Repository files navigation

• This is the 2024 version. For previous year' course materials, go tothis branch
• Lecture and seminar materials for each week are in ./week* folders, see README.md for materials and instructions
• Any technical issues, ideas, bugs in course materials, contribution ideas - add anissue
• Installing libraries and troubleshooting:this thread.

• week01Word Embeddings

  • Lecture: Word embeddings. Distributional semantics. Count-based (pre-neural) methods. Word2Vec: learn vectors. GloVe: count, then learn. Evaluation: intrinsic vs extrinsic. Analysis and Interpretability.Interactive lecture materials and more.
  • Seminar: Playing with word and sentence embeddings
  • Homework: Embedding-based machine translation system

• week02Text Classification

  • Lecture: Text classification: introduction and datasets. General framework: feature extractor + classifier. Classical approaches: Naive Bayes, MaxEnt (Logistic Regression), SVM. Neural Networks: General View, Convolutional Models, Recurrent Models. Practical Tips: Data Augmentation. Analysis and Interpretability.Interactive lecture materials and more.
  • Seminar: Text classification with convolutional NNs.
  • Homework: Statistical & neural text classification.

• week03Language Modeling

  • Lecture: Language Modeling: what does it mean? Left-to-right framework. N-gram language models. Neural Language Models: General View, Recurrent Models, Convolutional Models. Evaluation. Practical Tips: Weight Tying. Analysis and Interpretability.Interactive lecture materials and more.
  • Seminar: Build a N-gram language model from scratch
  • Homework: Neural LMs & smoothing in count-based models.

• week04Seq2seq and Attention

  • Lecture: Seq2seq Basics: Encoder-Decoder framework, Training, Simple Models, Inference (e.g., beam search). Attention: general, score functions, models. Transformer: self-attention, masked self-attention, multi-head attention; model architecture. Subword Segmentation (BPE). Analysis and Interpretability: functions of attention heads; probing for linguistic structure.Interactive lecture materials and more.
  • Seminar: Basic sequence to sequence model
  • Homework: Machine translation with attention

• week05Transfer Learning

  • Lecture: What is Transfer Learning? Great idea 1: From Words to Words-in-Context (CoVe, ELMo). Great idea 2: From Replacing Embeddings to Replacing Models (GPT, BERT). (A Bit of) Adaptors. Analysis and Interpretability.Interactive lecture materials and more.
  • Homework: fine-tuning a pre-trained BERT model

• week06LLMs and Prompting

  • Lecture: Scaling laws. Emergent abilities. Prompting (aka "in-context learning"): techiques that work; questioning whether model "understands" prompts. Hypotheses for why and how in-context learning works. Analysis and Interpretability.
  • Homework: manual prompt engneering and chain-of-thought reasoning

• week07Transformer architecture and training

  • Lecture: training tips for transformers; the evolution of transformer architecture from Vaswani et al (2017) to modern LLMs; parameter-efficient fine-tuning (PEFT)
  • Homework: fine-tuning a large language model with PEFT algorithms

• week08Reinforcement Learning from Human Feedback

  • Lecture: model alignment, RLHF, case study of InstructGPT and ChatGPT
  • Homework: fine-tune your own language model with RL (using HuggingFacetrl)

• week09 (extra)Domain Adaptation in NLP

  • Lecture: why do domain adaptation? Methods: reweighting, proxy labels, adversarial domain adaptation
  • Optional homework: implement domain adaptation when fine-tuning BERT models

• week10_Efficient Inference in NLP

  • Lecture: how NLP models are deployed, a survey of compression and acceleration: quantization, sparsification, ACT & more
  • Practice: implement RTN and GPTQ for 4-bit LLM quantization

• week11 (extra)_Retrieval Augmented Language Models

  • Guest lecture: retrieval in LMs, token-level retrieval (KNNLM & more), RAG, RETRO, tools: langchain , HF Agents, open problems

Course materials and teaching performed by

• Elena Voita - course author
• [Mikhail Diskin] [Ignat Romanov] [Ruslan Svirschevski] - lectures
• Valentina Broner - course admin for on-campus students
• Boris Kovarsky,David Talbot,Sergey Gubanov,Just Heuristic - help build course materials and/or held some classes
• 30+ volunteers who contributed and refined the notebooks and course materials. Without their help, the course would not be what it is today
• A mighty host of TAs who stoically grade hundreds of homework submissions from on-campus students each year