtonormalize values and detectoutliers.

The phraseout of distribution refers to a value that doesn't appear in thedataset or is very rare. For example, an image of the planet Saturn would beconsidered out of distribution for a dataset consisting of cat images.

divisive clustering

Seehierarchical clustering.

downsampling

Overloaded term that can mean either of the following:

• Reducing the amount of information in afeature inorder totrain a model more efficiently. For example,before training an image recognition model, downsampling high-resolutionimages to a lower-resolution format.
• Training on a disproportionately low percentage of over-representedclassexamples in order to improve model training on under-represented classes.For example, in aclass-imbalanceddataset, models tend to learn a lot about themajority class and not enough about theminority class. Downsampling helpsbalance the amount of training on the majority and minority classes.

SeeDatasets: Imbalanceddatasetsin Machine Learning Crash Course for more information.

DQN

Abbreviation forDeep Q-Network.

dropout regularization

A form ofregularization useful in trainingneural networks. Dropout regularizationremoves a random selection of a fixed number of the units in a networklayer for a single gradient step. The more units dropped out, the strongerthe regularization. This is analogous to training the network to emulatean exponentially largeensemble of smaller networks.For full details, seeDropout: A Simple Way to Prevent Neural Networks fromOverfitting.

dynamic

Something done frequently or continuously.The termsdynamic andonline are synonyms in machine learning.The following are common uses ofdynamic andonline in machinelearning:

• Adynamic model (oronline model) is a modelthat is retrained frequently or continuously.
• Dynamic training (oronline training) is the process of trainingfrequently or continuously.
• Dynamic inference (oronline inference) is the process ofgenerating predictions on demand.

dynamic model

Amodel that is frequently (maybe even continuously)retrained. A dynamic model is a "lifelong learner" thatconstantly adapts to evolving data. A dynamic model is also known as anonline model.

Contrast withstatic model.

E

eager execution

A TensorFlow programming environment in whichoperationsrun immediately. In contrast, operations called ingraph execution don't run until they are explicitlyevaluated. Eager execution is animperative interface, muchlike the code in most programming languages. Eager execution programs aregenerally far easier to debug than graph execution programs.

early stopping

A method forregularization that involves endingtrainingbefore training loss finishesdecreasing. In early stopping, you intentionally stop training the modelwhen the loss on avalidation dataset starts toincrease; that is, whengeneralization performance worsens.

Contrast withearly exit.

earth mover's distance (EMD)

A measure of the relative similarity of twodistributions.The lower the earth mover's distance, the more similar the distributions.

edit distance

A measurement of how similar two text strings are to each other.In machine learning, edit distance is useful for the following reasons:

• Edit distance is easy to compute.
• Edit distance can compare two strings known to be similar to each other.
• Edit distance can determine the degree to which different strings aresimilar to a given string.

Several definitions of edit distance exist, each using different stringoperations. SeeLevenshtein distance for anexample.

Einsum notation

An efficient notation for describing how twotensors are to becombined. The tensors are combined by multiplying the elements of one tensorby the elements of the other tensor and then summing the products.Einsum notation uses symbols to identify the axes of each tensor, and thosesame symbols are rearranged to specify the shape of the new resulting tensor.

NumPy provides a common Einsum implementation.

embedding layer

A specialhidden layer that trains on ahigh-dimensionalcategorical feature togradually learn a lower dimension embedding vector. Anembedding layer enables a neural network to train far moreefficiently than training just on the high-dimensional categorical feature.

For example, Earth currently supports about 73,000 tree species. Supposetree species is afeature in your model, so your model'sinput layer includes aone-hot vector 73,000elements long.For example, perhapsbaobab would be represented something like this:

A 73,000-element array is very long. If you don't add an embedding layerto the model, training is going to be very time consuming due tomultiplying 72,999 zeros. Perhaps you pick the embedding layer to consistof 12 dimensions. Consequently, the embedding layer will gradually learna new embedding vector for each tree species.

In certain situations,hashing is a reasonable alternativeto an embedding layer.

SeeEmbeddingsin Machine Learning Crash Course for more information.

embedding space

The d-dimensional vector space that features from a higher-dimensionalvector space are mapped to. Embedding space is trained to capture structurethat is meaningful for the intended application.

Thedot productof two embeddings is a measure of their similarity.

embedding vector

Broadly speaking, an array of floating-point numbers taken fromanyhidden layer that describe the inputs to that hidden layer.Often, an embedding vector is the array of floating-point numbers trained inan embedding layer. For example, suppose an embedding layer must learn anembedding vector for each of the 73,000 tree species on Earth. Perhaps thefollowing array is the embedding vector for a baobab tree:

An embedding vector is not a bunch of random numbers. An embedding layerdetermines these values through training, similar to the way aneural network learns other weights during training. Each element of thearray is a rating along some characteristic of a tree species. Whichelement represents which tree species' characteristic? That's very hardfor humans to determine.

The mathematically remarkable part of an embedding vector is that similaritems have similar sets of floating-point numbers. For example, similartree species have a more similar set of floating-point numbers thandissimilar tree species. Redwoods and sequoias are related tree species,so they'll have a more similar set of floating-pointing numbers thanredwoods and coconut palms. The numbers in the embedding vector willchange each time you retrain the model, even if you retrain the modelwith identical input.

empirical cumulative distribution function (eCDF orEDF)

Acumulative distribution functionbased onempirical measurements from a real dataset. The value of thefunction at any point along the x-axis is the fraction of observations inthe dataset that are less than or equal to the specified value.

empirical risk minimization (ERM)

Choosing the function that minimizes loss on the training set. Contrastwithstructural risk minimization.

encoder

In general, any ML system that converts from a raw, sparse, or externalrepresentation into a more processed, denser, or more internal representation.

Encoders are often a component of a larger model, where they are frequentlypaired with adecoder. SomeTransformerspair encoders with decoders, though other Transformers use only the encoderor only the decoder.

Some systems use the encoder's output as the input to a classification orregression network.

Insequence-to-sequence tasks, an encodertakes an input sequence and returns an internal state (a vector). Then, thedecoder uses that internal state to predict the next sequence.

Refer toTransformer for the definition of an encoder inthe Transformer architecture.

SeeLLMs: What's a large languagemodelin Machine Learning Crash Course for more information.

endpoints

A network-addressable location (typically a URL) where a service can be reached.

ensemble

A collection ofmodels trained independently whose predictionsare averaged or aggregated. In many cases, an ensemble produces betterpredictions than a single model. For example, arandom forest is an ensemble built from multipledecision trees. Note that not alldecision forests are ensembles.

SeeRandomForestin Machine Learning Crash Course for more information.

entropy

Ininformation theory,a description of how unpredictable a probabilitydistribution is. Alternatively, entropy is also defined as how muchinformation eachexample contains. A distribution hasthe highest possible entropy when all values of a random variable areequally likely.

The entropy of a set with two possible values "0" and "1" (for example,the labels in abinary classification problem)has the following formula:

  H = -p log p - q log q = -p log p - (1-p) * log (1-p)

where:

• H is the entropy.
• p is the fraction of "1" examples.
• q is the fraction of "0" examples. Note that q = (1 - p)
• log is generally log₂. In this case, the entropyunit is a bit.

For example, suppose the following:

• 100 examples contain the value "1"
• 300 examples contain the value "0"

Therefore, the entropy value is:

• p = 0.25
• q = 0.75
• H = (-0.25)log₂(0.25) - (0.75)log₂(0.75) = 0.81 bits per example

A set that is perfectly balanced (for example, 200 "0"s and 200 "1"s)would have an entropy of 1.0 bit per example. As a set becomes moreimbalanced, its entropy moves towards 0.0.

Indecision trees, entropy helps formulateinformation gain to help thesplitter select theconditionsduring the growth of a classification decision tree.

Compare entropy with:

• gini impurity
• cross-entropy loss function

Entropy is often calledShannon's entropy.

SeeExact splitter for binary classification with numericalfeaturesin the Decision Forests course for more information.

environment

In reinforcement learning, the world that contains theagentand allows the agent to observe that world'sstate. For example,the represented world can be a game like chess, or a physical world like amaze. When the agent applies anaction to the environment,then the environment transitions between states.

episode

In reinforcement learning, each of the repeated attempts by theagent to learn anenvironment.

epoch

A full training pass over the entiretraining setsuch that eachexample has been processed once.

An epoch representsN/batch sizetrainingiterations, whereN is thetotal number of examples.

For instance, suppose the following:

• The dataset consists of 1,000 examples.
• The batch size is 50 examples.

Therefore, a single epoch requires 20 iterations:

1 epoch = (N/batch size) = (1,000 / 50) = 20 iterations

SeeLinear regression:Hyperparametersin Machine Learning Crash Course for more information.

epsilon greedy policy

In reinforcement learning, apolicy that either follows arandom policy with epsilon probability or agreedy policy otherwise. For example, if epsilon is0.9, then the policy follows a random policy 90% of the time and a greedypolicy 10% of the time.

Over successive episodes, the algorithm reduces epsilon's value in orderto shift from following a random policy to following a greedy policy. Byshifting the policy, the agent first randomly explores the environment andthen greedily exploits the results of random exploration.

equality of opportunity

Afairness metric to assess whether a model ispredicting the desirable outcome equally well for all values of asensitive attribute. In other words, if thedesirable outcome for a model is thepositive class,the goal would be to have thetrue positive rate be thesame for all groups.

Equality of opportunity is related toequalized odds,which requires thatboth the true positive rates andfalse positive rates are the same for all groups.

Suppose Glubbdubdrib University admits both Lilliputians and Brobdingnagiansto a rigorous mathematics program. Lilliputians' secondary schools offer arobust curriculum of math classes, and the vast majority of students arequalified for the university program. Brobdingnagians' secondary schools don'toffer math classes at all, and as a result, far fewer of their students arequalified. Equality of opportunity is satisfied for the preferred label of"admitted" with respect to nationality (Lilliputian or Brobdingnagian) ifqualified students are equally likely to be admitted irrespective of whetherthey're a Lilliputian or a Brobdingnagian.

For example, suppose 100 Lilliputians and 100 Brobdingnagians apply toGlubbdubdrib University, and admissions decisions are made as follows:

Table 1. Lilliputian applicants (90% are qualified)

Table 2. Brobdingnagian applicants (10% are qualified):

The preceding examples satisfy equality of opportunity for acceptance ofqualified students because qualified Lilliputians and Brobdingnagians bothhave a 50% chance of being admitted.

While equality of opportunity is satisfied, the following two fairness metricsare not satisfied:

• demographic parity: Lilliputians and Brobdingnagians are admitted to the university at different rates; 48% of Lilliputians students are admitted, but only 14% of Brobdingnagian students are admitted.
• equalized odds: While qualified Lilliputian and Brobdingnagian students both have the same chance of being admitted, the additional constraint that unqualified Lilliputians and Brobdingnagians both have the same chance of being rejected is not satisfied. Unqualified Lilliputians have a 70% rejection rate, whereas unqualified Brobdingnagians have a 90% rejection rate.

SeeFairness: Equality ofopportunityin Machine Learning Crash Course for more information.

equalized odds

A fairness metric to assess whether a model is predicting outcomes equallywell for all values of asensitive attribute withrespect to both thepositive class andnegative class—not just one class or the otherexclusively. In other words, both thetrue positive rateandfalse negative rate should be the same forall groups.

Equalized odds is related toequality of opportunity, which only focuseson error rates for a single class (positive or negative).

For example, suppose Glubbdubdrib University admits both Lilliputians andBrobdingnagians to a rigorous mathematics program. Lilliputians' secondaryschools offer a robust curriculum of math classes, and the vast majority ofstudents are qualified for the university program. Brobdingnagians' secondaryschools don't offer math classes at all, and as a result, far fewer oftheir students are qualified. Equalized odds is satisfied provided that nomatter whether an applicant is a Lilliputian or a Brobdingnagian, if theyare qualified, they are equally as likely to get admitted to the program,and if they are not qualified, they are equally as likely to get rejected.

Suppose 100 Lilliputians and 100 Brobdingnagians apply to GlubbdubdribUniversity, and admissions decisions are made as follows:

Table 3. Lilliputian applicants (90% are qualified)

Table 4. Brobdingnagian applicants (10% are qualified):

Equalized odds is satisfied because qualified Lilliputian and Brobdingnagianstudents both have a 50% chance of being admitted, and unqualified Lilliputianand Brobdingnagian have an 80% chance of being rejected.

Equalized odds is formally defined in"Equality ofOpportunity in Supervised Learning" as follows:"predictor Ŷ satisfies equalized odds with respectto protected attribute A and outcome Y if Ŷ and A are independent,conditional on Y."

Estimator

A deprecated TensorFlow API. Usetf.keras insteadof Estimators.

evals

Primarily used as an abbreviation forLLM evaluations.More broadly,evals is an abbreviation for any form ofevaluation.

evaluation

The process of measuring a model's quality or comparing different modelsagainst each other.

To evaluate asupervised machine learningmodel, you typically judge it against avalidation setand atest set.Evaluating a LLMtypically involves broader quality and safety assessments.

exact match

An all-or-nothing metric in which the model's output either matchesground truth or thereference textexactly or it doesn't. For example, if ground truth isorange, the onlymodel output that satisfies exact match isorange.

Exact match can also evaluate models whose output is a sequence(a ranked list of items). In general, exact match requires the generatedranked list to exactly match ground truth; that is, each itemin both lists must be in the same order. That said, if ground truthconsists ofmultiple correct sequences, then exact match only requiresmodel's output matchesone of the correct sequences.

example

The values of one row offeatures and possiblyalabel. Examples insupervised learning fall into twogeneral categories:

• Alabeled example consists of one or more featuresand a label. Labeled examples are used during training.
• Anunlabeled example consists of one ormore features but no label. Unlabeled examples are used during inference.

For instance, suppose you are training a model to determine the influenceof weather conditions on student test scores. Here are three labeled examples:

Here are three unlabeled examples:

The row of adataset is typically the raw source for an example.That is, an example typically consists of a subset of the columns inthe dataset. Furthermore, the features in an example can also includesynthetic features, such asfeature crosses.

SeeSupervised Learning inthe Introduction to Machine Learning course for more information.

experience replay

In reinforcement learning, aDQN technique used toreduce temporal correlations in training data. Theagentstores state transitions in areplay buffer, and thensamples transitions from the replay buffer to create training data.

experimenter's bias

Seeconfirmation bias.

exploding gradient problem

The tendency forgradients indeep neural networks (especiallyrecurrent neural networks) to becomesurprisingly steep (high). Steep gradients often cause very large updatesto theweights of eachnode in adeep neural network.

Models suffering from the exploding gradient problem become difficultor impossible to train.Gradient clippingcan mitigate this problem.

Compare tovanishing gradient problem.

Extreme Summarization (xsum)

A dataset for evaluating an LLM's ability to summarize a single document.Each entry in the dataset consists of:

• A document authored by the British Broadcasting Corporation (BBC).
• A one-sentence summary of that document.

For details, seeDon't Give Me the Details, Just the Summary! Topic-Aware Convolutional Neural Networks for Extreme Summarization.

F

F₁

A "roll-up"binary classification metric thatrelies on bothprecision andrecall.Here is the formula:

factuality

Within the ML world, a property describing a model whose output is basedon reality. Factuality is a concept rather than a metric.For example, suppose you send the followingpromptto alarge language model:

What is the chemical formula for table salt?

A model optimizing factuality would respond:

NaCl

It is tempting to assume that all models should be based on factuality.However, some prompts, such as the following, should cause a generative AI modelto optimizecreativity rather thanfactuality.

Tell me a limerick about an astronaut and a caterpillar.

It is unlikely that the resulting limerick would be based on reality.

Contrast withgroundedness.

fairness constraint

• Post-processing your model's output.
• Altering theloss function to incorporate a penalty for violating afairness metric.
• Directly adding a mathematical constraint to an optimization problem.

fairness metric

A mathematical definition of "fairness" that is measurable.Some commonly used fairness metrics include:

• equalized odds
• predictive parity
• counterfactual fairness
• demographic parity

Many fairness metrics are mutually exclusive; seeincompatibility of fairness metrics.

false negative (FN)

An example in which the model mistakenly predicts thenegative class. For example, the modelpredicts that a particular email message isnot spam(the negative class), but that email messageactually is spam.

false negative rate

The proportion of actual positive examples for which the model mistakenlypredicted the negative class. The following formula calculates the falsenegative rate:

SeeThresholds and the confusionmatrixin Machine Learning Crash Course for more information.

false positive (FP)

An example in which the model mistakenly predicts thepositive class. For example, the model predictsthat a particular email message isspam (the positive class), but thatemail message isactually not spam.

SeeThresholds and the confusionmatrixin Machine Learning Crash Course for more information.

false positive rate (FPR)

The proportion of actual negative examples for which the model mistakenlypredicted the positive class. The following formula calculates the falsepositive rate:

The false positive rate is the x-axis in anROC curve.

SeeClassification: ROC andAUCin Machine Learning Crash Course for more information.

fast decay

Atraining technique to improve the performance ofLLMs. Fast decay involves rapidly decreasingthelearning rate during training.This strategy helps prevent the model fromoverfitting tothe training data, and improvesgeneralization.

feature

An input variable to a machine learning model. Anexampleconsists of one or more features. For instance, suppose you are training amodel to determine the influence of weather conditions on student test scores.The following table shows three examples, each of which containsthree features and one label:

Contrast withlabel.

SeeSupervised Learningin the Introduction to Machine Learning course for more information.

feature cross

Asynthetic feature formed by "crossing"categorical orbucketed features.

For example, consider a "mood forecasting" model that representstemperature in one of the following four buckets:

• freezing
• chilly
• temperate
• warm

And represents wind speed in one of the following three buckets:

• still
• light
• windy

Without feature crosses, the linear model trains independently on each of thepreceding seven various buckets. So, the model trains on, for example,freezing independently of the training on, for example,windy.

Alternatively, you could create a feature cross of temperature andwind speed. This synthetic feature would have the following 12 possiblevalues:

• freezing-still
• freezing-light
• freezing-windy
• chilly-still
• chilly-light
• chilly-windy
• temperate-still
• temperate-light
• temperate-windy
• warm-still
• warm-light
• warm-windy

Thanks to feature crosses, the model can learn mood differencesbetween afreezing-windy day and afreezing-still day.

If you create a synthetic feature from two features that each have a lot ofdifferent buckets, the resulting feature cross will have a huge numberof possible combinations. For example, if one feature has 1,000 buckets andthe other feature has 2,000 buckets, the resulting feature cross has 2,000,000buckets.

Formally, a cross is aCartesian product.

Feature crosses are mostly used with linear models and are rarely usedwith neural networks.

SeeCategorical data: Featurecrossesin Machine Learning Crash Course for more information.

feature engineering

A process that involves the following steps:

1. Determining whichfeatures might be usefulin training a model.
2. Converting raw data from the dataset into efficient versions ofthose features.

For example, you might determine thattemperature might be a usefulfeature. Then, you might experiment withbucketingto optimize what the model can learn from differenttemperature ranges.

Feature engineering is sometimes calledfeature extraction orfeaturization.

SeeNumerical data: How a model ingests data using featurevectorsin Machine Learning Crash Course for more information.

feature extraction

Overloaded term having either of the following definitions:

• Retrieving intermediate feature representations calculated by anunsupervised or pretrained model(for example,hidden layer values in aneural network) for use in another model as input.
• Synonym forfeature engineering.

feature importances

Synonym forvariable importances.

feature set

The group offeatures your machine learningmodel trains on.For example, a simple feature set for a model that predicts housing pricesmight consist of postal code, property size, and property condition.

feature spec

Describes the information required to extractfeatures datafrom thetf.Example protocol buffer. Because thetf.Example protocol buffer is just a container for data, you must specifythe following:

• The data to extract (that is, the keys for the features)
• The data type (for example, float or int)
• The length (fixed or variable)

feature vector

The array offeature values comprising anexample. The feature vector is input duringtraining and duringinference.For example, the feature vector for a model with two discrete featuresmight be:

[0.92, 0.56]

Each example supplies different values for the feature vector, so thefeature vector for the next example could be something like:

[0.73, 0.49]

Feature engineering determines how to representfeatures in the feature vector. For example, a binary categorical feature withfive possible values might be represented withone-hot encoding. In this case, the portion of thefeature vector for a particular example would consist of four zeroes anda single 1.0 in the third position, as follows:

[0.0, 0.0, 1.0, 0.0, 0.0]

As another example, suppose your model consists of three features:

• a binary categorical feature withfive possible values represented withone-hot encoding; for example:[0.0, 1.0, 0.0, 0.0, 0.0]
• another binary categorical feature withthree possible values representedwith one-hot encoding; for example:[0.0, 0.0, 1.0]
• a floating-point feature; for example:8.3.

In this case, the feature vector for each example would be representedbynine values. Given the example values in the preceding list, thefeature vector would be:

0.01.00.00.00.00.00.01.08.3

SeeNumerical data: How a model ingests data using featurevectorsin Machine Learning Crash Course for more information.

featurization

The process of extractingfeatures from an input source,such as a document or video, and mapping those features into afeature vector.

Some ML experts use featurization as a synonym forfeature engineering orfeature extraction.

federated learning

A distributed machine learning approach thattrainsmachine learningmodels using decentralizedexamples residing on devices such as smartphones.In federated learning, a subset of devices downloads the current modelfrom a central coordinating server. The devices use the examples storedon the devices to make improvements to the model. The devices then uploadthe model improvements (but not the training examples) to the coordinatingserver, where they are aggregated with other updates to yield an improvedglobal model. After the aggregation, the model updates computed by devicesare no longer needed, and can be discarded.

Since the training examples are never uploaded, federated learning follows theprivacy principles of focused data collection and data minimization.

See theFederated Learning comic(yes, a comic) for more details.

feedback loop

In machine learning, a situation in which a model's predictions influence thetraining data for the same model or another model. For example, a model thatrecommends movies will influence the movies that people see, which will theninfluence subsequent movie recommendation models.

SeeProduction ML systems: Questions toaskin Machine Learning Crash Course for more information.

feedforward neural network (FFN)

A neural network without cyclic or recursive connections. For example,traditionaldeep neural networks arefeedforward neural networks. Contrast withrecurrent neuralnetworks, which are cyclic.

few-shot learning

A machine learning approach, often used for object classification,designed to train effectiveclassification modelsfrom only a small number of training examples.

See alsoone-shot learning andzero-shot learning.

few-shot prompting

Aprompt that contains more than one (a "few") exampledemonstrating how thelarge language modelshould respond. For example, the following lengthy prompt contains twoexamples showing a large language model how to answer a query.

Few-shot prompting generally produces more desirable results thanzero-shot prompting andone-shot prompting. However, few-shot promptingrequires a lengthier prompt.

Few-shot prompting is a form offew-shot learningapplied toprompt-based learning.

SeePromptengineeringin Machine Learning Crash Course for more information.

Fiddle

A Python-firstconfiguration library that sets thevalues of functions and classes without invasive code or infrastructure.In the case ofPax—and other ML codebases—these functions andclasses representmodels andtraininghyperparameters.

Fiddleassumes that machine learning codebases are typically divided into:

• Library code, which defines the layers and optimizers.
• Dataset "glue" code, which calls the libraries and wires everything together.

Fiddle captures the call structure of the glue code in an unevaluated andmutable form.

fine-tuning

A second, task-specific training pass performed on apre-trained model to refine its parameters for aspecific use case. For example, the full training sequence for somelarge language models is as follows:

1. Pre-training: Train a large language model on a vastgeneral dataset,such as all the English language Wikipedia pages.
2. Fine-tuning: Train the pre-trained model to perform aspecific task,such as responding to medical queries. Fine-tuning typically involveshundreds or thousands of examples focused on the specific task.

As another example, the full training sequence for a large image model is asfollows:

1. Pre-training: Train a large image model on a vastgeneral imagedataset, such as all the images in Wikimedia commons.
2. Fine-tuning: Train the pre-trained model to perform aspecific task,such as generating images of orcas.

Fine-tuning can entail any combination of the following strategies:

• Modifyingall of the pre-trained model's existingparameters. This is sometimes calledfull fine-tuning.
• Modifying onlysome of the pre-trained model's existing parameters(typically, the layers closest to theoutput layer),while keeping other existing parameters unchanged (typically, the layersclosest to theinput layer). Seeparameter-efficient tuning.
• Adding more layers, typically on top of the existing layers closest to theoutput layer.

Fine-tuning is a form oftransfer learning.As such, fine-tuning might use a different loss function or a different modeltype than those used to train the pre-trained model. For example, you couldfine-tune a pre-trained large image model to produce a regression model thatreturns the number of birds in an input image.

Compare and contrast fine-tuning with the following terms:

• distillation
• prompt-based learning

SeeFine-tuningin Machine Learning Crash Course for more information.

Flash model

A family of relatively smallGemini models optimized for speedand lowlatency. Flash models are designed for a widerange of applications where quick responses and high throughput are crucial.

Flax

A high-performance open-source library fordeep learning built on top ofJAX. Flax provides functionsfortrainingneural networks, as wellas methods for evaluating their performance.

Flaxformer

An open-sourceTransformerlibrary,built onFlax, designed primarily for natural language processingand multimodal research.

forget gate

The portion of aLong Short-Term Memorycell that regulates the flow of information through the cell.Forget gates maintain context by deciding which information to discardfrom the cell state.

foundation model

A very largepre-trained model trained on an enormousand diversetraining set. A foundation model can do bothof the following:

• Respond well to a wide range of requests.
• Serve as abase model for additionalfine-tuning or other customization.

In other words, a foundation model is already very capable in a general sensebut can be further customized to become even more useful for a specific task.

fraction of successes

A metric for evaluating an ML model'sgenerated text.The fraction of successes is the number of "successful" generated textoutputs divided by the total number of generated text outputs. For example,if alarge language model generated 10 blocksof code, five of which were successful, then the fraction of successeswould be 50%.

Although fraction of successes is broadly useful throughout statistics,within ML, this metric is primarily useful for measuring verifiable taskslike code generation or math problems.

full softmax

Synonym forsoftmax.

Contrast withcandidate sampling.

SeeNeural networks: Multi-classclassificationin Machine Learning Crash Course for more information.

fully connected layer

Ahidden layer in which eachnode isconnected toevery node in the subsequent hidden layer.

A fully connected layer is also known as adense layer.

function transformation

A function that takes a function as input and returns a transformed functionas output.JAX uses function transformations.

G

GAN

Abbreviation forgenerative adversarialnetwork.

Gemini

The ecosystem comprising Google's most advanced AI. Elements of this ecosysteminclude:

• VariousGemini models.
• The interactive conversational interface to a Gemini model.Users type prompts and Gemini responds to those prompts.
• Various Gemini APIs.
• Various business products based on Gemini models; for example,Gemini for Google Cloud.

Gemini models

Google's state-of-the-artTransformer-basedmultimodal models. Gemini models are specificallydesigned to integrate withagents.

Users can interact with Gemini models in a variety of ways, including throughan interactive dialog interface and through SDKs.

Gemma

A family of lightweight open models built from the sameresearch and technology used to create theGemini models. Severaldifferent Gemma models are available, each providing different features, such asvision, code, and instruction following. SeeGemmafor details.

GenAI or genAI

Abbreviation forgenerative AI.

generalization

Amodel's ability to make correct predictions on new,previously unseen data. A model that can generalize is the oppositeof a model that isoverfitting.

SeeGeneralizationin Machine Learning Crash Course for more information.

generalization curve

A plot of bothtraining loss andvalidation loss as a function of the number ofiterations.

A generalization curve can help you detect possibleoverfitting. For example, the followinggeneralization curve suggests overfitting because validation lossultimately becomes significantly higher than training loss.

SeeGeneralizationin Machine Learning Crash Course for more information.

generalized linear model

A generalization ofleast squares regressionmodels, which are based onGaussiannoise, to othertypes of models based on other types of noise, such asPoisson noiseorcategorical noise. Examples of generalized linear models include:

• logistic regression
• multi-class regression
• least squares regression

The parameters of a generalized linear model can be found throughconvex optimization.

Generalized linear models exhibit the following properties:

• The average prediction of the optimal least squares regression model isequal to the average label on the training data.
• The average probability predicted by the optimal logistic regressionmodel is equal to the average label on the training data.

The power of a generalized linear model is limited by its features. Unlikea deep model, a generalized linear model cannot "learn new features."

generated text

In general, the text that an ML model outputs. When evaluating largelanguage models, some metrics compare generated text againstreference text. For example, suppose you aretrying to determine how effectively an ML model translates from Frenchto Dutch. In this case:

• Thegenerated text is the Dutch translation that the ML model outputs.
• Thereference text is the Dutch translation that a human translator (orsoftware) creates.

Note that some evaluation strategies don't involve reference text.

generative adversarial network (GAN)

A system to create new data in which agenerator createsdata and adiscriminator determines whether thatcreated data is valid or invalid.

See theGenerative Adversarial Networks coursefor more information.

generative AI

An emerging transformative field with no formal definition.That said, most experts agree that generative AI models cancreate ("generate") content that is all of the following:

• complex
• coherent
• original

Examples of generative AI include:

• Large language models, which can generatesophisticated original text and answer questions.
• Image generation model, which can produce unique images.
• Audio and music generation models, which can compose original music orgenerate realistic speech.
• Video generation models, which can generate original videos.

Some earlier technologies, includingLSTMsandRNNs, can also generate original andcoherent content. Some experts view these earlier technologies asgenerative AI, while others feel that true generative AI requires more complexoutput than those earlier technologies can produce.

Contrast withpredictive ML.

generative model

Practically speaking, a model that does either of the following:

• Creates (generates) new examples from the training dataset.For example, a generative model could create poetry after trainingon a dataset of poems. Thegenerator part of agenerative adversarial networkfalls into this category.
• Determines the probability that a new example comes from thetraining set, or was created from the same mechanism that createdthe training set. For example, after training ona dataset consisting of English sentences, a generative model coulddetermine the probability that new input is a valid English sentence.

A generative model can theoretically discern the distribution of examplesor particular features in a dataset. That is:

p(examples)

Unsupervised learning models are generative.

Contrast withdiscriminative models.

generator

The subsystem within agenerative adversarialnetworkthat creates newexamples.

Contrast withdiscriminative model.

gini impurity

A metric similar toentropy.Splittersuse values derived from either gini impurity or entropy to composeconditions for classificationdecision trees.Information gain is derived from entropy.No universally accepted equivalent term for the metric derivedfrom gini impurity exists; however, this unnamed metric is just as important asinformation gain.

Gini impurity is also calledgini index, or simplygini.

golden dataset

A set of manually curated data that capturesground truth.Teams can use one or more golden datasets to evaluate a model's quality.

Some golden datasets capture different subdomains of ground truth. For example,a golden dataset for image classification might capture lighting conditionsand image resolution.

golden response

Aresponse known to be good. For example, given the followingprompt:

2 + 2

The golden response is hopefully:

4

Google AI Studio

A Google tool providing a user-friendly interfacefor experimenting with and building applications using Google'slarge language models.See theGoogle AI Studio home pagefor details.

GPT (Generative Pre-trained Transformer)

A family ofTransformer-basedlarge language models developed byOpenAI.

GPT variants can apply to multiplemodalities, including:

• image generation (for example, ImageGPT)
• text-to-image generation (for example,DALL-E).

gradient

The vector ofpartial derivatives with respect toall of the independent variables. In machine learning, the gradient isthe vector of partial derivatives of the model function. The gradient pointsin the direction of steepest ascent.

gradient accumulation

Abackpropagation technique that updates theparameters onlyonce per epoch rather than once periteration. After processing eachmini-batch, gradientaccumulation simply updates a running total of gradients. Then, afterprocessing the last mini-batch in the epoch, the system finally updatesthe parameters based on the total of all gradient changes.

Gradient accumulation is useful when thebatch size isvery large compared to the amount of available memory for training.When memory is an issue, the natural tendency is to reduce batch size.However, reducing the batch size in normal backpropagationincreasesthe number of parameter updates. Gradient accumulation enables the modelto avoid memory issues but still train efficiently.

gradient boosted (decision) trees (GBT)

A type ofdecision forest in which:

• Training relies ongradient boosting.
• The weak model is adecision tree.

SeeGradient Boosted DecisionTrees in theDecision Forests course for more information.

gradient boosting

A training algorithm where weak models are trained to iterativelyimprove the quality (reduce the loss) of a strong model. For example,a weak model could be a linear or small decision tree model.The strong model becomes the sum of all the previously trained weak models.

In the simplest form of gradient boosting, at each iteration, a weak modelis trained to predict the loss gradient of the strong model. Then, thestrong model's output is updated by subtracting the predicted gradient,similar togradient descent.

where:

• $F_{0}$ is the starting strong model.
• $F_{i+1}$ is the next strong model.
• $F_{i}$ is the current strong model.
• $\xi$ is a value between 0.0 and 1.0 calledshrinkage, which is analogous to thelearning rate in gradient descent.
• $f_{i}$ is the weak model trained to predict the loss gradient of $F_{i}$.

Modern variations of gradient boosting also include the second derivative(Hessian) of the loss in their computation.

Decision trees are commonly used as weak models ingradient boosting. Seegradient boosted (decision) trees.

gradient clipping

A commonly used mechanism to mitigate theexploding gradient problem by artificiallylimiting (clipping) the maximum value of gradients when usinggradient descent totrain a model.

gradient descent

A mathematical technique to minimizeloss.Gradient descent iteratively adjustsweights andbiases,gradually finding the best combination to minimize loss.

Gradient descent is older—much, much older—than machine learning.

See theLinear regression: Gradientdescentin Machine Learning Crash Course for more information.

graph

In TensorFlow, a computation specification. Nodes in the graphrepresent operations. Edges are directed and represent passing the resultof an operation (aTensor) as anoperand to another operation. UseTensorBoard to visualize a graph.

graph execution

A TensorFlow programming environment in which the program first constructsagraph and then executes all or part of that graph. Graphexecution is the default execution mode in TensorFlow 1.x.

Contrast witheager execution.

greedy policy

In reinforcement learning, apolicy that always chooses theaction with the highest expectedreturn.

groundedness

Note that a grounded model is not always afactual model.For example, the input physics textbook could contain mistakes.

ground truth

Reality.

The thing that actually happened.

For example, consider abinary classificationmodel that predicts whether a student in their first year of universitywill graduate within six years. Ground truth for this model is whether ornot that student actually graduated within six years.

group attribution bias

Assuming that what is true for an individual is also true for everyonein that group. The effects of group attribution bias can be exacerbatedif aconvenience samplingis used for data collection. In a non-representative sample, attributionsmay be made that don't reflect reality.

See alsoout-group homogeneity biasandin-group bias. Also, seeFairness: Types of biasin Machine Learning Crash Course for more information.

H

hallucination

The production of plausible-seeming but factually incorrect output by agenerative AI model that purports to be making anassertion about the real world.For example, a generative AI model that claims that Barack Obama died in 1865ishallucinating.

hashing

In machine learning, a mechanism for bucketingcategorical data, particularly when the numberof categories is large, but the number of categories actually appearingin the dataset is comparatively small.

For example, Earth is home to about 73,000 tree species. You couldrepresent each of the 73,000 tree species in 73,000 separate categoricalbuckets. Alternatively, if only 200 of those tree species actually appearin a dataset, you could use hashing to divide tree species intoperhaps 500 buckets.

A single bucket could contain multiple tree species. For example, hashingcould placebaobab andred maple—two genetically dissimilarspecies—into the same bucket. Regardless, hashing is still a good way tomap large categorical sets into the selected number of buckets. Hashing turns acategorical feature having a large number of possible values into a muchsmaller number of values by grouping values in adeterministic way.

SeeCategorical data: Vocabulary and one-hotencodingin Machine Learning Crash Course for more information.

heuristic

A simple and quickly implemented solution to a problem. For example,"With a heuristic, we achieved 86% accuracy. When we switched to adeep neural network, accuracy went up to 98%."

hidden layer

A layer in aneural network between theinput layer (the features) and theoutput layer (the prediction).Each hidden layer consists of one or moreneurons.For example, the following neural network contains two hidden layers,the first with three neurons and the second with two neurons:

Adeep neural network contains more than onehidden layer. For example, the preceding illustration is a deep neuralnetwork because the model contains two hidden layers.

SeeNeural networks: Nodes and hiddenlayersin Machine Learning Crash Course for more information.

hierarchical clustering

A category ofclustering algorithms that create a treeof clusters. Hierarchical clustering is well-suited to hierarchical data,such as botanical taxonomies. There are two types of hierarchicalclustering algorithms:

• Agglomerative clustering first assigns every example to its own cluster,and iteratively merges the closest clusters to create a hierarchicaltree.
• Divisive clustering first groups all examples into one cluster and theniteratively divides the cluster into a hierarchical tree.

Contrast withcentroid-based clustering.

SeeClusteringalgorithmsin the Clustering course for more information.

hill climbing

An algorithm for iteratively improving ("walking uphill") an ML model untilthe model stops improving ("reaches the top of a hill"). The general formof the algorithm is as follows:

1. Build a starting model.
2. Create new candidate models by making small adjustments to the way youtrain orfine-tune. This might entailworking with a slightly differenttraining set ordifferent hyperparameters.
3. Evaluate the new candidate models and take one of thefollowing actions:
   • If a candidate model outperforms the starting model, then that candidatemodel becomes the new starting model. In this case, repeat Steps 1, 2,and 3.
   • If no model outperforms the starting model, then you've reached the topof the hill and should stop iterating.

SeeDeep Learning Tuning Playbookfor guidance on hyperparameter tuning. See the Data modules ofMachine Learning Crash Course for guidanceon feature engineering.

hinge loss

A family ofloss functions forclassification designed to find thedecision boundary as distant as possiblefrom each training example,thus maximizing the margin between examples and the boundary.KSVMs use hinge loss (or a related function, such assquared hinge loss). For binary classification, the hinge loss functionis defined as follows:

wherey is the true label, either -1 or +1, andy' is the raw outputof theclassification model:

Consequently, a plot of hinge loss versus (y * y') looks as follows:

historical bias

A type ofbias that already exists in the world and hasmade its way into a dataset. These biases have a tendency to reflect existingcultural stereotypes, demographic inequalities, and prejudices against certainsocial groups.

For example, consider aclassification model thatpredicts whether or not a loan applicant will default on their loan, which wastrained on historical loan-default data from the 1980s from local banks in twodifferent communities. If past applicants from Community A were six times morelikely to default on their loans than applicants from Community B, the modelmight learn a historical bias resulting in the model being less likely toapprove loans in Community A, even if the historical conditions that resultedin that community's higher default rates were no longer relevant.

SeeFairness: Types ofbiasin Machine Learning Crash Course for more information.

holdout data

Examples intentionally not used ("held out") during training.Thevalidation dataset andtest dataset are examples of holdout data. Holdout datahelps evaluate your model's ability to generalize to data other than thedata it was trained on. The loss on the holdout set provides a betterestimate of the loss on an unseen dataset than does the loss on thetraining set.

host

When training an ML model onaccelerator chips(GPUs orTPUs), the part of the systemthat controls both of the following:

• The overall flow of the code.
• The extraction and transformation of the input pipeline.

The host typically runs on a CPU, not on an accelerator chip; thedevice manipulatestensors on theaccelerator chips.

human evaluation

A process in whichpeople judge the quality of an ML model's output;for example, having bilingual people judge the quality of an ML translationmodel. Human evaluation is particularly useful for judging models that haveno one right answer.

Contrast withautomatic evaluation andautorater evaluation.

human in the loop (HITL)

A loosely-defined idiom that could mean either of the following:

• A policy of viewinggenerative AI output critically orskeptically.
• A strategy or system for ensuring that people help shape, evaluate, and refinea model's behavior. Keeping a human in the loop enables an AI to benefit fromboth machine intelligence and human intelligence. For example, a system inwhich an AI generates code which software engineers then review is ahuman-in-the-loop system.

hyperparameter

The variables that you or a hyperparameter tuning serviceadjust during successive runs of training a model. For example,learning rate is a hyperparameter. You couldset the learning rate to 0.01 before one training session. If youdetermine that 0.01 is too high, you could perhaps set the learningrate to 0.003 for the next training session.

In contrast,parameters are the variousweights andbias that the modellearns during training.

SeeLinear regression:Hyperparametersin Machine Learning Crash Course for more information.

hyperplane

A boundary that separates a space into two subspaces. For example, a line is ahyperplane in two dimensions and a plane is a hyperplane in three dimensions.More typically in machine learning, a hyperplane is the boundary separating ahigh-dimensional space.Kernel Support Vector Machines usehyperplanes to separate positive classes from negative classes, often in a veryhigh-dimensional space.

I

i.i.d.

Abbreviation forindependently and identically distributed.

image recognition

A process that classifies object(s), pattern(s), or concept(s) in an image.Image recognition is also known asimage classification.

imbalanced dataset

Synonym forclass-imbalanced dataset.

implicit bias

Automatically making an association or assumption based on one's mindmodels and memories. Implicit bias can affect the following:

• How data is collected and classified.
• How machine learning systems are designed and developed.

For example, when building aclassification modelto identify wedding photos, an engineer may use the presence of a white dressin a photo as a feature. However, white dresses have been customary only duringcertain eras and in certain cultures.

See alsoconfirmation bias.

imputation

Short form ofvalue imputation.

incompatibility of fairness metrics

The idea that some notions of fairness are mutually incompatible andcannot be satisfied simultaneously. As a result, there is no singleuniversalmetric for quantifying fairnessthat can be applied to all ML problems.

While this may seem discouraging, incompatibility of fairness metricsdoesn't imply that fairness efforts are fruitless. Instead, it suggeststhat fairness must be defined contextually for a given ML problem, withthe goal of preventing harms specific to its use cases.

See"On the(im)possibility of fairness" for a more detailed discussion of theincompatibility of fairness metrics.

in-context learning

Synonym forfew-shot prompting.

independently and identically distributed (i.i.d)

Data drawn from a distribution that doesn't change, and where each valuedrawn doesn't depend on values that have been drawn previously. An i.i.d.is theideal gasof machinelearning—a useful mathematical construct but almost never exactly foundin the real world. For example, the distribution of visitors to a web pagemay be i.i.d. over a brief window of time; that is, the distribution doesn'tchange during that brief window and one person's visit is generallyindependent of another's visit. However, if you expand that window of time,seasonal differences in the web page's visitors may appear.

See alsononstationarity.

individual fairness

A fairness metric that checks whether similar individuals are classifiedsimilarly. For example, Brobdingnagian Academy might want to satisfyindividual fairness by ensuring that two students with identical gradesand standardized test scores are equally likely to gain admission.

Note that individual fairness relies entirely on how you define "similarity"(in this case, grades and test scores), and you can run the risk ofintroducing new fairness problems if your similarity metric misses importantinformation (such as the rigor of a student's curriculum).

See"Fairness ThroughAwareness" for a more detailed discussion of individual fairness.

inference

In traditional machine learning, the process of making predictions byapplying a trained model tounlabeled examples.SeeSupervised Learningin the Intro to ML course to learn more.

Inlarge language models, inference is theprocess of using a trained model to generate aresponseto an inputprompt.

Inference has a somewhat different meaning in statistics. See theWikipedia article on statistical inference for details.

inference path

In adecision tree, duringinference,the route a particularexample takes from theroot to otherconditions, terminating withaleaf. For example, in the following decision tree, thethicker arrows show the inference path for an example with the followingfeature values:

• x = 7
• y = 12
• z = -3

The inference path in the following illustration travels through threeconditions before reaching the leaf (Zeta).

The three thick arrows show the inference path.

SeeDecision treesin the Decision Forests course for more information.

information gain

Indecision forests, the difference betweena node'sentropy and the weighted (by number of examples)sum of the entropy of its children nodes. A node's entropy is the entropyof the examples in that node.

For example, consider the following entropy values:

• entropy of parent node = 0.6
• entropy of one child node with 16 relevant examples = 0.2
• entropy of another child node with 24 relevant examples = 0.1

So 40% of the examples are in one child node and 60% are in theother child node. Therefore:

• weighted entropy sum of child nodes = (0.4 * 0.2) + (0.6 * 0.1) = 0.14

So, the information gain is:

• information gain = entropy of parent node - weighted entropy sum of child nodes
• information gain = 0.6 - 0.14 = 0.46

Mostsplitters seek to createconditionsthat maximize information gain.

in-group bias

Showing partiality to one's own group or own characteristics.If testers or raters consist of the machine learning developer's friends,family, or colleagues, then in-group bias may invalidate product testingor the dataset.

In-group bias is a form ofgroup attribution bias.See alsoout-group homogeneity bias.

SeeFairness: Types ofbias inMachine Learning Crash Course for more information.

input generator

A mechanism by which data is loaded intoaneural network.

An input generator can be thought of as a component responsible for processingraw data into tensors which are iterated over to generate batches fortraining, evaluation, and inference.

input layer

Thelayer of aneural network thatholds thefeature vector. That is, the input layerprovidesexamples fortraining orinference. For example, the input layer in the followingneural network consists of two features:

in-set condition

In adecision tree, aconditionthat tests for the presence of one item in a set of items.For example, the following is an in-set condition:

  house-style in [tudor, colonial, cape]

During inference, if the value of the house-stylefeatureistudor orcolonial orcape, then this condition evaluates to Yes. Ifthe value of the house-style feature is something else (for example,ranch),then this condition evaluates to No.

In-set conditions usually lead to more efficient decision trees thanconditions that testone-hot encoded features.

instance

Synonym forexample.

instruction tuning

A form offine-tuning that improves agenerative AI model's ability to followinstructions. Instruction tuning involves training a model on a seriesof instruction prompts, typically covering a widevariety of tasks. The resulting instruction-tuned model then tends togenerate usefulresponses tozero-shot prompts across a variety of tasks.

Compare and contrast with:

• parameter-efficient tuning
• prompt tuning

interpretability

The ability to explain or to present an MLmodel's reasoning inunderstandable terms to a human.

Mostlinear regression models, for example, are highlyinterpretable. (You merely need to look at the trained weights for eachfeature.) Decision forests are also highly interpretable. Some models, however,require sophisticated visualization to become interpretable.

You can use theLearning Interpretability Tool (LIT)to interpret ML models.

inter-rater agreement

A measurement of how often human raters agree when doing a task.If raters disagree, the task instructions may need to be improved.Also sometimes calledinter-annotator agreement orinter-rater reliability. See alsoCohen'skappa,which is one of the most popular inter-rater agreement measurements.

SeeCategorical data: Commonissuesin Machine Learning Crash Course for more information.

intersection over union (IoU)

The intersection of two sets divided by their union. In machine-learningimage-detection tasks, IoU is used to measure the accuracy of the model'spredictedbounding box with respect to theground-truth bounding box. In this case, the IoU for thetwo boxes is the ratio between the overlapping area and the total area, andits value ranges from 0 (no overlap of predicted bounding box and ground-truthbounding box) to 1 (predicted bounding box and ground-truth bounding box havethe exact same coordinates).

For example, in the image below:

• The predicted bounding box (the coordinates delimiting where the modelpredicts the night table in the painting is located) is outlined in purple.
• The ground-truth bounding box (the coordinates delimiting where the nighttable in the painting is actually located) is outlined in green.

Here, the intersection of the bounding boxes for prediction and ground truth(below left) is 1, and the union of the bounding boxes for prediction andground truth (below right) is 7, so the IoU is \(\frac{1}{7}\).

IoU

Abbreviation forintersection over union.

item matrix

Inrecommendation systems, amatrix ofembedding vectors generated bymatrix factorizationthat holds latent signals about eachitem.Each row of the item matrix holds the value of a single latentfeature for all items.For example, consider a movie recommendation system. Each columnin the item matrix represents a single movie. The latent signalsmight represent genres, or might be harder-to-interpretsignals that involve complex interactions among genre, stars,movie age, or other factors.

The item matrix has the same number of columns as the targetmatrix that is being factorized. For example, given a movierecommendation system that evaluates 10,000 movie titles, theitem matrix will have 10,000 columns.

items

In arecommendation system, the entities thata system recommends. For example, videos are the items that a video storerecommends, while books are the items that a bookstore recommends.

iteration

A single update of amodel's parameters—the model'sweights andbiases—duringtraining. Thebatch size determineshow many examples the model processes in a single iteration. For instance,if the batch size is 20, then the model processes 20 examples beforeadjusting the parameters.

When training aneural network, a single iterationinvolves the following two passes:

1. A forward pass to evaluate loss on a single batch.
2. A backward pass (backpropagation) to adjust themodel's parameters based on the loss and the learning rate.

SeeGradientdescentin Machine Learning Crash Course for more information.

J

JAX

An array computing library, bringing togetherXLA (Accelerated Linear Algebra) and automatic differentiationfor high-performance numerical computing. JAX provides a simple and powerfulAPI for writing accelerated numerical code with composable transformations.JAX provides features such as:

• grad (automatic differentiation)
• jit (just-in-time compilation)
• vmap (automatic vectorization or batching)
• pmap (parallelization)

JAX is a language for expressing and composing transformations of numericalcode, analogous—but much larger in scope—to Python'sNumPylibrary. (In fact, the .numpy library under JAX is a functionally equivalent,but entirely rewritten version of the Python NumPy library.)

JAX is particularly well-suited for speeding up many machine learning tasksby transforming the models and data into a form suitable for parallelismacross GPU andTPUaccelerator chips.

Flax,Optax,Pax, and many otherlibraries are built on the JAX infrastructure.

K

Keras

A popular Python machine learning API.Kerasruns onseveral deep learning frameworks, including TensorFlow, where it is madeavailable astf.keras.

Kernel Support Vector Machines (KSVMs)

A classification algorithm that seeks to maximize the margin betweenpositive andnegative classes by mapping input data vectorsto a higher dimensional space. For example, consider a classificationproblem in which the input datasethas a hundred features. To maximize the margin betweenpositive and negative classes, a KSVM could internally map those features intoa million-dimension space. KSVMs uses a loss function calledhinge loss.

keypoints

The coordinates of particular features in an image. For example, for animage recognition model that distinguishesflower species, keypoints might be the center of each petal, the stem,the stamen, and so on.

k-fold cross validation

An algorithm for predicting a model's ability togeneralize to new data. Thek in k-fold refers to thenumber of equal groups you divide a dataset's examples into; that is, you trainand test your model k times. For each round of training and testing, adifferent group is the test set, and all remaining groups become the trainingset. After k rounds of training and testing, you calculate the mean andstandard deviation of the chosen test metric(s).

For example, suppose your dataset consists of 120 examples. Further suppose,you decide to set k to 4. Therefore, after shuffling the examples,you divide the dataset into four equal groups of 30 examples and conduct fourtraining and testing rounds:

For example,Mean Squared Error (MSE) mightbe the most meaningful metric for a linear regression model. Therefore, youwould find the mean and standard deviation of the MSE across all four rounds.

k-means

A popularclustering algorithm that groups examplesin unsupervised learning. The k-means algorithm basically does the following:

• Iteratively determines the best k center points (knownascentroids).
• Assigns each example to the closest centroid. Those examples nearestthe same centroid belong to the same group.

The k-means algorithm picks centroid locations to minimize the cumulativesquare of the distances from each example to its closest centroid.

For example, consider the following plot of dog height to dog width:

If k=3, the k-means algorithm will determine three centroids. Each exampleis assigned to its closest centroid, yielding three groups:

Imagine that a manufacturer wants to determine the ideal sizes for small,medium, and large sweaters for dogs. The three centroids identify the meanheight and mean width of each dog in that cluster. So, the manufacturershould probably base sweater sizes on those three centroids. Note thatthe centroid of a cluster is typicallynot an example in the cluster.

The preceding illustrations shows k-means for examples with onlytwo features (height and width). Note that k-means can group examplesacross many features.

SeeWhat is k-means clustering?in the Clustering course for more information.

k-median

A clustering algorithm closely related tok-means. Thepractical difference between the two is as follows:

• In k-means, centroids are determined by minimizing the sum of thesquares of the distance between a centroid candidate and each ofits examples.
• In k-median, centroids are determined by minimizing the sum of thedistance between a centroid candidate and each of its examples.

Note that the definitions of distance are also different:

• k-means relies on theEuclidean distance fromthe centroid to an example. (In two dimensions, the Euclideandistance means using the Pythagorean theorem to calculatethe hypotenuse.) For example, the k-means distance between (2,2)and (5,-2) would be:

• k-median relies on the Manhattan distancefrom the centroid to an example. This distance is the sum of theabsolute deltas in each dimension. For example, the k-mediandistance between (2,2) and (5,-2) would be:

L

L₀ regularization

A type ofregularization thatpenalizes thetotal number of nonzeroweightsin a model. For example, a model having 11 nonzero weightswould be penalized more than a similar model having 10 nonzero weights.

L₀ regularization is sometimes calledL0-norm regularization.

L₁ loss

Aloss function that calculates the absolute valueof the difference between actuallabel values andthe values that amodel predicts. For example, here's thecalculation of L₁ loss for abatch of fiveexamples:

L₁ loss is less sensitive tooutliersthanL₂ loss.

TheMean Absolute Error is the averageL₁ loss per example.

SeeLinear regression: Lossin Machine Learning Crash Course for more information.

L₁ regularization

A type ofregularization that penalizesweights in proportion to the sum of the absolute value ofthe weights. L₁ regularization helps drive the weights of irrelevantor barely relevant features toexactly 0. Afeature witha weight of 0 is effectively removed from the model.

Contrast withL₂ regularization.

L₂ loss

Aloss function that calculates the squareof the difference between actuallabel values andthe values that amodel predicts. For example, here's thecalculation of L₂ loss for abatch of fiveexamples:

Due to squaring, L₂ loss amplifies the influence ofoutliers.That is, L₂ loss reacts more strongly to bad predictions thanL₁ loss. For example, the L₁ lossfor the preceding batch would be 8 rather than 16. Notice that a singleoutlier accounts for 9 of the 16.

Regression models typically use L₂ lossas the loss function.

TheMean Squared Error is the averageL₂ loss per example.Squared loss is another name for L₂ loss.

SeeLogistic regression: Loss andregularizationin Machine Learning Crash Course for more information.

L₂ regularization

A type ofregularization that penalizesweights in proportion to the sum of thesquares of the weights.L₂ regularization helps driveoutlier weights (thosewith high positive or low negative values) closer to 0 butnot quite to 0.Features with values very close to 0 remain in the modelbut don't influence the model's prediction very much.

L₂ regularization always improves generalization inlinear models.

Contrast withL₁ regularization.

SeeOverfitting: L2regularizationin Machine Learning Crash Course for more information.

label

Insupervised machine learning, the"answer" or "result" portion of anexample.

Eachlabeled example consists of one or morefeatures and a label. For example, in a spamdetection dataset, the label would probably be either "spam" or"not spam." In a rainfall dataset, the label might be the amount ofrain that fell during a certain period.

SeeSupervised Learningin Introduction to Machine Learning for more information.

labeled example

An example that contains one or morefeatures and alabel. For example, the following table shows threelabeled examples from a house valuation model, each with three featuresand one label:

Insupervised machine learning,models train on labeled examples and make predictions onunlabeled examples.

Contrast labeled example with unlabeled examples.

SeeSupervised Learningin Introduction to Machine Learning for more information.

label leakage

A model design flaw in which afeature is a proxy for thelabel. For example, consider abinary classification model that predictswhether or not a prospective customer will purchase a particular product.Suppose that one of the features for the model is a Boolean namedSpokeToCustomerAgent. Further suppose that a customer agent is onlyassignedafter the prospective customer has actually purchased theproduct. During training, the model will quickly learn the associationbetweenSpokeToCustomerAgent and the label.

SeeMonitoringpipelinesin Machine Learning Crash Course for more information.

lambda

Synonym forregularization rate.

Lambda is an overloaded term. Here we're focusing on the term'sdefinition withinregularization.

LaMDA (Language Model for Dialogue Applications)

ATransformer-basedlarge language model developed by Google trained ona large dialogue dataset that can generate realistic conversationalresponses.

LaMDA: our breakthrough conversationtechnology provides an overview.

landmarks

Synonym forkeypoints.

language model

Amodel that estimates the probability of atokenor sequence of tokens occurring in a longer sequence of tokens.

SeeWhat is a language model?in Machine Learning Crash Course for more information.

large language model

At a minimum, alanguage model having a very high numberofparameters. More informally, anyTransformer-based language model, such asGemini orGPT.

SeeLarge language models (LLMs)in Machine Learning Crash Course for more information.

latency

The time it takes for a model to process input and generate a response.Ahigh latency response takes takes longer to generate than alow latency response.

Factors that influence latency oflarge language models include:

• Input and outputtoken lengths
• Model complexity
• The infrastructure the model runs on

Optimizing for latency is crucial for creating responsive and user-friendlyapplications.

latent space

Synonym forembedding space.

layer

A set ofneurons in aneural network. Three common types of layersare as follows:

• Theinput layer, which provides values for all thefeatures.
• One or morehidden layers, which findnonlinear relationships between the features and the label.
• Theoutput layer, which provides the prediction.

For example, the following illustration shows a neural network withone input layer, two hidden layers, and one output layer:

InTensorFlow,layers are also Python functions that takeTensors and configuration options as input andproduce other tensors as output.

Layers API (tf.layers)

A TensorFlow API for constructing adeep neural networkas a composition of layers. The Layers API lets you build differenttypes oflayers, such as:

• tf.layers.Dense for afully-connected layer.
• tf.layers.Conv2D for a convolutional layer.

The Layers API follows theKeras layers API conventions.That is, aside from a different prefix, all functions in the Layers APIhave the same names and signatures as their counterparts in the Keraslayers API.

leaf

Any endpoint in adecision tree. Unlike acondition, a leaf doesn't perform a test.Rather, a leaf is a possible prediction. A leaf is also the terminalnode of aninference path.

For example, the following decision tree contains three leaves:

SeeDecision treesin the Decision Forests course for more information.

Learning Interpretability Tool (LIT)

A visual, interactive model-understanding and data visualization tool.

You can use open-sourceLIT tointerpret models or to visualize text, image, andtabular data.

learning rate

A floating-point number that tells thegradient descentalgorithm how strongly to adjust weights and biases on eachiteration. For example, a learning rate of 0.3 wouldadjust weights and biases three times more powerfully than a learning rateof 0.1.

Learning rate is a keyhyperparameter. If you setthe learning rate too low, training will take too long. Ifyou set the learning rate too high, gradient descent often has troublereachingconvergence.

SeeLinear regression:Hyperparametersin Machine Learning Crash Course for more information.

least squares regression

Alinear regression model trained by minimizingL₂ Loss.

Levenshtein Distance

Anedit distance metric that calculates the fewest delete,insert, and substitute operations required to change one word to another.For example, the Levenshtein distance between the words "heart" and "darts"is three because the following three edits are the fewest changes to turnone word into the other:

1. heart → deart (substitute "h" with "d")
2. deart → dart (delete "e")
3. dart → darts (insert "s")

Note that the preceding sequence isn't the only path of three edits.

linear

A relationship between two or more variables that can be represented solelythrough addition and multiplication.

The plot of a linear relationship is a line.

Contrast withnonlinear.

linear model

Amodel that assigns oneweight perfeature to makepredictions.(Linear models also incorporate abias.) In contrast,the relationship of features to predictions indeep modelsis generallynonlinear.

Linear models are usually easier to train and moreinterpretable than deep models. However,deep models can learn complex relationshipsbetween features.

Linear regression andlogistic regression are two types of linear models.

y' = 7 + (-2.5)(x1) + (-1.2)(x2) + (1.4)(x3)

y' = 7 + (-2.5)(4) + (-1.2)(-10) + (1.4)(5)y' = 16

linear regression

A type of machine learning model in which both of the following are true:

• The model is alinear model.
• The prediction is a floating-point value. (This is theregression part oflinear regression.)

Contrast linear regression withlogistic regression.Also, contrast regression withclassification.

SeeLinear regressionin Machine Learning Crash Course for more information.

LIT

Abbreviation for theLearning Interpretability Tool (LIT),which was previously known as the Language Interpretability Tool.

LLM

Abbreviation forlarge language model.

LLM evaluations (evals)

A set of metrics and benchmarks for assessing the performance oflarge language models (LLMs). At a high level,LLM evaluations:

• Help researchers identify areas where LLMs need improvement.
• Are useful in comparing different LLMs and identifying the best LLM for aparticular task.
• Help ensure that LLMs are safe and ethical to use.

SeeLarge languagemodels (LLMs)in Machine Learning Crash Course for more information.

logistic regression

A type ofregression model that predicts a probability.Logistic regression models have the following characteristics:

• The label iscategorical. The term logisticregression usually refers tobinary logistic regression, that is,to a model that calculates probabilities for labels with two possible values.A less common variant,multinomial logistic regression, calculatesprobabilities for labels with more than two possible values.
• The loss function during training isLog Loss.(Multiple Log Loss units can be placed in parallel for labelswith more than two possible values.)
• The model has a linear architecture, not a deep neural network.However, the remainder of this definition also applies todeep models that predict probabilitiesfor categorical labels.

For example, consider a logistic regression model that calculates theprobability of an input email being either spam or not spam.During inference, suppose the model predicts 0.72. Therefore, themodel is estimating:

• A 72% chance of the email being spam.
• A 28% chance of the email not being spam.

A logistic regression model uses the following two-step architecture:

1. The model generates a raw prediction (y') by applying a linear functionof input features.
2. The model uses that raw prediction as input to asigmoid function, which converts the rawprediction to a value between 0 and 1, exclusive.

Like any regression model, a logistic regression model predicts a number.However, this number typically becomes part of a binary classificationmodel as follows:

• If the predicted number isgreater than theclassification threshold, thebinary classification model predicts the positive class.
• If the predicted number isless than the classification threshold,the binary classification model predicts the negative class.

SeeLogistic regressionin Machine Learning Crash Course for more information.

logits

The vector of raw (non-normalized) predictions that a classificationmodel generates, which is ordinarily then passed to a normalization function.If the model is solving amulti-class classificationproblem, logits typically become an input to thesoftmax function.The softmax function then generates a vector of (normalized)probabilities with one value for each possible class.

Log Loss

Theloss function used in binarylogistic regression.

SeeLogistic regression: Loss and regularizationin Machine Learning Crash Course for more information.

log-odds

The logarithm of the odds of some event.

Long Short-Term Memory (LSTM)

A type of cell in arecurrent neural network used to processsequences of data in applications such as handwriting recognition,machine translation, and image captioning. LSTMsaddress thevanishing gradient problem thatoccurs when training RNNs due to long data sequences by maintaining history inan internal memory state based on new input and context from previous cells inthe RNN.

LoRA

Abbreviation forLow-Rank Adaptability.

loss

During thetraining of asupervised model, a measure of how far amodel'sprediction is from itslabel.

Aloss function calculates the loss.

SeeLinear regression: Lossin Machine Learning Crash Course for more information.

loss aggregator

A type ofmachine learning algorithm thatimproves theperformance of amodelby combining thepredictions of multiple models andusing those predictions to make a single prediction. As a result,a loss aggregator can reduce the variance of the predictions andimprove theaccuracy of the predictions.

loss curve

A plot ofloss as a function of the number of trainingiterations. The following plot shows a typical losscurve:

Loss curves can help you determine when your model isconverging oroverfitting.

Loss curves can plot all of the following types of loss:

• training loss
• validation loss
• test loss

See alsogeneralization curve.

SeeOverfitting: Interpreting loss curvesin Machine Learning Crash Course for more information.

loss function

Duringtraining or testing, amathematical function that calculates theloss on abatch of examples. A loss function returns a lower lossfor models that makes good predictions than for models that makebad predictions.

The goal of training is typically to minimize the loss that a loss functionreturns.

Many different kinds of loss functions exist. Pick the appropriate lossfunction for the kind of model you are building. For example:

• L₂ loss (orMean Squared Error)is the loss function forlinear regression.
• Log Loss is the loss function forlogistic regression.

loss surface

A graph of weight(s) versus loss.Gradient descent aimsto find the weight(s) for which the loss surface is at a local minimum.

lost-in-the-middle effect

AnLLM's tendency to use information from the start and end of alongcontext window more effectively than informationfrom the middle. That is, given a long context, the lost-in-the-middle effectcauses accuracy to be:

• Relatively high when the relevant information to form aresponse is near thebeginning orend of the context.
• Relatively low when the relevant information to form aresponse is in themiddle of the context.

The term comes fromLost in the Middle: How Language Models Use Long Contexts.

Low-Rank Adaptability (LoRA)

Aparameter-efficient technique forfine tuning that "freezes" the model's pre-trainedweights (such that they can no longer be modified) and then inserts a small setof trainable weights into the model. This set of trainable weights (also knownas "update matrixes") is considerably smaller than the base model and istherefore much faster to train.

LoRA provides the following benefits:

• Improves the quality of a model's predictions for the domain where the finetuning is applied.
• Fine-tunes faster than techniques that require fine-tuningall of a model'sparameters.
• Reduces the computational cost ofinference by enablingconcurrent serving of multiple specialized models sharing the same basemodel.

LSTM

Abbreviation forLong Short-Term Memory.

M

machine learning

A program or system thattrains amodel from input data. The trained model canmake useful predictions from new (never-before-seen) data drawn fromthe same distribution as the one used to train the model.

Machine learning also refers to the field of study concernedwith these programs or systems.

See theIntroduction to Machine Learningcourse for more information.

machine translation

Using software (typically, a machine learning model) to convert text fromone human language to another human language, for example, from English toJapanese.

majority class

The more common label in aclass-imbalanced dataset. For example,given a dataset containing 99% negative labels and 1% positive labels, thenegative labels are the majority class.

Contrast withminority class.

SeeDatasets: Imbalanced datasetsin Machine Learning Crash Course for more information.

Markov decision process (MDP)

A graph representing the decision-making model where decisions(oractions) are taken to navigate a sequence ofstates under the assumption that theMarkov property holds. Inreinforcement learning, these transitionsbetween states return a numericalreward.

Markov property

A property of certainenvironments, where statetransitions are entirely determined by information implicit in thecurrentstate and the agent'saction.

masked language model

Alanguage model that predicts the probability ofcandidate tokens to fill in blanks in a sequence. For example, amasked language model can calculate probabilities for candidate word(s)to replace the underline in the following sentence:

The ____ in the hat came back.

The literature typically uses the string "MASK" instead of an underline.For example:

The "MASK" in the hat came back.

Most modern masked language models arebidirectional.

math-pass@k

A metric to determine an LLM's accuracy in solving a math problem withinK attempts. For example, math-pass@2 measures an LLM's ability to solve mathproblems within two attempts. An accuracy of 0.85 on math-pass@2 indicates thatan LLM was able to solve math problems 85% of the time within two attempts.

math-pass@k is identical to thepass@k metric, except thatthe term math-pass@k is specifically used for math evaluation.

matplotlib

An open-source Python 2D plotting library.matplotlib helps you visualizedifferent aspects of machine learning.

matrix factorization

In math, a mechanism for finding the matrixes whose dot product approximates atarget matrix.

Inrecommendation systems, the target matrixoften holds users' ratings onitems. For example, the targetmatrix for a movie recommendation system might look something like thefollowing, where the positive integers are user ratings and 0means that the user didn't rate the movie:

The movie recommendation system aims to predict user ratings forunrated movies. For example, will User 1 likeBlack Panther?

One approach for recommendation systems is to use matrixfactorization to generate the following two matrixes:

• Auser matrix, shaped as the number of users X thenumber of embedding dimensions.
• Anitem matrix, shaped as the number of embeddingdimensions X the number of items.

For example, using matrix factorization on our three users and five itemscould yield the following user matrix and item matrix:

User Matrix                 Item Matrix1.1   2.3           0.9   0.2   1.4    2.0   1.20.6   2.0           1.7   1.2   1.2   -0.1   2.12.5   0.5

The dot product of the user matrix and item matrix yields a recommendationmatrix that contains not only the original user ratings but also predictionsfor the movies that each user hasn't seen.For example, consider User 1's rating ofCasablanca, which was 5.0. The dotproduct corresponding to that cell in the recommendation matrix shouldhopefully be around 5.0, and it is:

(1.1* 0.9) + (2.3 * 1.7) = 4.9

More importantly, will User 1 likeBlack Panther? Taking the dot productcorresponding to the first row and the third column yields a predictedrating of 4.3:

(1.1* 1.4) + (2.3 * 1.2) = 4.3

Matrix factorization typically yields a user matrix and item matrix that,together, are significantly more compact than the target matrix.

MBPP

Abbreviation forMostly Basic Python Problems.

Mean Absolute Error (MAE)

The average loss per example whenL₁ loss isused. Calculate Mean Absolute Error as follows:

1. Calculate the L₁ loss for a batch.
2. Divide the L₁ loss by the number of examples in the batch.

For example, consider the calculation of L₁ loss on thefollowing batch of five examples:

So, L₁ loss is 8 and the number of examples is 5.Therefore, the Mean Absolute Error is:

Mean Absolute Error = L1 loss / Number of ExamplesMean Absolute Error = 8/5 = 1.6

Contrast Mean Absolute Error withMean Squared Error andRoot Mean Squared Error.

mean average precision at k (mAP@k)

The statistical mean of allaverage precision at k scores acrossa validation dataset. One use of mean average precision at k is to judgethe quality of recommendations generated by arecommendation system.

Although the phrase "mean average" sounds redundant, the name ofthe metric is appropriate. After all, this metric finds the meanof multipleaverage precision at k values.

Mean Squared Error (MSE)

The average loss per example whenL₂ loss isused. Calculate Mean Squared Error as follows:

1. Calculate the L₂ loss for a batch.
2. Divide the L₂ loss by the number of examples in the batch.

For example, consider the loss on the following batch of five examples:

Therefore, the Mean Squared Error is:

Mean Squared Error = L2 loss / Number of ExamplesMean Squared Error = 16/5 = 3.2

Mean Squared Error is a popular trainingoptimizer,particularly forlinear regression.

Contrast Mean Squared Error withMean Absolute Error andRoot Mean Squared Error.

TensorFlow Playground uses Mean Squared Errorto calculate loss values.

mesh

In ML parallel programming, a term associated with assigning the data andmodel to TPU chips, and defining how these values will be sharded or replicated.

Mesh is an overloaded term that can mean either of the following:

• A physical layout of TPU chips.
• An abstract logical construct for mapping the data and model to the TPUchips.

In either case, a mesh is specified as ashape.

meta-learning

A subset of machine learning that discovers or improves a learning algorithm.A meta-learning system can also aim to train a model to quickly learn a newtask from a small amount of data or from experience gained in previous tasks.Meta-learning algorithms generally try to achieve the following:

• Improve or learn hand-engineered features (such as an initializer oran optimizer).
• Be more data-efficient and compute-efficient.
• Improve generalization.

Meta-learning is related tofew-shot learning.

metric

A statistic that you care about.

Anobjective is a metric that a machine learning systemtries to optimize.

Metrics API (tf.metrics)

A TensorFlow API for evaluating models. For example,tf.metrics.accuracydetermines how often a model's predictions match labels.

mini-batch

A small, randomly selected subset of abatch processed in oneiteration.Thebatch size of a mini-batch is usuallybetween 10 and 1,000 examples.

For example, suppose the entire training set (the full batch)consists of 1,000 examples. Further suppose that you set thebatch size of each mini-batch to 20. Therefore, eachiteration determines the loss on a random 20 of the 1,000 examples and thenadjusts theweights andbiases accordingly.

It is much more efficient to calculate the loss on a mini-batch than theloss on all the examples in the full batch.

SeeLinear regression:Hyperparametersin Machine Learning Crash Course for more information.

mini-batch stochastic gradient descent

Agradient descent algorithm that usesmini-batches. In other words, mini-batch stochasticgradient descent estimates the gradient based on a small subset of thetraining data. Regularstochastic gradient descent uses amini-batch of size 1.

minimax loss

A loss function forgenerative adversarial networks,based on thecross-entropy between the distributionof generated data and real data.

Minimax loss is used in thefirst paper to describegenerative adversarial networks.

SeeLoss Functions in theGenerative Adversarial Networks course for more information.

minority class

The less common label in aclass-imbalanced dataset. For example,given a dataset containing 99% negative labels and 1% positive labels, thepositive labels are the minority class.

Contrast withmajority class.

SeeDatasets: Imbalanced datasetsin Machine Learning Crash Course for more information.

mixture of experts

A scheme to increaseneural network efficiency byusing only a subset of its parameters (known as anexpert) to processa given inputtoken orexample. Agating network routes each input token or example to the proper expert(s).

For details, see either of the following papers:

• Outrageously Large Neural Networks: TheSparsely-Gated Mixture-of-Experts Layer
• Mixture-of-Experts with Expert ChoiceRouting

ML

Abbreviation formachine learning.

MMIT

Abbreviation formultimodal instruction-tuned.

MNIST

A public-domain dataset compiled by LeCun, Cortes, and Burges containing60,000 images, each image showing how a human manually wrote a particulardigit from 0–9. Each image is stored as a 28x28 array of integers, whereeach integer is a grayscale value between 0 and 255, inclusive.

MNIST is a canonical dataset for machine learning, often used to test newmachine learning approaches. For details, seeThe MNIST Database of Handwritten Digits.

modality

A high-level data category. For example, numbers, text, images, video, andaudio are five different modalities.

model

In general, any mathematical construct that processes input data and returnsoutput. Phrased differently, a model is the set of parameters and structureneeded for a system to make predictions.Insupervised machine learning,a model takes anexample as input and infers aprediction as output. Within supervised machine learning,models differ somewhat. For example:

• A linear regression model consists of a set ofweightsand abias.
• Aneural network model consists of:
  • A set ofhidden layers, each containing one ormoreneurons.
  • The weights and bias associated with each neuron.
• Adecision tree model consists of:
  • The shape of the tree; that is, the pattern in which the conditionsand leaves are connected.
  • The conditions and leaves.

You can save, restore, or make copies of a model.

Unsupervised machine learning alsogenerates models, typically a function that can map an input example tothe most appropriatecluster.

  f(x, y) = 3x -5xy + y2 + 17

def half_of_greater(x, y):  if (x > y):    return(x / 2)  else    return(y / 2)

model capacity

The complexity of problems that a model can learn. The more complex theproblems that a model can learn, the higher the model's capacity. A model'scapacity typically increases with the number of model parameters. For a formaldefinition ofclassification model capacity, seeVC dimension.

model cascading

A system that picks the idealmodel for a specific inferencequery.

Imagine a group of models, ranging from very large (lots ofparameters) to much smaller (far fewer parameters).Very large models consume more computational resources atinference time than smaller models. However, very largemodels can typically infer more complex requests than smaller models.Model cascading determines the complexity of the inference query and thenpicks the appropriate model to perform the inference.The main motivation for model cascading is to reduce inference costs bygenerally selecting smaller models, and only selecting a larger model for morecomplex queries.

Imagine that a small model runs on a phone and a larger version of that modelruns on a remote server. Good model cascading reduces cost andlatency by enabling the smaller model to handle simple requestsand only calling the remote model to handle complex requests.

See alsomodel router.

model parallelism

A way of scaling training or inference that puts different parts of onemodel on differentdevices. Model parallelismenables models that are too big to fit on a single device.

To implement model parallelism, a system typically does the following:

1. Shards (divides) the model into smaller parts.
2. Distributes the training of those smaller parts across multiple processors.Each processor trains its own part of the model.
3. Combines the results to create a single model.

Model parallelism slows training.

See alsodata parallelism.

model router

The algorithm that determines the idealmodel forinference inmodel cascading.A model router is itself typically a machine learning model thatgradually learns how to pick the best model for a given input.However, a model router could sometimes be a simpler,non-machine learning algorithm.

model training

The process of determining the bestmodel.

MOE

Abbreviation formixture of experts.

Momentum

A sophisticated gradient descent algorithm in which a learning step dependsnot only on the derivative in the current step, but also on the derivativesof the step(s) that immediately preceded it. Momentum involves computing anexponentially weighted moving average of the gradients over time, analogousto momentum in physics. Momentum sometimes prevents learning from gettingstuck in local minima.

Mostly Basic Python Problems (MBPP)

A dataset for evaluating an LLM's proficiency in generating Python code.Mostly Basic Python Problemsprovides about 1,000 crowd-sourced programming problems.Each problem in the dataset contains:

• A task description
• Solution code
• Three automated test cases

MT

Abbreviation formachine translation.

multi-class classification

In supervised learning, aclassification problemin which the dataset containsmore than twoclasses of labels.For example, the labels in the Iris dataset must be one of the followingthree classes:

• Iris setosa
• Iris virginica
• Iris versicolor

A model trained on the Iris dataset that predicts Iris type on new examplesis performing multi-class classification.

In contrast, classification problems that distinguish between exactly twoclasses arebinary classification models.For example, an email model that predicts eitherspam ornot spamis a binary classification model.

In clustering problems, multi-class classification refers to more thantwo clusters.

SeeNeural networks: Multi-class classificationin Machine Learning Crash Course for more information.

multi-class logistic regression

Usinglogistic regression inmulti-class classification problems.

multi-head self-attention

An extension ofself-attention that applies theself-attention mechanism multiple times for each position in the input sequence.

Transformers introduced multi-head self-attention.

multimodal instruction-tuned

Aninstruction-tuned model that can process inputbeyond text, such as images, video, and audio.

multimodal model

A model whose inputs, outputs, or both include more than onemodality. For example, consider a model that takes both animage and a text caption (two modalities) asfeatures, andoutputs a score indicating how appropriate the text caption is for the image.So, this model's inputs are multimodal and the output is unimodal.

multinomial classification

Synonym formulti-class classification.

multinomial regression

Synonym formulti-class logistic regression.

Multi-sentence Reading Comprehension (MultiRC)

A dataset to evaluate an LLM's ability to answer multiple choice exercises.Each example in the dataset contains:

• A context paragraph
• A question about that paragraph
• Multiple answers to the question. Each answer is labeled True or False.Multiple answers may be True.

For example:

• Context paragraph:

  Susan wanted to have a birthday party. She called all of her friends. She hasfive friends. Her mom said that Susan can invite them all to the party. Herfirst friend could not go to the party because she was sick. Her second friendwas going out of town. Her third friend was not so sure if her parents wouldlet her. The fourth friend said maybe. The fifth friend could go to the partyfor sure. Susan was a little sad. On the day of the party, all five friendsshowed up. Each friend had a present for Susan. Susan was happy and sent eachfriend a thank you card the next week.

• Question: Did Susan's sick friend recover?

• Multiple answers:

  • Yes, she recovered. (True)
  • No. (False)
  • Yes. (True)
  • No, she didn't recover. (False)
  • Yes, she was at Susan's party. (True)

MultiRC is a component of theSuperGLUE ensemble.

For details, seeLooking Beyond the Surface: A Challenge Set for Reading Comprehension over Multiple Sentences.

multitask

A machine learning technique in which a singlemodel istrained to perform multipletasks.

Multitask models are created by training on data that is appropriate foreach of the different tasks. This allows the model to learn to shareinformation across the tasks, which helps the model learn more effectively.

A model trained for multiple tasks often has improved generalization abilitiesand can be more robust at handling different types of data.

N

Nano

A relatively smallGemini model designed for on-deviceuse. SeeGemini Nano for details.

See alsoPro andUltra.

NaN trap

When one number in your model becomes aNaNduring training, which causes many or all other numbers in your model toeventually become a NaN.

NaN is an abbreviation forNotaNumber.

natural language processing

natural language understanding

A subset ofnatural language processingthat determines theintentions of something said or typed. Natural languageunderstanding can go beyond natural language processing to consider complexaspects of language like context, sarcasm, and sentiment.

negative class

Inbinary classification, one class istermedpositive and the other is termednegative. The positive class isthe thing or event that the model is testing for and the negative class is theother possibility. For example:

• The negative class in a medical test might be "not tumor."
• The negative class in an emailclassification model might be "not spam."

Contrast withpositive class.

negative sampling

Synonym forcandidate sampling.

Neural Architecture Search (NAS)

A technique for automatically designing the architecture of aneural network. NAS algorithms can reduce the amountof time and resources required to train a neural network.

NAS typically uses:

• A search space, which is a set of possible architectures.
• A fitness function, which is a measure of how well a particulararchitecture performs on a given task.

NAS algorithms often start with a small set of possible architectures andgradually expand the search space as the algorithm learns more about whatarchitectures are effective. The fitness function is typically based on theperformance of the architecture on a training set, and the algorithm istypically trained using areinforcement learning technique.

NAS algorithms have proven effective in finding high-performingarchitectures for a variety of tasks, including imageclassification, text classification,andmachine translation.

neural network

Amodel containing at least onehidden layer.Adeep neural network is a type of neural networkcontaining more than one hidden layer. For example, the following diagramshows a deep neural network containing two hidden layers.

Each neuron in a neural network connects to all of the nodes in the next layer.For example, in the preceding diagram, notice that each of the three neuronsin the first hidden layer separately connect to both of the two neurons in thesecond hidden layer.

Neural networks implemented on computers are sometimes calledartificial neural networks to differentiate them fromneural networks found in brains and other nervous systems.

Some neural networks can mimic extremely complex nonlinear relationshipsbetween different features and the label.

See alsoconvolutional neural network andrecurrent neural network.

SeeNeural networksin Machine Learning Crash Course for more information.

neuron

In machine learning, a distinct unit within ahidden layerof aneural network. Each neuron performs the followingtwo-step action:

1. Calculates theweighted sum of input values multipliedby their corresponding weights.
2. Passes the weighted sum as input to anactivation function.

A neuron in the first hidden layer accepts inputs from the feature valuesin theinput layer. A neuron in any hidden layer beyondthe first accepts inputs from the neurons in the preceding hidden layer.For example, a neuron in the second hidden layer accepts inputs from theneurons in the first hidden layer.

The following illustration highlights two neurons and theirinputs.

A neuron in a neural network mimics the behavior of neurons in brains andother parts of nervous systems.

N-gram

An ordered sequence of N words. For example,truly madly is a 2-gram. Becauseorder is relevant,madly truly is a different 2-gram thantruly madly.

Manynatural language understandingmodels rely on N-grams to predict the next word that the user will typeor say. For example, suppose a user typedhappily ever.An NLU model based on trigrams would likely predict that theuser will next type the wordafter.

Contrast N-grams withbag of words, which areunordered sets of words.

SeeLarge language modelsin Machine Learning Crash Course for more information.

NLP

Abbreviation fornatural language processing.

NLU

Abbreviation fornatural language understanding.

node (decision tree)

In adecision tree, anycondition orleaf.

SeeDecision Treesin the Decision Forests course for more information.

node (neural network)

Aneuron in ahidden layer.

SeeNeural Networksin Machine Learning Crash Course for more information.

node (TensorFlow graph)

An operation in a TensorFlowgraph.

noise

Broadly speaking, anything that obscures the signal in a dataset. Noisecan be introduced into data in a variety of ways. For example:

• Human raters make mistakes in labeling.
• Humans and instruments mis-record or omit feature values.

non-binary condition

Acondition containing more than two possible outcomes.For example, the following non-binary condition contains three possibleoutcomes:

SeeTypes of conditionsin the Decision Forests course for more information.

nonlinear

A relationship between two or more variables that can't be represented solelythrough addition and multiplication. Alinear relationshipcan be represented as a line; anonlinear relationship can't berepresented as a line. For example, consider two models that each relatea single feature to a single label. The model on the left is linearand the model on the right is nonlinear:

SeeNeural networks: Nodes and hidden layersin Machine Learning Crash Course to experiment with different kindsof nonlinear functions.

non-response bias

Seeselection bias.

nonstationarity

A feature whose values change across one or more dimensions, usually time.For example, consider the following examples of nonstationarity:

• The number of swimsuits sold at a particular store varies with the season.
• The quantity of a particular fruit harvested in a particular regionis zero for much of the year but large for a brief period.
• Due to climate change, annual mean temperatures are shifting.

Contrast withstationarity.

no one right answer (NORA)

Aprompt havingmultiple correctresponses.For example, the following prompt has no one right answer:

Tell me a funny joke about elephants.

Evaluating the responses to no one right answer promptsis usually far more subjective than evaluating prompts withone right answer. For example, evaluating an elephantjoke requires a systematic way to determine how funny the joke is.

NORA

Abbreviation forno one right answer.

normalization

Broadly speaking, the process of converting a variable's actual rangeof values into a standard range of values, such as:

• -1 to +1
• 0 to 1
• Z-scores (roughly, -3 to +3)

For example, suppose the actual range of values of a certain feature is800 to 2,400. As part offeature engineering,you could normalize the actual values down to a standard range, suchas -1 to +1.

Normalization is a common task infeature engineering. Models usually train faster(and produce better predictions) when every numerical feature in thefeature vector has roughly the same range.

See alsoZ-score normalization.

SeeNumerical Data: Normalizationin Machine Learning Crash Course for more information.

Notebook LM

A Gemini-based tool that enables users to upload documents and thenuseprompts to ask questions about, summarize, or organizethose documents. For example, an author could upload several short storiesand ask Notebook LM to find their common themes or to identify which one wouldmake the best movie.

novelty detection

The process of determining whether a new (novel) example comes from the samedistribution as thetraining set. In other words, aftertraining on the training set, novelty detection determines whether anewexample (during inference or during additional training) is anoutlier.

Contrast withoutlier detection.

numerical data

Features represented as integers or real-valued numbers.For example, a house valuation model would probably represent the sizeof a house (in square feet or square meters) as numerical data. Representinga feature as numerical data indicates that the feature's values haveamathematical relationship to the label.That is, the number of square meters in a house probably has somemathematical relationship to the value of the house.

Not all integer data should be represented as numerical data. For example,postal codes in some parts of the world are integers; however, integer postalcodes shouldn't be represented as numerical data in models. That's because apostal code of20000 is not twice (or half) as potent as a postal code of10000. Furthermore, although different postal codesdo correlate to differentreal estate values, we can't assume that real estate values at postal code20000 are twice as valuable as real estate values at postal code 10000.Postal codes should be represented ascategorical datainstead.

Numerical features are sometimes calledcontinuous features.

SeeWorking with numerical datain Machine Learning Crash Course for more information.

NumPy

Anopen-source math librarythat provides efficient array operations in Python.pandas is built on NumPy.

O

objective

Ametric that your algorithm is trying to optimize.

objective function

The mathematical formula ormetric that a model aims to optimize.For example, the objective function forlinear regression is usuallyMean Squared Loss. Therefore, when training alinear regression model, training aims to minimize Mean Squared Loss.

In some cases, the goal is tomaximize the objective function.For example, if the objective function is accuracy, the goal isto maximize accuracy.

See alsoloss.

oblique condition

In adecision tree, acondition that involves more than onefeature. For example, if height and width are both features,then the following is an oblique condition:

  height > width

Contrast withaxis-aligned condition.

SeeTypes of conditionsin the Decision Forests course for more information.

offline

Synonym forstatic.

offline inference

The process of a model generating a batch ofpredictionsand then caching (saving) those predictions. Apps can then access the inferredprediction from the cache rather than rerunning the model.

For example, consider a model that generates local weather forecasts(predictions) once every four hours. After each model run, the systemcaches all the local weather forecasts. Weather apps retrieve the forecastsfrom the cache.

Offline inference is also calledstatic inference.

Contrast withonline inference.SeeProduction ML systems: Static versus dynamic inferencein Machine Learning Crash Course for more information.

one-hot encoding

Representing categorical data as a vector in which:

• One element is set to 1.
• All other elements are set to 0.

One-hot encoding is commonly used to represent strings or identifiers thathave a finite set of possible values.For example, suppose a certain categorical feature namedScandinavia has five possible values:

• "Denmark"
• "Sweden"
• "Norway"
• "Finland"
• "Iceland"

One-hot encoding could represent each of the five values as follows:

Thanks to one-hot encoding, a model can learn different connectionsbased on each of the five countries.

Representing a feature asnumerical data is analternative to one-hot encoding. Unfortunately, representing theScandinavian countries numerically is not a good choice. For example,consider the following numeric representation:

• "Denmark" is 0
• "Sweden" is 1
• "Norway" is 2
• "Finland" is 3
• "Iceland" is 4