Classifications

• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/08—Learning methods
  • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/04—Architecture, e.g. interconnection topology
  • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/004—Artificial life, i.e. computing arrangements simulating life
  • G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/004—Artificial life, i.e. computing arrangements simulating life
  • G06N3/008—Artificial life, i.e. computing arrangements simulating life based on physical entities controlled by simulated intelligence so as to replicate intelligent life forms, e.g. based on robots replicating pets or humans in their appearance or behaviour
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/04—Architecture, e.g. interconnection topology
  • G06N3/042—Knowledge-based neural networks; Logical representations of neural networks
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/04—Architecture, e.g. interconnection topology
  • G06N3/044—Recurrent networks, e.g. Hopfield networks
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/04—Architecture, e.g. interconnection topology
  • G06N3/044—Recurrent networks, e.g. Hopfield networks
  • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/04—Architecture, e.g. interconnection topology
  • G06N3/045—Combinations of networks
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/04—Architecture, e.g. interconnection topology
  • G06N3/0464—Convolutional networks [CNN, ConvNet]
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/04—Architecture, e.g. interconnection topology
  • G06N3/0495—Quantised networks; Sparse networks; Compressed networks
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/08—Learning methods
  • G06N3/09—Supervised learning
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/08—Learning methods
  • G06N3/092—Reinforcement learning
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/08—Learning methods
  • G06N3/098—Distributed learning, e.g. federated learning
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N5/00—Computing arrangements using knowledge-based models
  • G06N5/01—Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound

Definitions

• This specificationrelates to the characterization of activity in a recurrent artificial neural network.
• the characterization of activitycan be used, e.g., in the identification of decision moments, as well as in encoding/decoding signals in contexts such as transmission, encryption, and data storage. It also relates to encoding and decoding information, and to systems and techniques that use the encoded information in various contexts.
• the encoded informationcan represent the activity in a neural network, e.g., a recurrent neural network.
• Artificial neural networksare devices that are inspired by the structure and functional aspects of networks of biological neurons.
• artificial neural networksmimic the information encoding and other processing capabilities of networks of biological neurons using a system of interconnected constructs called nodes.
• the arrangement and strength of connections between nodes in an artificial neural networkdetermines the results of information processing or information storage by the artificial neural network.
• Neural networkscan be trained to produce a desired signal flow within the network and achieve desired information processing or information storage results. In general, training a neural network will change the arrangement and/or strength of connections between nodes during a learning phase. A neural network can be considered trained when sufficiently appropriate processing results are achieved by the neural network for given sets of inputs.
• Non-linear data processingdoes not satisfy the superposition principle, i.e., the variables that are to be determined cannot be written as a linear sum of independent components.
• Examples of contexts in which non-linear data processing is usefulinclude pattern and sequence recognition, speech processing, novelty detection and sequential decision making, complex system modelling, and systems and techniques in a variety of other contexts. Encoding and decoding both convert information from one form or representation into another.
• the different representationscan provide different features that are more or less useful in different applications. For example, some forms or representations of information (e.g., natural language) may be easier for a human to understand. Other forms or representations may be smaller in size (e.g.,“compressed”) and easier to transmit or store. Still other forms or representations may intentionally obscure the information content (e.g., information can be cryptographically encoded).
• an encoding or decoding processwill generally follow a predefined set of rules or an algorithm that establishes the correspondence between the information in the different forms or representations.
• encoding process that yields a binary codemay assign a role or meaning to individual bits according to their position in a binary sequence or vector.
• This specificationdescribes technologies relating to the characterization of activity in an artificial neural network.
• a method for identifying decision moments in a neural networkincludes determining a complexity of patterns of activity in the recurrent artificial neural network, wherein the activity is responsive to input into the recurrent artificial neural network, determining a timing of activity having a complexity that is distinguishable from other activity that is responsive to the input, and identifying the decision moments based on the timing of the activity that has the distinguishable complexity.
• a method for characterizing activity in a recurrent artificial neural networkincludes identifying pre-defined clique patterns of activity of the recurrent artificial neural network. The method is performed by data processing apparatus.
• a methodcan include outputting a binary sequence of zeros and ones from a recurrent artificial neural network, wherein each digit in the sequence represents whether or not a particular group of nodes in the recurrent artificial neural network displays a respective pattern of activity.
• a method of structuring a recurrent artificial neural networkcan include characterizing a complexity of patterns of activity that can arise in the recurrent artificial neural network that comprises a structured collection of nodes and links between the nodes, and evolving the structure of the recurrent artificial neural network to increase the complexity of the patterns of activity. This method of structuring can also be used, e.g., a part of a method of training the recurrent artificial neural network.
• decision moments in artificial recurrent neural networkscan be identified based on characteristics of the dynamic properties of the neural network during information processing. Rather than waiting for the artificial neural network to reach a pre-defined end of a logic sequence, decision moments in artificial neural networks can be identified based on characteristics of the functional states of the artificial neural network in the midst of information processing.
• the characteristics of the dynamic properties of recurrent artificial neural networks during information processingcan be used in a variety of signalling operations, including signal transmission, encoding, encryption, and storage.
• the characteristics of the activity in a recurrent artificial neural network during information processingreflects the input and can be considered to be an encoded form of the input (i.e., the“output” of a recurrent artificial neural network in encoding processes).
• These characteristicscan be, e.g., transmitted to a remote receiver which could decode the transmitted characteristics to reconstitute the input or a portion of the input.
• activity in different groups of nodes a recurrent artificial neural networkcan be represented as a binary sequence of zeros and ones, each digit indicating whether or not that activity comports with a pattern. Since the activity can be the output of a recurrent artificial neural network in some contexts, the output of a recurrent artificial neural network can be represented as a vector of binary digits and compatible with digital data processing.
• such characterization of the dynamic properties of a recurrent artificial neural networkcan be used prior to and/or during training to increase the likelihood of complex patterns of activity arising during information processing.
• the links between nodes in a recurrent neural networkcan be intentionally evolved to increase the complexity of activity patterns.
• the links between nodes in a recurrent artificial neural networkcan be intentionally evolved to increase the likelihood that, e.g., clique patterns and directed clique patterns of activity arise during information processing. This can reduce the time and effort needed to train a recurrent artificial neural network.
• such characterization of the dynamic properties of a recurrent artificial neural networkcan be used to determine a degree of completeness in the training of a recurrent neural network.
• a recurrent artificial neural network that displays particular types of ordering in the activitye. g, clique patterns and directed clique patterns
• the extent of trainingcan be quantified by quantifying the degree of ordering of the activity in the recurrent artificial neural network.
• FIG. 1is a schematic illustration of the structure of recurrent artificial neural network device.
• FIGS. 2 and 3are schematic illustrations of the function of recurrent artificial neural network device in different windows of time.
• FIG. 4is a flowchart of a process for identifying decision moments in a recurrent artificial neural network based on characterization of the activity in the network.
• FIG. 5is a schematic illustration of patterns of activity that can be identified and used for identifying decision moments in a recurrent artificial neural network.
• FIG. 6is a schematic illustration of patterns of activity that can be identified and used for identifying decision moments in a recurrent artificial neural network.
• FIG. 7is a schematic illustration of patterns of activity that can be identified and used for identifying decision moments in a recurrent artificial neural network.
• FIG. 8is a schematic illustration of a data table that can be used in a determination of the complexity or degree of ordering in the activity patterns in a recurrent artificial neural network device.
• FIG. 9is a schematic illustration of a determination of the timing of activity patterns that have a distinguishable complexity.
• FIG. 10is a flowchart of a process for encoding signals using a recurrent artificial neural network based on characterization of the activity in the network.
• FIG. 11is a flowchart of a process for decoding signals using a recurrent artificial neural network based on characterization of the activity in the network.
• FIGS. 12, 13, and 14are schematic illustrations of an binary form or representation of topological structures.
• FIGS. 15 and 16schematically illustrate an example of how the presence or absence of features that correspond to different bits are not independent of one another.
• FIGS. 17, 18, 19, 20are schematic illustrations of the use of representations of the occurrence of topological structures in the activity in a neural network in four different classification systems.
• FIGS. 21, 22are schematic illustration of edge devices that include a local artificial neural network that can be trained using representations of the occurrence of topological structures corresponding to activity in a source neural network.
• FIG. 23is a schematic representation of a system in which local neural networks can be trained using representations of the occurrence of topological structures corresponding to activity in a source neural network.
• FIGS. 24, 25, 26, 27are schematic illustrations of the use of representations of the occurrence of topological structures in the activity in a neural network in four different systems.
• FIG. 28is a schematic illustration of a system 0 that includes an artificial neural network that can be trained using representations of the occurrence of topological structures corresponding to activity in a source neural network.
• FIG. 1is a schematic illustration of the structure of a recurrent artificial neural network device 100.
• Recurrent artificial neural network device 100is a device that mimics the information encoding and other processing capabilities of networks of biological neurons using a system of interconnected nodes.
• Recurrent artificial neural network device 100can be implemented in hardware, in software, or in combinations thereof.
• the illustration of recurrent artificial neural network device 100includes a plurality of nodes 101, 102, ... , 107 that are interconnected by a plurality of structural links 110.
• Nodes 101, 102, ... , 107are discrete information processing constructs that are analogous to neurons in biological networks.
• Nodes 101, 102, ... , 107generally process one or more input signals received over one or more of links 110 to produce one or more output signals that are output over one or more of links 110.
• nodes 101, 102, ... , 107can be artificial neurons that weight and sum multiple input signals, pass the sum through one or more non-linear activation functions, and output one or more output signals.
• nodes 101, 102, ... , 107can operate as accumulators.
• nodes 101, 102, ... , 107can operate in accordance with an integrate-and-fire model in which one or more signals accumulate in a first node until a threshold is reached. After the threshold is reached, the first node fires by transmitting an output signal to a connected, second node along one or more of links 110.
• the second node 101, 102, ... , 107accumulates the received signal and, if a threshold is reached, then the second node 101, 102, ... , 107 transmits yet another output signal to a further connected node.
• Structural links 110are connections that are capable of transmitting signals between nodes 101, 102, ... , 107.
• all structural links 110are treated herein as identical bidirectional links that convey a signal from every first of nodes 101, 102, ... , 107 to every second of nodes 101, 102, ... , 107 in identically the same manner as a signal is conveyed from the second to the first.
• thisis not necessarily the case.
• structural links 110can be unidirectional links that convey a signal from a first of nodes 101, 102, ... , 107 to a second of nodes 101, 102, ... , 107 without conveying signals from the second to the first.
• structural links 110can have diverse properties other than or in addition to directionality.
• different structural links 110can carry signals of different magnitudes— resulting in a different strengths of interconnection between respective of nodes 101, 102, ... , 107.
• different structural links 110can carry different types of signal (e.g., inhibitory and/or excitatory signals).
• structural links 110can be modelled on the links between soma in biological systems and reflect at least a portion of the enormous morphological, chemical, and other diversity of such links.
• recurrent artificial neural network device 100is a clique network (or subnetwork) in that every node 101, 102, ... , 107 is connected to every other node 101, 102, ... , 107. This is not necessarily the case. Rather, in some
• each node 101, 102, ... , 107can be connected to a proper subset of nodes 101, 102, ... , 107 (by identical links or diverse links, as the case may be).
• recurrent artificial neural network device 100is illustrated with only seven nodes.
• real-world neural network deviceswill include significantly larger numbers of nodes.
• neural network devicescan include hundreds of thousands, millions, or even billions of nodes.
• recurrent neural network device 100can be a fraction of a larger recurrent artificial neural network (i.e., a subnetwork).
• accumulation and signal transmission processesrequire the passage of time in the real world.
• soma of a neuronintegrates input received over time
• signal transmission from neuron to neuronrequires times that are determined by, e.g., the signal transmission velocity and the nature and length of the links between neurons.
• the state of a biological neural network deviceis dynamic and changes over time.
• timeis artificial and represented using mathematical constructs. For example, rather than requiring a real world passage of time for signals to transmit from node to node, such signals can be represented in terms of artificial units that are generally unrelated to the real world passage of time— as measured in computer clock cycles or otherwise. Nevertheless, the state of an artificial recurrent neural network device can be described as“dynamic” in that it changes with respect to these artificial units.
• time unitsare referred to herein as“time” units. Nevertheless, it is to be understood that these units are artificial and generally do not correspond to the real world passage of time.
• FIGS. 2 and 3are schematic illustrations of the function of recurrent artificial neural network device 100 in different windows of time. Because the state of device 100 is dynamic, the functioning of device 100 can be represented using the signal transmission activity that occurs within a window. Such a functional illustration generally shows activity in only a fraction of links 110. In particular, since in general not every link 110 conveys a signal within a particular window, not every link 110 is illustrated as actively contributing to the functioning of the device 100 in these illustrations.
• an active link 110is illustrated as a relatively thick solid line connecting a pair of nodes 101, 102, ... , 107.
• inactive links 110are illustrated as dashed lines. This is for the sake of illustration only. In other words, the structural connections formed by links 110 exist whether or not links 110 are active.
• the direction of the activityis also schematically illustrated.
• the relatively thick solid lines that illustrate active of links 110also include arrow heads that denote the direction of signal transmission along the link during the relevant window.
• the direction of signal transmission in a single windowdoes not conclusively constrain the link to being a unidirectional link having the indicated directionality.
• a linkin a first functional illustration for a first window of time, a link can be active in a first direction.
• a linkcan be active in the opposite direction.
• the directionality of signal transmissionwill conclusively indicate the directionality of the link.
• feedforward neural network devicesinformation moves exclusively in a single direction (i.e., forward) to an output layer of nodes that is at the end of the network.
• Feedforward neural network devicesindicate that a“decision” has been reached and that information processing is complete by the propagation of the signals through the network to the output layer.
• the connections between nodesform cycles and the activity of the network dynamically progresses without a readily identifiable decision.
• the first nodecan transmit a signal to the second node, which in response can transmit a signal to the third.
• the third nodecan transmit a signal back to the first.
• the signals received by the first nodecan be responsive— at least in part— to the signals transmitted from that same node.
• FIGS. 2 and 3illustrate this in a network that is only slightly larger than a three-node recurrent neural network.
• the functional illustrations shown in FIG. 2can be illustrative of activity within a first window and FIG. 3 can be illustrative of activity within a second, immediately following.
• a collection of signal transmission activityappears to originate in node 104 and progress in a generally clockwise direction through device 100 during the first window.
• the second windowat least some of the signal transmission activity generally appears to return to node 104. Even in such a simplistic illustration, signal transmission does not proceed in a manner that yields a clearly identifiable output or end.
• a recurrent neural networkof, e.g., thousands of nodes or more
• signal propagationcan occur over a huge number of paths and that these signals lack a clearly identifiable“output” location or time.
• the networkmay by design return to a quiescent state in which only background or even no signal transmission activity occurs, the quiescent state itself does not indicate the results of information processing.
• the recurrent neural networkalways returns to the quiescent state regardless of the input.
• The“output” or the result of the information processingis thus encoded in the activity that occurs within the recurrent neural network in response to a particular input.
• FIG. 4is a flowchart of a process 400 for identifying decision moments in a recurrent artificial neural network based on characterization of the activity in the network.
• a decision momentis a point of time when the activity in a recurrent artificial neural network is indicative of the results of information processing by the network in response to an input.
• Process 400can be performed by a system of one or more data processing apparatus that perform operations in accordance with the logic of one or more sets of machine-readable instructions.
• process 400can be performed by that same system of one or more computers that executes software for implementing the recurrent artificial neural network used in process 400.
• the system performing process 400receives a notification that a signal has been input into the recurrent artificial neural network at 405.
• the input of the signalis a discrete injection event in which, e.g., information is injected into one or more nodes and/or one or more links of the neural network.
• the input of the signalis a stream of information that is injected into the one or more nodes and/or links of the neural network over a period of time.
• the notificationindicates that the artificial neural network is actively processing information and not, e.g., in a quiescent state.
• the notificationis received from the neural network itself, e.g., such as when the neural network exits an identifiable quiescent state.
• the system performing process 400divides the responsive activity in the network into a collection of windows at 410.
• the windowscan subdivide the time between injection and a return to a quiescent state into a number of periods during which the activity displays variable complexities.
• the duration of the injection(and optionally the time to return to a quiescent state after injection is complete) can be subdivided into windows during which the activity displays variable complexities.
• the windowsall have the same duration, but this is not necessarily the case. Rather, in some implementations, the windows can have different durations. For example, in some implementations, duration can increase as time since a discrete injection event has occurred increases.
• the windowscan be a successive series of separate windows. In other implementations, the windows overlap in time so that one window begins before a previous window ends. In some cases, the windows can be a moving window that moves in time.
• different durations of windowsare defined for different determinations of the complexity of activity. For example, for activity patterns that define activity occurring between relatively larger numbers of nodes, the windows can have a relatively longer duration than windows that are defined for activity patterns that define activity occurring between relatively smaller numbers of nodes. For example, in the context of patterns 500 of activity (FIG. 5), a window that is defined for identifying activity that comports with pattern 530 can be longer than a window that is defined for identifying activity that comports with pattern 505.
• the system performing process 400identifies patterns in the activity in the network in the different windows at 415.
• patterns in the activitycan be identified by treating a functional graph as a topological space with nodes as points.
• the activity patterns that are identifiedare cliques, e.g., directed cliques, in a functional graph of the network.
• the system performing process 400determines the complexity of the activity patterns in different windows at 420.
• the complexitycan be a measure of the likelihood that an ordered pattern of activity arises within a window.
• activity patterns that arise randomlywould be relatively simple.
• activity patterns that show non- random orderare relatively complex.
• the complexity of an activity patterncan be measured using, e.g., the simplex counts or the Betti numbers of the activity pattern.
• the system performing process 400determines the timing of activity patterns having a distinguishable complexity at 425.
• a particular activity patterncan be distinguishable based on a complexity that deviates upwards or deviates downward, e.g., from a fixed or a variable baseline.
• the timing of activity patterns that indicate particularly high levels or particularly low levels of non-random order in the activitycan be determined.
• signal inputis a discrete injection event
• deviationse.g., from a stable baseline or from a curve that is characteristic of the neural network’s average response to a variety of different discrete injection events can be used to determine the timing of distinguishable complex activity patterns.
• deviationse.g., from a stable baseline or from a curve that is characteristic of the neural network’s average response to a variety of different discrete injection events can be used to determine the timing of distinguishable complex activity patterns.
• the system performing process 400times the reading of the output from the neural network based on the timing of distinguishably complex activity patterns at 430.
• the output of the neural networkcan be read at the same time that distinguishable complex activity patterns arise.
• the observed activity patternsthemselves can also be taken as the output of the recurrent artificial neural network.
• FIG. 5is an illustration of patterns 500 of activity that can be identified and used for identifying decision moments in a recurrent artificial neural network.
• patterns 500can be identified at 415 in process 400 (FIG. 4).
• Patterns 500are illustrations of activity within a recurrent artificial neural network.
• a functional graphis treated as a topological space with nodes as points.
• Activity in nodes and links that comports with patterns 500can be recognized as ordered regardless of the identity of the particular nodes and/or links that participate in the activity.
• first pattern 505can represent the activity between nodes 101, 104, 105 in FIG. 2, with point 0 in pattern 505 as node 104, point 1 as node 105, and point 2 as node 101.
• first pattern 505can also represent the activity between nodes 104, 105, 106 in FIG.
• patterns 500are all directed cliques or directed simplices. In such patterns, activity originates from a source node that transmits signals to every other node in the pattern. In patterns 500, such source nodes are designated as point 0 whereas the other nodes are designated as points 1, 2, .... Further, in directed cliques or simplices, one of the nodes acts a sink and receives signals transmitted from every other node in the pattern. In patterns 500, such sink nodes are designated as the highest numbered point in the pattern. For example, in pattern 505, the sink node is designated as point 2. In pattern 510, the sink node is designated as point 3. In pattern 515, the sink node is designated as point 3, and so on. The activity represented by patterns 500 is thus ordered in a
• Each of patterns 500has a different number of points and reflects ordered activity in a different number of nodes.
• pattern 505is a 2D-simplex and reflects activity in three nodes
• pattern 510is a 3D-simplex and reflects activity in four nodes, and so on.
• the degree of ordering and the complexity of the activityFor example, for a large collection of nodes that have a certain level of random activity within a window, some of that activity may comport with pattern 505 out of happenstance. However, it is progressively more unlikely that random activity will comport with the respective of patterns 510, 515, 520....
• the presence of activity that comports with pattern 530indicates a relatively higher degree of ordering and complexity in the activity that the presence of activity that comports with pattern 505.
• different duration windowscan be defined for different determinations of the complexity of activity. For example, when activity that comports with pattern 530 is to be identified, longer duration windows can be used than when activity that comports with pattern 505 is to be identified.
• FIG. 6is an illustration of patterns 600 of activity that can be identified and used for identifying decision moments in a recurrent artificial neural network.
• patterns 600can be identified at 415 in process 400 (FIG. 4).
• patterns 600are illustrations of activity within a recurrent artificial neural network. However, patterns 600 depart from the strict ordering of patterns 500 in that patterns 600 are not all directed cliques or directed simplices. In particular, patterns 605, 610 have a lower directionality than pattern 515. Indeed, pattern 605 lacks a sink node altogether. Nevertheless, patterns 605, 610 indicate degree of ordered activity that exceeds that expected through random happenstance and can be used to determine the complexity of activity in a recurrent artificial neural network.
• FIG. 7is an illustration of patterns 700 of activity that can be identified and used for identifying decision moments in a recurrent artificial neural network.
• patterns 700can be identified at 415 in process 400 (FIG. 4).
• Patterns 700are groups of directed cliques or directed simplices of the same dimension (i.e., have the same number of points) that define patterns involving more points than the individual cliques or simplices and enclose cavities within the group of directed simplices.
• pattern 705includes six different three point, 2-dimensions patterns 505 that together define a homology class of degree two
• pattern 710includes eight different three point, 2-dimensions patterns 505 that together define a second homology class of degree two.
• Each of the three point, 2-dimensions patterns 505 in patterns 705, 710can be thought of as enclosing a respective cavity.
• the n lh Betti number associated with a directed graphprovides a count of such homology classes within a topological representation.
• the activity illustrated by patterns such as patterns 700illustrates a relatively high degree of ordering of the activity within a network that is unlikely to arise by random happenstance. Patterns 700 can be used to characterize the complexity of that activity.
• the patterns of activity that are identifiedare discarded or otherwise ignored during the identification of decision moments.
• activity that comports with the five point, 4-dimensional simplex pattern 515inherently includes activity that comports with the four point, 3 -dimensional and three point, 2-dimension simplex patterns 510, 505.
• points 0, 2, 3, 4 and points 1, 2, 3, 4 in 4-dimensional simplex pattern 515 of FIG. 5both comport with 3-dimensional simplex pattern 510.
• patterns that include fewer points— and hence are of a lower dimension—can be discarded or otherwise ignored during the identification of decision moments.
• FIG. 8is a schematic illustration of a data table 800 that can be used in such a determination.
• Data table 800can be used to determine the complexity of the activity patterns in isolation or in conjunction with other activities.
• data table 800can be used at 420 in process 400 (FIG. 4).
• table 800includes a number count of pattern occurrences during a window“N,” where the number counts of activity that matches patterns of different dimensions are presented in different rows.
• row 805includes a number count (i.e.,“2032”) of the occurrences of activity that matches one or more three point, 2-dimensional patterns
• row 810includes a number count (i.e., “877”) of the occurrences of activity that matches one or more four point, 3-dimensional patterns. Since an occurrence of the patterns indicates that the activity has an order which is non-random, the number counts also provide generalized characterization of the overall complexity of the activity patterns.
• a table that is analogous to table 800can be formed for each window that is defined, e.g., at 410 in process 400 (FIG. 4).
• table 800includes a separate row and a separate entry for every type of activity pattern, this is not necessarily the case.
• one of more countscan be omitted (e.g., counts of simpler patterns) can be omitted from table 800 and from a determination of the complexity.
• a single row or entrycan include counts of occurrences of multiple activity patterns.
• FIG. 8presents the number count in a table 800, this is not necessarily the case.
• the number countcan be presented as a vector (e.g., ⁇ 2032, 877, 133, 66, 48, ...>).
• the countscan be expressed in binary and can be compatible with digital data processing infrastructure.
• number counts of the occurrences of the patternscan be weighted or combined to determine the degree or complexity of the ordering, e.g., at 420 in process 400 (FIG. 4).
• the Euler characteristiccan provide an approximation of the complexity of the activity and is given by:
• the patternscan be, e.g., the directed clique patterns 500 (FIG. 5).
• pattern occurrencescan be weighted based on the weights of the links that are active.
• the strength of connection between nodes in an artificial neural networkcan vary, e.g., as a consequence of how active the connection was during training.
• An occurrence of a pattern of activity along a collection of relatively stronger linkscan be weighted differently from the occurrence of that same pattern of activity along a collection of relatively weaker links.
• the sum of the weight of the links that are activecan be used to weight the occurrence.
• the Euler characteristic or other measure of complexitycan be normalized by the total number of patterns that are matched within a particular window and/or the total number of patterns that it is possible for a given network to form given its structure.
• An example of a normalization with regard to the total number of patterns that it is possible for a network to formis given below in Equations 2, 3.
• occurrences of higher dimension patterns involving larger numbers of nodescan be weighted more heavily than occurrences of lower dimension patterns involving smaller numbers of nodes. For example, the probability of forming directed cliques decreases rapidly with increasing dimension. In particular, to form an n- clique from n+l nodes, one needs (n+l)n/2 edges all oriented correctly. This probability can be reflected in the weighting.
• both the dimension and the directionality of patternscan be used to weight occurrences of patterns and determine the complexity of the activity. For example, with reference to FIG. 6, occurrences of five point, 4-dimensional pattern 515 can be weighted more heavily than occurrences of five point, 4-dimensional patterns 605, 610 in accordance with the differences in directionality of those patterns.
• Sx actlveindicates the number of active occurrences of a pattern of n points
• ERNis the calculation for an equivalent random network, i.e., a network with the same number of nodes randomly connected.
• SCis given by where Sx sllent indicates the number of occurrences of a pattern of n points when the recurrent artificial neural network is silent and can be thought of as embodying the total number of patterns that it is possible for the network to form.
• the patternscan be, e.g., the directed clique patterns 500 (FIG. 5).
• FIG. 9is a schematic illustration of a determination of the timing of activity patterns that have a distinguishable complexity.
• the determination illustrated in FIG. 9can be performed in isolation or in conjunction with other activities. For example, the determination can be performed at 425 in process 400 (FIG. 4).
• FIG. 9includes a graph 905 and a graph 910.
• Graph 905illustrates occurrences of patterns as a function of time along the x-axis.
• individual occurrencesare illustrated schematically as vertical lines 906, 907, 908, 909.
• Each row of occurrencescan be instances where activity matches a respective pattern or class of pattern.
• the top row of occurrencescan be instances where activity matches pattern 505 (FIG. 5)
• the second row of occurrencescan be instances where activity matches pattern 510 (FIG. 5)
• the third row of occurrencescan be instances where activity matches pattern 515 (FIG. 5), and so on.
• Graph 905also includes dashed rectangles 915, 920, 925 that schematically delineate different windows of time when the activity patterns have a distinguishable complexity. As shown, the likelihood that activity in the recurrent artificial neural network matches a pattern indicative of complexity is higher during the windows delineated by dashed rectangles 915, 920, 925 than outside those windows.
• Graph 910illustrates the complexity associated with these occurrences as a function of time along the x-axis.
• Graph 910includes a first peak 930 in complexity that coincides with the window delineated by dashed rectangle 915 and a second peak 935 in complexity that coincides with the window delineated by dashed rectangles 920, 925.
• the complexity illustrated by peaks 930, 925is distinguishable from what can be considered to be a baseline level 940 of complexity.
• the times at which the output of a recurrent artificial neural network is to be readcoincide with the occurrences of activity patterns that have a distinguishable complexity.
• the output of a recurrent artificial neural networkcan be read at peaks 930, 925, i.e., during the windows delineated by dashed rectangles 915, 920, 925.
• the identification of distinguishable levels of complexity in a recurrent artificial neural networkis particularly beneficial when the input is a stream of data.
• data streamsinclude, e.g., video or audio data.
• data streamshave a beginning, it is generally desirable to process information in the data stream that does not have a pre-defined relationship with the beginning of the data stream.
• a neural networkcould perform object recognition such as, e.g., recognizing bicyclists in the vicinity of an automobile.
• object recognitionsuch as, e.g., recognizing bicyclists in the vicinity of an automobile.
• Such a neural networksshould be able to recognizing bicyclists regardless of when those bicyclists appear in the video stream, i.e., without regard to time since the beginning of the video.
• any patterns of activity in the neural networkwill generally display a low or quiescent level of complexity. These low or quiescent level of complexity are displayed regardless of the continuous (or nearly continuous) input of streaming data into the neural network device.
• the complexity of the activitywill become distinguishable and indicate the time at which an object is recognized in the video stream.
• the timing of a distinguishable level of complexity of the activitycan also act as a yes/no output as to whether the data in the data stream satisfies certain criteria.
• the activity patternsthat have a distinguishable complexity.
• the identity and activity of the nodes that participate in activity that comports with the activity patternscan be considered the output of the recurrent artificial neural network.
• the identified activity patternscan thus illustrate the result of processing by the neural network, as well as the timing when this decision is to be read.
• the content of the decisioncan be expressed in a variety of different forms. For example, in some implementations and as discussed in further detail below, the content of the decision can be expressed as a binary vector or matrix of ones and zeros.
• Each digitcan indicate whether or not a pattern of activity is present or not, e.g., for a pre-defined group of nodes and/or a predefined duration.
• the content of the decisionis expressed in binary and can be compatible with traditional digital data processing
• FIG. 10is a flowchart of a process 1000 for encoding signals using a recurrent artificial neural network based on characterization of the activity in the network.
• Signalscan be encoded in a variety of different contexts such as, e.g., transmission, encryption, and data storage.
• Process 1000can be performed by a system of one or more data processing apparatus that perform operations in accordance with the logic of one or more sets of machine-readable instructions.
• process 1000can be performed by that same system of one or more computers that executes software for implementing the recurrent artificial neural network used in process 1000.
• process 1000can be performed by the same data processing apparatus that performs process 400.
• process 1000can be performed by, e.g., the encoder in a signal transmission system or the encoder of a data storage system.
• the system performing process 1000inputs a signal into a recurrent artificial neural network at 1005.
• the input of the signalis a discrete injection event.
• the input signalis streamed into the recurrent artificial neural network.
• the system performing process 1000identifies one or more decision moments in the recurrent artificial neural network at 1010. For example, the system can identify one or more decision moments by performing process 400 (FIG. 4).
• the system performing process 1000reads the output of the recurrent artificial neural network at 1015.
• the content of the output of the recurrent artificial neural networkis the activity in the neural network that matches the patterns used to identify the decision point(s).
• individual“reader nodes”can be added to a neural network to identify occurrences of a particular pattern of activity at a particular collection of nodes and hence to read the output of the recurrent artificial neural network at 1015.
• the reader nodescan fire if and only if the activity at a particular collection of nodes satisfies timing (and possibly magnitude, as well) criteria. For example, in order to read an occurrence of pattern 505 (FIG. 5) at nodes 104, 105, 106 (FIGS, 2, 3), a reader node could be connected to nodes 104, 105, 106 (or the links 110 between them). The reader node would itself only become active if a pattern of activity involving nodes 104, 105, 106 (or their links) occurred.
• the use of such reader nodeswould eliminate the need to define windows of time for the recurrent artificial neural network as a whole.
• individual reader nodescan be connected to different nodes and/or numbers of nodes (or the links between them).
• the individual reader nodescan be set to have tailored responses (e.g., different decay times in an integrate-and-fire model) to identify different activity patterns.
• the system performing process 1000transmits or stores the output of the recurrent artificial neural network at 1020.
• the particular action performed at 1020can reflect the context in which process 1000 is being used. For example, in contexts where secure or compressed communications are desired, the system performing process 1000 can transmit the output of the recurrent neural network to a receiver that has access to the same or to a similar recurrent neural network. As another example, in contexts where secure or compressed data storage is desired, the system performing process 1000 can record the output of the recurrent neural network in one or more machine-readable data storage devices for later access.
• the complete output of the recurrent neural networkis not transmitted or stored.
• the content of the output of the recurrent neural networkis the activity in the neural network that matches the patterns indicative of complexity in the activity
• only activity that matches relatively more complex or higher dimensional activitymay be transmitted or stored.
• patterns 500FIG. 5
• only activity that matches patterns 515, 520, 525, and 530is transmitted or stored whereas activity that matches patterns 505, 510 is ignored or discarded. In this way, a lossy process allows the volume of data that is transmitted or stored to be reduced at the cost of the completeness of the information being encoded.
• Process 1100is a flowchart of a process 1100 for decoding signals using a recurrent artificial neural network based on characterization of the activity in the network.
• Signalscan be decoded in a variety of different contexts such as, e.g., signal reception, decryption, and reading data from storage.
• Process 1100can be performed by a system of one or more data processing apparatus that perform operations in accordance with the logic of one or more sets of machine-readable instructions.
• process 1100can be performed by that same system of one or more computers that executes software for implementing the recurrent artificial neural network used in process 1100.
• process 1100can be performed by the same data processing apparatus that performs process 400 and/or process 1000.
• process 1100can be performed by, e.g., the decoder in a signal reception system or the decoder of a data storage system.
• the system performing process 1100receives at least a portion of the output of a recurrent artificial neural network at 1105.
• the particular action performed at 1105can reflect the context in which process 1100 is being used.
• the system performing process 1000can receive a transmitted signal that includes the output of the recurrent artificial neural network or read a machine -readable data storage device that stores the output of the recurrent artificial neural network.
• the system performing process 1100reconstructs the input of the recurrent artificial neural network from the received output at 1110.
• Reconstructioncan proceed in a variety of different ways.
• a second artificial neural network(recurrent or not) can be trained to reconstruct the input into the recurrent neural network from the output received at 1105.
• a decoderthat has been trained using machine learning (including but not limited to deep learning) can be trained to reconstruct the input into the recurrent neural network from the output received at 1105.
• machine learningincluding but not limited to deep learning
• process 1100can include receiving user input specifying an extent to which the input is to be reconstructed and, in response, adjust the reconstruction at 1110 accordingly.
• the user inputcould specify that a complete reconstruction is not needed.
• the system performing process 1100adjusts the reconstruction.
• the content of the output of the recurrent neural networkis the activity in the neural network that matches the patterns indicative of complexity in the activity
• only the output that characterizes activity that matches relatively more complex or higher dimensional activitywould be used to reconstruct the input.
• patterns 500FIG. 5
• only activity that matches patterns 515, 520, 525, and 530could be used to reconstruct the input, whereas activity that matches patterns 505, 510 could be ignored or discarded. In this way, a lossy reconstruction can proceed in selected circumstances.
• processes 1000, 1100can be used for peer-to-peer encrypted communications.
• both the sender (i.e., the encoder) and the receiver (i.e., the decoder)can be provided with the same recurrent artificial neural network.
• the shared recurrent artificial neural networkcan be tailored to ensure that a third party cannot reverse-engineer it and decrypt the signal, including:
• the decision moment time pointscan be used as keys to decrypt the signal.
• processes 1000, 1100are presented in terms of encoding and decoding a single recurrent artificial neural network, processes 1000, 1100 can also be applied in systems and processes that rely upon multiple recurrent artificial neural networks. These recurrent artificial neural networks can operate either in parallel or in series. As an example of series operation, the output of a first recurrent artificial neural network can serve as the input of a second recurrent artificial neural network. The resultant output of the second recurrent artificial neural network is a twice encoded (or twice encrypted) version of the input into the first recurrent artificial neural network.
• Such a series arrangement of recurrent artificial neural networkscan be useful in circumstances where different parties have different levels of access to information, e.g., in medical record systems where patient identity information may not be accessible to a party that will be using and have access to the remainder of the medical record.
• the same informationcan be input into multiple, different recurrent artificial neural networks.
• the different outputs of those neural networkscan be used, e.g., to ensure that the input can be reconstructed with high fidelity.
• activity within a recurrent artificial neural networkshould match a pattern indicative of ordering, this is not necessarily the case. Rather, in some implementations, activity within a recurrent artificial neural network can comport with a pattern without necessarily displaying activity that matches the pattern. For example, an increase in the likelihood that a recurrent neural network is to display activity that would match a pattern can be treated as non-random ordering of the activity.
• different groups of patternscan be tailored for use in characterizing the activity in different recurrent artificial neural networks.
• the patternscan be tailored, e.g., according to the effectiveness of the patterns in
• Effectivenesscan be quantified, e.g., based on the size of a table or vector that represents the occurrence counts of different patterns.
• the patterns used to characterize the activity in a recurrent artificial neural networkscan consider the strength of a connection between nodes.
• the patterns described previously hereintreat all signal transmission activity between two nodes in a binary manner, i.e., either the activity exists or it doesn’t. This is not necessarily the case. Rather, in some implementations, comporting with a pattern can require activity of a certain level or strength of connection to be taken as indicative of ordered complexity in the activity of a recurrent artificial neural network.
• the content of the output of the recurrent artificial neural networkcan include activity patterns that occur outside windows of time in which the activity in a neural network has a distinguishable level of complexity.
• the output of the recurrent artificial neural network that is read at 1015 and transmitted or stored at 1020 (FIG. 10)can include information encoding activity patterns that occur, e.g., outside dashed rectangles 915, 920, 925 in graph 905 (FIG. 9).
• the output of the recurrent artificial neural networkcould characterize only the highest dimensional patterns of activity, regardless of when those patterns of activity occur.
• the output of the recurrent artificial neural networkcould characterize only patterns of activity that enclose cavities, regardless of when those patterns of activity occur.
• FIGS. 12, 13, and 14are schematic illustrations of a binary form or representation 1200 of topological structures such as, e.g., patterns of activity in a neural network.
• the topological structures illustrated in FIGS. 12, 13, and 14all include the same information, namely, an indication the presence or absence of features in a graph.
• the featurescan be, e.g., activity in a neural network device.
• the activityis identified based on or during periods of time in which the activity in the neural network has a complexity that is distinguishable from other activity that is responsive to an input.
• binary representation 1200includes bits 1205, 1207, 1211, 1293, 1294,
• bits 1205, 1207, 1211, 1293, 1294, 1297 ...are illustrated as discrete rectangular shapes that are either filled or unfilled to indicate the binary value of the bit.
• representation 1200superficially appears to be either a one- dimensional vector of bits (FIGS. 12, 13) or a two-dimensional matrix of bits (FIG. 14).
• representation 1200differs from a vector, from a matrix, or from other ordered collection of bits in that the same information can be encoded regardless of the order of the bits— i.e., regardless of the location of individual bits within the collection.
• each individual bit 1205, 1207, 1211, 1293, 1294, 1297 ...can represent the presence or absence of a topological feature— regardless of the location of that feature in the graph.
• a bitsuch as bit 1207 can indicate the presence of a topological feature that comports with pattern 505 (FIG. 5), regardless of whether that activity occurs between nodes 104, 105, 101 or between nodes 105, 101, 102.
• each individual bit 1205, 1207, 1211, 1293, 1294, 1297, ...can be associated with a particular feature, the location of that feature in the graph need not be encoded, e.g., by a corresponding location of the bit in representation 1200.
• representation 1200may only provide an isomorphic topological reconstruction of the graph.
• the location of individual bits 1205, 1207, 1211, 1293, 1294, 1297, ...does indeed encode information such as, e.g., the location of a feature in the graph.
• the source graphcan be reconstructed using representation 1200.
• such an encodingis not necessarily present.
• bit 1205appears before bit 1207, which appears before bit 1211 at the start of representation 1200.
• bit 1211appears before bit 1211 at the start of representation 1200.
• FIGS. 2 and 3the order of bits 1205, 1207, and 1211 within representation 1200— and the position of bits 1205, 1207, and 1211 relative to other bits within
• binary representation 1200has changed. Nevertheless, binary representation 1200 remains the same— as does the set of rules or algorithm that defines the process for encoding information in binary representation 1200. So long as the correspondence between the bit and the feature is known, the location of the bits in the representation 1200 is irrelevant.
• each bit 1205, 1207, 1211, 1293, 1294, 1297 ...individually represents the presence or absence of a feature in a graph.
• a graphis a set of nodes and a set of edges between those nodes.
• the nodescan correspond to objects. Examples of objects can include, e.g., artificial neurons in a neural network, individuals in a social network, or the like. Edges can correspond to some relation between the objects. Examples of relations include, e.g., a structural connection or activity along the connection. In the context of a neural network, artificial neurons can be related by a structural connection between neurons or by transmission of information along a structural connection.
• edgescan thus characterize relatively long-lived structural characteristics of the set of nodes or relatively transient activity characteristics that occur within a defined time frame. Further, edges can either be directed or bidirectional. Directed edges indicate directionality of the relation between the objects. For example, the transmission of information from a first neuron to a second neuron can be represented by a directed edge that denotes the direction of
• a relational connectionmay indicate that second user is to receive information from the first but not that the first is to receive information from the second.
• a graphcan be expressed as a set of unit intervals [0, 1] where 0 and 1 are identified with respective nodes that are connected by an edge.
• bits 1205, 1207, 1211, 1293, 1294, 1297can be, e.g., a node, a set of nodes, a set of sets of nodes, a set of edges, a set of sets of edges, and/or additional hierarchically-more-complex features (e.g., a set of sets of sets of nodes).
• Bits 1205, 1207, 1211, 1293, 1294, 1297generally represent the presence or absence of features that are at different hierarchical levels. For example, bit 1205 may represent the presence or absence of a node, whereas bit 1205 may represent the presence or absence of a set of nodes.
• bits 1205, 1207, 1211, 1293, 1294, 1297may represent features in a graph that have a threshold level of some characteristic.
• bits 1205, 1207, 1211, 1293, 1294, 1297can represent not only that there is activity in a set of edges, but also that this activity is weighted either above or below a threshold level.
• the weightscan, e.g., embody the training of a neural network device to a particular purpose or can be an innate characteristic of the edges.
• FIGS. 5, 6, and 8 aboveillustrate features whose presence or absence can be represented by bits 1205, 1207, 1211, 1293, 1294, 1297 ...
• the directed simplices in collections 500, 600, 700treat functional or structural graphs as a topological space with nodes as points. Structure or activity involving one or more nodes and links that comports with simplices in collection 500, 600, 700 can be represented in a bit regardless of the identity of the particular nodes and/or links that participate in the activity.
• 4-dimensional simplex pattern 515inherently includes structure or activity that comports with the four point, 3 -dimensional and three point, 2-dimension simplex patterns 510, 505.
• points 0, 2, 3, 4 and points 1, 2, 3, 4 in 4- dimensional simplex pattern 515 of FIG. 5both comport with 3-dimensional simplex pattern 510.
• simplex patterns that include fewer points— and hence are of a lower dimension—can be discarded or otherwise ignored.
• bits 1205, 1207, 1211, 1293, 1294, 1297may not be independent of one another.
• bits 1205, 1207, 1211, 1293, 1294, 1297represent the presence or absence of OD-simplices that each reflect the existence or activity of a single node, then bits 1205, 1207, 1211, 1293, 1294, 1297 are independent of one another.
• bits 1205, 1207, 1211, 1293, 1294, 1297represent the presence or absence of higher-dimensional simplices that each reflect the existence or activity of multiple nodes, then the information encoded by the presence or absence of each individual feature may not be independent of the presence or absence of the other features.
• FIG. 15schematically illustrates an example of how the presence or absence of features that correspond to different bits are not independent of one another.
• a subgraph 1500that includes four nodes 1505, 1510, 1515, 1520 and six directed edges 1525, 1530, 1535, 1540, 1545, 1550 is illustrated.
• edge 1525is directed from node 1525 to node 1510
• edge 1530is directed from node 1515 to node 1505
• edge 1535is directed from node 1520 to node 1505
• edge 1540is directed from node 1520 to node 1510
• edge 1545is directed from node 1515 to node 1510
• edge 1550is directed from node 1515 to node 1520.
• a single bit in representation 1200may indicate the presence of a directed 3D-simplex.
• a bitcould indicate the presence of the 3D-simplex formed by nodes 1505, 1510, 1515, 1520 and edges 1525, 1530, 1535, 1540, 1545, 1550.
• a second bit in representation 1200e.g., filled bit 1293 in FIGS.
• bit 1293may indicate the presence of a directed 2D-simplex.
• a bitcould indicate the presence of the 2D-simplex formed by nodes 1515, 1505, 1510 and edges 1525, 1530, 1545.
• the information encoded by bit 1293is
• the information encoded by bit 1293may also be redundant with the information encoded by still further bit.
• the information encoded by bit 1293would be redundant with both a third bit and a fourth bit that indicates the presence of additional directed 2D-simplices. Examples of those simplices are formed by nodes 1515, 1520, 1510 and edges 1540, 1545, 1550 and nodes 1520, 1505, 1510 and edges 1525, 1535,
• FIG. 16schematically illustrates another example of how the presence or absence of features that correspond to different bits are not independent of one another.
• a subgraph 1600that includes four nodes 1605, 1610, 1615, 1620 and five directed edges 1625, 1630, 1635, 1640, 1645 is illustrated.
• 1635, 1640, 1645generally correspond to nodes 1505, 1510, 1515, 1520 and edges 1525, 1530, 1535, 1540, 1545 in subgraph 1500 (FIG. 15). However, in contrast with subgraph 1500 in which nodes 1515, 1520 are connected by edge 1550, nodes 1615, 1620 are not connected by an edge.
• a single bit in representation 1200may indicate the absence of a directed 3D-simplex, such as, e.g., the directed 3D-simplex that encompasses nodes 1605, 1610, 1615, 1620.
• a second bit in representation 1200(e.g., filled bit 1293 in FIGS. 12, 13, 14) may indicate the presence of a 2D-simplex.
• An example directed 2D-simplexis formed by nodes 1615, 1605, 1610 and edges 1625, 1630, 1645.
• This combination of a filled bit 1293 and an unfilled bit 1205provides information indicative of presence or absence of other features (and the state of other bits) that may or may not be present in representation 1200.
• the combination of the absence of a directed 3D-simplex and the presence of a directed 2D-simplexindicates that at least one edge is absent from either:
• a representation 1200that includes a collection of bits that corresponds only to, e.g., the presence or absence of 3D-simplices is possible.
• the encoding of informationis fault tolerant and provides “graceful degradation” of the encoded information.
• the loss of a particular bit (or group of bits)may increase the uncertainty as to the presence or absence of a feature.
• estimates of the likelihood that a feature is present or absentwill still be possible from the other bits that indicate the presence or absence of adjacent features.
• the ordering or arrangement of bitsis irrelevant to isomorphic reconstruction of the graph that is represented by the bit. All that is required is a known correspondence between the bits and particular nodes/structures in the graph.
• the patterns of activity in a neural networkcan be encoded in a representation 1200 (FIGS. 12, 13, and 14).
• the patterns of activity in a neural networkare a result of a number of characteristics of the neural network such as, e.g., the structural connections between nodes of the neural network, the weights between nodes, as well as a whole host of possible other parameters.
• the neural networkcould have been trained prior to the encoding of the patterns of activity in representation 1200.
• the responsive pattern of activitycan be thought of a“representation” or an
• representation 1200can appear to be a straightforward-appearing collection of (in some cases, binary) digits, each of the digits can encode the relationship or correspondence between a particular input and relevant activity in the neural network.
• FIGS. 17, 18, 19, 20are schematic illustrations of the use of representations of the occurrence of topological structures in the activity in a neural network in four different classification systems 1700, 1800, 1900, 2000.
• Classification systems 1700, 1800each classify representations of the patterns of activity in a neural network as part of the classification of input.
• Classification systems 1900, 2000each classify approximations of representations of the patterns of activity in a neural network as part of the classification of input.
• the patterns of activity that are representedoccur in and are read from a source neural network device 1705 that is part of the classification system 1700, 1800.
• classification systems 1900, 2000the patterns of activity that are approximately represented occur in a source neural network device that is not part of the classification system 1700, 1800. Nevertheless, the approximation of the representation of those patterns of activity are read from an approximator 1905 that is part of classification systems 1900, 2000.
• classification system 1700includes a source neural network 1705 and a linear classifier 1710.
• Source neural network 1705is a neural network device that is configured to receive an input and present representations of the occurrence of topological structures in the activity within source neural network 1705.
• source neural network 1705includes an input layer 1715 that receives the input.
• input layer 1715that receives the input.
• Source neural network 1705can be any of variety of different types of neural network.
• source neural network 1705is a recurrent neural network such as, e.g., a recurrent neural network that is modelled on a biological system. In some cases, source neural network 1705 can model a degree of the morphological, chemical, and other characteristics of a biological system. In general, source neural network 1705 is
• classification system 1700will generally be a dispersed system in which a remote classifier 1710 communicates with source neural network 1705, e.g., via a data communications network.
• source neural network 1705can be untrained and the activity that is represented can be the innate activity of source neural network 1705. In other implementations, source neural network 1705 can be trained and the activity that is represented can embody this training.
• the representations read from source neural network 1705can be representations such as representation 1200 (FIGS. 12, 13, 14).
• the representationscan be read from source neural network 1705 in a number of ways.
• source neural network 1705includes“reader nodes” that read patterns of activity between other nodes within source neural network 1705.
• the activity within source neural network 1705are read by a data processing component of that is programmed to monitor source neural network 1705 for relatively highly-ordered patterns of activity.
• source neural network 1705can include an output layer from which representation 1200 can be read, e.g., when source neural network 1705 is
• Linear classifier 1710is a device that classifies an object— namely, representations of the patterns of activity in source neural network 1705— based on a linear combination of the object’s characteristics.
• Linear classifier 1710includes an input 1720 and an output 1725.
• Input 1720is coupled to receive representations of the patterns of activity in source neural network 1705.
• the representations of the patterns of activity in source neural network 1705is a feature vector that represents the characteristics of the input into source neural network 1705 that are used by linear classifier 1710 to classify that input.
• Linear classifier 1710can receive the representations of the patterns of activity in source neural network 1705 in a variety of ways.
• the representations of the patterns of activitycan be received as discrete events or as a continuous stream over a real time or non- real time communication channel.
• Output 1725is coupled to output the classification result from linear classifier 1710.
• output 1725is schematically illustrated as a parallel port with multiple channels. This is not necessarily the case.
• output 1725can output classification result over a serial port or a port with combined parallel and serial capabilities.
• linear classifier 1710can be implemented on one or more computing devices with relatively limited computational performance.
• linear classifier 1710can be implemented on a personal computer or a mobile computing device such as a smart phone or tablet.
• classification system 1800includes source neural network 1705 and a neural network classifier 1810.
• Neural network classifier 1810is a neural network device classifies an object— namely, representations of the patterns of activity in source neural network 1705— based on a non-linear combination of the object’s characteristics.
• neural network classifier 1810is a feedforward network that includes an input layer 1820 and an output layer 1825.
• neural network classifier 1810can receive the representations of the patterns of activity in source neural network 1705 in a variety of ways. For example, the representations of the patterns of activity can be received as discrete events or as a continuous stream over a real time or non- real time communication channel.
• neural network classifier 1810can perform inferences on one or more computing devices with relatively limited computational performance.
• neural network classifier 1810can be implemented on a personal computer or a mobile computing device such as a smart phone or tablet, e.g., in a Neural Processing Unit of such a device.
• classification system 1800will generally be a dispersed system in which a remote neural network classifier 1810 communicates with source neural network 1705, e.g., via a data communications network.
• neural network classifier 1810can be, e.g., a deep neural network such as a convolutional neural network that includes convolutional layers, pooling layers, and fully-connected layers.
• Convolutional layerscan generate feature maps, e.g., using linear convolutional filters and/or nonlinear activation functions. Pooling layers reduce the number of parameters and control overfitting.
• the computations performed by the different layers in image classifier 1820can be defined in different ways in different implementations of image classifier 1820.
• classification system 1900includes source approximator 1905 and a linear classifier 1710.
• source approximator 1905is a relatively simple neural network that is trained to receive input— either at an input layer 1915 or elsewhere— and output a vector that approximates a representation of topological structures that arise in the patterns of activity in a relatively more complex neural network.
• source approximator 1905can be trained to approximate a recurrent source neural network such as, e.g., a recurrent neural network that is modelled on a biological system and includes a degree of the morphological, chemical, and other characteristics of a biological system.
• source approximator 1905includes an input layer 1915 and an output layer 1920.
• Input layer 1915is couplable to receive the input data.
• Output layer 1920is coupled to output an approximation of a representation of the activity within a neural network device for receipt by input 1720 of linear classifier.
• output layer 1920can output an approximation 1200’ of representation 1200 (FIGS. 12, 13, 14).
• the representation 1200 schematically illustrated in FIGS. 17, 18 and the approximation 1200’ of representation 1200 schematically illustrated in FIGS. 19, 20are identical. This is for the sake of convenience only. In general, approximation 1200’ will differ from
• linear classifier 1710can classify approximation 1200’.
• source approximator 1905can perform inferences on one or more computing devices with relatively limited computational performance.
• source approximator 1905can be implemented on a personal computer or a mobile computing device such as a smart phone or tablet, e.g., in a Neural Processing Unit of such a device.
• classification system 1900will generally be housed in a single housing, e.g., with source approximator 1905 and linear classifier 1710 implemented on either the same data processing devices or on data processing devices coupled by a hardwired connection.
• classification system 2000includes source approximator 1905 and a neural network classifier 1810.
• Output layer 1920 of source approximator 1905is coupled to output an approximation 1200’ of a representation of the activity within a neural network device for receipt by input 1820 of neural network classifier 1810.
• neural network classifier 1810can classify approximation 1200’.
• classification system 1900will generally be housed in a single housing, e.g., with source approximator 1905 and neural network classifier 1810 implemented on either the same data processing devices or on data processing devices coupled by a hardwired connection.
• FIG. 21is a schematic illustration of an edge device 2100 that includes a local artificial neural network that can be trained using representations of the occurrence of topological structures corresponding to activity in a source neural network.
• a local artificial neural networkcan be, e.g., an artificial neural network that is executed entirely on one or more local processors that do not require a communications network to exchange data.
• the local processorswill be connected by hardwire connections.
• the local processorscan be housed in a single housing, such as a single personal computer or a single handheld, mobile device. In some instances, the local processors can be under the control of and accessible by a single individual or a limited number of individuals. In effect, by using a representation of the occurrence of topological structures in a more complex source neural network to train (e.g., using a supervised learning or reinforcement learning technique) a simpler and/or less highly trained but more exclusive second neural network, even individuals with limited computing resources and limited numbers of training samples can train a neural network as desired. Storage requirements and computational complexity during training are reduced and resources like battery lifespan are spared.
• edge device 2100is schematically illustrated as a security-camera device that includes an optical imaging system 2110, image processing electronics 2115, a source approximator 2120, a representation classifier 2125, and a communications controller and interface 2130.
• Optical imaging system 2110can include, e.g., one or more lenses (or even a pinhole) and a CCD device.
• Image processing electronics 2115can read the output of optical imaging system 2110 and in general can perform basic image processing functions.
• Communications controller and interface 2130is a device that is configured to control the flow of information to and from device 2100. As discussed further below, among the operations that
• communications controller and interface 2130can perform are the transmission of images of interest to other devices and the receipt of training information from other devices.
• Communications controller and interface 2130can thus include both a data transmitter and a receiver that can communicate over, e.g., a data port 2135.
• Data port 2135can be a wired port, a wireless port, an optical port, or the like.
• Source approximator 2120is relatively simple neural network that is trained to output a vector that approximates a representation of topological structures that arise in the patterns of activity in a relatively more complex neural network. For example, source
• approximator2l20can be trained to approximate a recurrent source neural network such as, e.g., a recurrent neural network that is modelled on a biological system and includes a degree of the morphological, chemical, and other characteristics of a biological system.
• a recurrent source neural networksuch as, e.g., a recurrent neural network that is modelled on a biological system and includes a degree of the morphological, chemical, and other characteristics of a biological system.
• Representation classifier 2125is either a linear classifier or a neural network classifier that is coupled to receive an approximation of a representation of the patterns of activity in a source neural network from source approximator 2120 and output a
• Representation classifier 2125can be, e.g., a deep neural network such as a convolutional neural network that includes convolutional layers, pooling layers, and fully-connected layers. Convolutional layers can generate feature maps, e.g., using linear convolutional filters and/or nonlinear activation functions. Pooling layers reduce the number of parameters and control overfitting. The computations performed by the different layers in representation classifier 2125 can be defined in different ways in different implementations of representation classifier 2125.
• optical imaging system 2110can generate raw digital images.
• Image processing electronics 2115can read the raw images and will generally perform at least some basic image processing functions.
• Source approximator 2120can receive images from image processing electronics 2115 and perform inference operations to output a vector that approximates a representation of topological structures that arise in the patterns of activity in a relatively more complex neural network. This
• approximation vectoris input into representation classifier 2125 which determines whether the approximation vector satisfies one or more sets of classification criteria. Examples include facial recognition and other machine vision operations. In the event that
• representation classifier 2125determines that the approximation vector satisfies a set of classification criteria, representation classifier 2125 can instruct communications controller and interface 2130 to transmit information regarding the images. For example,
• communications controller and interface 2130can transmit the image itself, the
• communications controller and interface 2130can receive a training set.
• the training setcan include raw or processed image data
• Such a training setcan be used to retrain source approximator 2120, e.g., using a supervised learning or reinforcement learning technique.
• the representationsare used as the target answer vectors and represent the desired result of source approximator 2120 processing the raw or processed image data.
• the training setcan include representations of topological structures that arise in the patterns of activity in a relatively more complex neural network and the desired classification of those representations of topological structures.
• Such a training setcan be used to retrain a neural network representation classifier 2125, e.g., using a supervised learning or reinforcement learning technique.
• the desired classificationare used as the target answer vectors and represent the desired result of representation classifier 2125 processing the representations of topological structures.
• inference operations at device 2100can be facilely adapted to changing circumstances and objectives without large sets of training data and time- and computing power-intensive iterative training.
• FIG. 22is a schematic illustration of a second edge device 2200 that includes a local artificial neural network that can be trained using representations of the occurrence of topological structures corresponding to activity in a source neural network.
• second edge device 2200is schematically illustrated as a mobile computing device such as smart phone or a tablet.
• Device 2200includes an optical imaging system (e.g., on the backside of device 2200, not shown), image processing electronics 2215, a representation classifier 2225, a communications controller and interface 2230, and a data port 2235. These components can have characteristics and perform actions that correspond to those of optical imaging system 2110, image processing electronics 2115, representation classifier 2125, communications controller and interface 2130, and data port 2135 in device 2100 (FIG. 21).
• the illustrated implementation of device 2200additionally includes one or more additional sensors 2240 and a multi-input source approximator 2245.
• Sensor(s) 2240can sense one of more characteristics of the environment surrounding device 2200 or of device 2200 itself.
• sensor 2240can be an accelerometer that senses the acceleration to which device 2200 is subject.
• sensor 2240can be an acoustic sensor such as a microphone that senses noise in the environment of device 2200.
• Still further examples of sensor 2240include chemical sensors (e.g.,“artificial noses” and the like), humidity sensors, radiation sensors, and the like.
• sensor 2240is coupled to processing electronics that can read the output of sensor 2240 (or other information such as, e.g., a contact list or map) and perform basic processing functions.
• processing electronicscan read the output of sensor 2240 (or other information such as, e.g., a contact list or map) and perform basic processing functions.
• sensor 2240can thus have different“modalities” in that the physical sensed physical parameter changes from sensor to sensor.
• Multi-input source approximator 2245is a relatively simple neural network that is trained to output a vector that approximates a representation of topological structures that arise in the patterns of activity in a relatively more complex neural network.
• multi-input source approximator 2245can be trained to approximate a recurrent source neural network such as, e.g., a recurrent neural network that is modelled on a biological system and includes a degree of the morphological, chemical, and other characteristics of a biological system.
• multi-input source approximator 2245is coupled to receive raw or processed sensor data from multiple sensors and return an approximation of a representation of topological structures that arise in the patterns of activity in a relatively more complex neural network based on that data.
• multi-input source approximator 2245can receive processed image data from image processing electronics 2215 as well as, e.g., acoustic, acceleration, chemical, or other data from one or more sensors 2240.
• Multi-input source approximator 2245can be, e.g., a deep neural network such as a convolutional neural network that includes convolutional layers, pooling layers, and fully- connected layers.
• the computations performed by the different layers in multi-input source approximator 2245can be dedicated to a single type of sensor data or to sensor data of multiple modalities.
• multi-input source approximator 2245is to return the approximation based on raw or processed sensor data from multiple sensors. For example, processed image data from image processing electronics 2215 and acoustic data from a microphone sensor 2240 can be used by multi-input source approximator 2245 to approximate a representation of topological structures that would arise in the patterns of activity in a relatively more complex neural network that received the same data.
• communications controller and interface 2230can receive a training set.
• the training setcan include raw or processed image, sounds, chemical or other data and representations of topological structures that arise in the patterns of activity in a relatively more complex neural network.
• Such a training setcan be used to retrain multi- input source approximator 2245, e.g., using a supervised learning or reinforcement learning technique.
• the representationsare used as the target answer vectors and represent the desired result of multi-input source approximator 2245 processing the raw or processed image or sensor data.
• the training setcan include representations of topological structures that arise in the patterns of activity in a relatively more complex neural network and the desired classification of those representations of topological structures.
• Such a training setcan be used to retrain a neural network representation classifier 2225, e.g., using a supervised learning or reinforcement learning technique.
• the desired classificationare used as the target answer vectors and represent the desired result of representation classifier 2225 processing the representations of topological structures.
• inference operations at device 2200can be facilely adapted to changing circumstances and objectives without large sets of training data and time- and computing power-intensive iterative training.
• FIG. 23is a schematic representation of a system 2300 in which local neural networks can be trained using representations of the occurrence of topological structures corresponding to activity in a source neural network.
• the target neural networksare implemented on relatively simple, less expensive data processing systems whereas the source neural network can implemented on a relatively complex, more expensive data processing system.
• System 2300includes a variety of devices 2305 with local neural networks, a telephone base station 2310, a wireless access point 2315, a server system 2320, and one or more data communications networks 2325.
• Local neural networks devices 2305are devices that are configured to process data using computationally-less-intensive target neural networks. As illustrated, local neural networks devices 2305 can be implemented as mobile computing devices, cameras, automobiles, or any of a legion of other appliances, fixtures, and mobile components, as well as different makes and models of devices within each category. Different local neural networks devices 2305 can belong to different owners. In some implementations, access to the data processing functionality of local neural networks devices 2305 will generally be restricted to these owners and/or the owner’s designates.
• Local neural networks devices 2305can each include one or more source
• the relatively more complex neural networkcan be a recurrent source neural network such as, e.g., a recurrent neural network that is modelled on a biological system and includes a degree of the morphological, chemical, and other characteristics of a biological system.
• local neural networks devices 2305can also be programmed to re-train the source approximators using representations of topological structures that arise in the patterns of activity in a relatively more complex neural network as the target answer vectors.
• local neural networks devices 2305can be programmed to perform one or more iterative training techniques (e.g., gradient descent or stochastic gradient descent).
• the source approximators in local neural networks devices 2305are trainable by, e.g., a dedicated training system or by a training system that is installed on a personal computer that can interact with the local neural networks devices 2305 to train source approximators.
• Each local neural networks device 2305includes one or more wireless or wired data communication components.
• each local neural networks device 2305includes at least one wireless data communication components such as a mobile phone transceiver, a wireless transceiver, or both.
• the mobile phone transceiversare able to exchange data with phone base station 2310.
• the wireless transceiversare able to exchange data with a wireless access point 2315.
• Each local neural networks device 2305may also be able to exchange data with peer mobile computing devices.
• Phone base station 2310 and wireless access point 2315are connected for data communication with one or more data communication networks 2325 and can exchange information with a server system 2320 over the network(s).
• Local neural networks devices 2305are thus generally also in data communication with server system 2320. However, this is not necessarily the case. For example, in implementations where local neural networks devices 2305 are trained by other data processing devices, local neural networks devices 2305 need only be in data communication with these other data processing devices at least once.
• Server system 2320is a system of one or more data processing devices that is programmed to perform data processing activities in accordance with one or more sets of machine-readable instructions. The activities can include serving training sets to training systems for mobile computing devices 2305. As discussed above, the training systems can be internal to mobile local neural networks devices 2305 themselves or on one or more other data processing devices.
• the training setscan include representations of the occurrence of topological structures corresponding to activity in a source neural network and corresponding input data.
• server system 2320also includes the source neural network. However, this is not necessarily the case and server system 2320 may receive the training sets from yet another system of data processing device(s) that implement the source neural network.
• server system 2320can serve the training set to trainers that train mobile computing devices 2305.
• the source approximators in target local neural networks devices 2305can be trained using the training set so that the target neural networks approximate the operations of the source neural network.
• FIGS. 24, 25, 26, 27are schematic illustrations of the use of representations of the occurrence of topological structures in the activity in a neural network in four different systems 2400, 2500, 2600, 2700.
• Systems 2400, 2500, 2600, 2700can be configured to perform any of a number of different operations.
• systems 2400, 2500, 2600, 2700can perform object localization operations, object detection operations, object segmentation operations, object detection operations, prediction operations, action selection operations, or the like.
• Object localization operationslocate an object within an image. For example, a bounding box can be constructed around an object.
• object localizationcan be combined with object recognition, in which the localized object is labeled with an
• Object detection operationsclassify image pixels as either belonging to a particular class (e.g., belonging to an object interest) or not.
• object detectionis performed by grouping pixels and forming bounding boxes around the pixel groups. The bounding box should be a tight fit around the object.
• Object segmentationgenerally assigns class labels to each image pixel. Thus, rather than a bounding box, object segmentation proceeds on a pixel-by-pixel basis and generally requires that only a single label be assigned to each pixel.
• Prediction operationsseek to draw conclusions that are outside the range of a observed data. Although prediction operations can seek to forecast future occurrences (e.g., based on information about the past and current state), prediction operations can also seek to draw conclusions about the past and current state based on incomplete information on those states.
• Action selection operationsseek to choose an action based on a set of conditions. Action selection operations have traditionally be broken down into different approaches such a symbol-based systems (classical planning), distributed solutions, and reactive or dynamic planning.
• Classification systems 2400, 2500each perform a desired operation on
• Systems 2600, 2700each perform a desired operation on approximations of representations of the patterns of activity in a neural network.
• systems 2400, 2500the patterns of activity that are represented occur in and are read from a source neural network device 1705 that is part of the system 2400, 2500.
• system 2400includes a source neural network 1705 and a linear processor 2410.
• Linear processor 2410is a device that performs operations based on a linear combination of the characteristics of representations of the patterns of activity in a neural network (or approximations of such representations).
• the operationscan be, e.g., object localization operations, object detection operations, object segmentation operations, object detection operations, prediction operations, action selection operations, or the like.
• Linear processor 2410includes an input 2420 and an output 2425.
• Input 2420is coupled to receive representations of the patterns of activity in source neural network 1705.
• Linear processor 2410can receive the representations of the patterns of activity in source neural network 1705 in a variety of ways.
• the representations of the patterns of activitycan be received as discrete events or as a continuous stream over a real time or non- real time communication channel.
• Output 2525is coupled to output the processing result from linear processor 2410.
• linear processor 2410can be implemented on one or more computing devices with relatively limited computational performance.
• linear processor 2410can be implemented on a personal computer or a mobile computing device such as a smart phone or tablet.
• system 2400includes a source neural network 1705 and a linear processor 2410.
• Linear processor 2410is a device that performs operations based on a linear combination of the characteristics of representations of the patterns of activity in a neural network (or approximations of such representations).
• the operationscan be, e.g., object localization operations, object detection operations, object segmentation operations, prediction operations, action selection operations, or the like.
• Linear processor 2410includes an input 2420 and an output 2425.
• Input 2420is coupled to receive representations of the patterns of activity in source neural network 1705.
• Linear processor 2410can receive the representations of the patterns of activity in source neural network 1705 in a variety of ways.
• the representations of the patterns of activitycan be received as discrete events or as a continuous stream over a real time or non- real time communication channel.
• Output 2525is coupled to output the processing result from linear processor 2410.
• linear processor 2410can be implemented on one or more computing devices with relatively limited computational performance.
• linear processor 2410can be implemented on a personal computer or a mobile computing device such as a smart phone or tablet.
• classification system 2500includes source neural network 1705 and a neural network 2510.
• Neural network 2510is a neural network device that is configured to perform operations based on a non-linear combination of the characteristics of representations of the patterns of activity in a neural network (or approximations of such representations). The operations can be, e.g., object localization operations, object detection operations, object segmentation operations, prediction operations, action selection operations, or the like.
• neural network 2510is a feedforward network that includes an input layer 2520 and an output layer 2525. As with linear processor 2410, neural network 2510 can receive the representations of the patterns of activity in source neural network 1705 in a variety of ways.
• neural network 2510can perform inferences on one or more computing devices with relatively limited computational performance.
• neural network 2510can be implemented on a personal computer or a mobile computing device such as a smart phone or tablet, e.g., in a Neural Processing Unit of such a device.
• system 2500will generally be a dispersed system in which a remote neural network 2510 communicates with source neural network 1705, e.g., via a data communications network.
• neural network 2510can be, e.g., a deep neural network such as a convolutional neural network.
• system 2600includes source approximator 1905 and a linear processor 2410. Despite any differences between approximation 1200’ representation 1200, processor 2410 can still perform operations on approximation 1200’.
• system 2700includes source approximator 1905 and a neural network
• neural network 2510can still perform operations on approximation 1200’.
• systems 2600, 2700can be implemented on an edge device, such as, e.g., edge devices 2100, 2200 (FIGS. 21, 22).
• systems 2600, 2700can be implemented as part of a system in which local neural networks can be trained using representations of the occurrence of topological structures corresponding to activity in a source neural network, such as system 2300 (FIG. 23).
• FIG. 28is a schematic illustration of a reinforcement learning system 2800 that includes an artificial neural network that can be trained using representations of the occurrence of topological structures corresponding to activity in a source neural network.
• Reinforcement learningis a type of machine learning in which an artificial neural network learns from feedback regarding the consequences of actions taken in response to the artificial neural network’s decisions.
• a reinforcement learning systemmoves from one state in the environment to another by performing actions and receiving information characterizing the new state and a reward and/or regret that characterizes the success (or lack of success) of the action.
• Reinforcement learningseek to maximize the total reward (or minimize the regret) through a learning process.
• the artificial neural network in reinforcement learning system 2800is a deep neural network 2805 (or other deep learning architecture) that is trained using a reinforcement learning approach.
• deep neural network 2805can be a local artificial neural network (such as neural network 2510 (FIGS.
• deep neural network 2805can be implemented on a system of networked devices.
• reinforcement learning system 2800includes an actuator 2810, one or more sensors 2815, and a teacher module 2820. In some implementations, reinforcement learning system 2800 also includes one or more sources 2825 of additional data.
• Actuator 2810is a device that controls a mechanism or a system that interacts with an environment 2830.
• actuator 2810controls a physical mechanism or system (e.g., the steering of an automobile or the positioning of a robot).
• actuator 2810can control a virtual mechanism or system (e.g., a virtual game board or an investment portfolio).
• environment 2830may also be physical or virtual.
• Sensor(s) 2815are devices that measure characteristics of the environment 2830. At least some of the measurements with characterize interactions between the controlled mechanism or system and other aspects of the environment 2830. For example, when actuator 2810 steers an automobile, sensor(s) 2815 may measure one or more of the speed, direction, and acceleration of the automobile, the proximity of the automobile to other features, and the response of other features to the automobile. As another example, when actuator 2810 controls an investment portfolio, sensor(s) 2815 may measure the value and risk associated with the portfolio.
• both source approximator 1905 and teacher module 2820are coupled to receive at least some of the measurements made by sensor(s) 2815.
• source approximator 1905can receive measurement data at input layer 1915 and output an approximation 1200’ of a representation of topological structures that arise in the patterns of activity in a source neural network.
• Teacher module 2820is a device that is configured to interpret the measurements received from sensor(s) 2815 and provide a reward and/or a regret to deep neural network 2805. Rewards are positive and indicate successful control of the mechanism or system.
• Regretsare negative and indicate unsuccessful or less than optimal control.
• teacher module 2820also provides a characterization of the measurements along with the reward/regret for reinforcement learning. In general, the characterization of the
• teacher module 2820may read approximations 1200’ output from source approximator 1905 and pair the read approximations 1200’ with corresponding reward/regret values.
• reinforcement learningdoes not occur in system 2800 in real time or during active control of actuator 2810 by deep neural network 2805. Rather, training feedback can be collected by teacher module 2820 and used for reinforcement training when deep neural network 2805 is not actively instructing actuator 2810.
• teacher module 2820can be remote from deep neural network 2805 and only in intermittent data communication with deep neural network 2805. Regardless of whether reinforcement learning is intermittent or continuous, deep neural network 2805 can be evolved, e.g., to optimize reward and/or reduce regret using the information received from teacher module 2820.
• system 2800also includes one or more sources 2825 of additional data.
• Source approximator 1905can also receive data from data sources 2825 at input layer 1915. In these instances, approximation 1200’ will result from processing both sensor data and the data from data sources 2825.
• the data collected by one reinforcement learning system 2800can be used for training or reinforcement learning of other systems, including other reinforcement learning systems.
• the characterization of the measurements along with the reward/regret valuescan be provided by teacher module 2820 to a data exchange system that collects such data from a variety of reinforcement learning systems and redistributes it among them.
• the characterization of the measurementscan be an approximation of a representation of topological structures that arise in the patterns of activity in a source neural network, such as approximation 1200’.
• reinforcement learning system 2800The particular operations that are performed by reinforcement learning system 2800 will of course depend on the particular operational context. For example, in contexts where source approximator 1905, deep neural network 2805, actuator 2810, and sensors 2815 are part of an automobile, deep neural network 2805 can perform object localization and/or detection operations while steering the automobile.
• reward/regret values and approximations 1200’that characterize the state of the environment when object localizations and/or detection operations were performed can be provided to the data exchange system.
• the data exchange systemcan then distribute the reward/regret values and approximations 1200’ to other reinforcement learning systems 2800 associated with other vehicles for reinforcement learning at those other vehicles.
• reinforcement learningcan be used to improve object localization and/or detection operations at a second vehicle using the reward/regret values and approximations 1200’.
• the operations that are learned at other vehiclesneed not be identical to the operations that are performed by deep neural network 2805.
• reward/regret values that are based on travel time and approximations 1200’ that result from the input of sensor data characterizing, e.g., an unexpectedly wet road at a location identified by a GPS data source 2825can be can be used for route planning operations at another vehicle.
• Embodiments of the subject matter and the operations described in this specificationcan be implemented in digital electronic circuitry, or in computer software, firmware, or hardware, including the structures disclosed in this specification and their structural equivalents, or in combinations of one or more of them.
• Embodiments of the subject matter described in this specificationcan be implemented as one or more computer programs, i.e., one or more modules of computer program instructions, encoded on computer storage medium for execution by, or to control the operation of, data processing apparatus.
• program instructionscan be encoded on an
• a computer storage mediumcan be, or be included in, a computer-readable storage device, a computer-readable storage substrate, a random or serial access memory array or device, or a combination of one or more of them.
• a computer storage mediumis not a propagated signal, a computer storage medium can be a source or destination of computer program instructions encoded in an artificially-generated propagated signal.
• the computer storage mediumcan also be, or be included in, one or more separate physical components or media (e.g., multiple CDs, disks, or other storage devices).
• the operations described in this specificationcan be implemented as operations performed by a data processing apparatus on data stored on one or more computer-readable storage devices or received from other sources.
• the term“data processing apparatus”encompasses all kinds of apparatus, devices, and machines for processing data, including by way of example a programmable processor, a computer, a system on a chip, or multiple ones, or combinations, of the foregoing
• the apparatuscan include special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application- specific integrated circuit).
• the apparatuscan also include, in addition to hardware, code that creates an execution environment for the computer program in question, e.g., code that constitutes processor firmware, a protocol stack, a database management system, an operating system, a cross-platform runtime environment, a virtual machine, or a combination of one or more of them.
• the apparatus and execution environmentcan realize various different computing model infrastructures, such as web services, distributed computing and grid computing infrastructures.
• a computer program(also known as a program, software, software application, script, or code) can be written in any form of programming language, including compiled or interpreted languages, declarative or procedural languages, and it can be deployed in any form, including as a stand-alone program or as a module, component, subroutine, object, or other unit suitable for use in a computing environment.
• a computer programmay, but need not, correspond to a file in a file system.
• a programcan be stored in a portion of a file that holds other programs or data (e.g., one or more scripts stored in a markup language document), in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files that store one or more modules, sub-programs, or portions of code).
• a computer programcan be deployed to be executed on one computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected by a communication network.
• the processes and logic flows described in this specificationcan be performed by one or more programmable processors executing one or more computer programs to perform actions by operating on input data and generating output.
• the processes and logic flowscan also be performed by, and apparatus can also be implemented as, special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application-specific integrated circuit).
• processors suitable for the execution of a computer programinclude, by way of example, both general and special purpose microprocessors, and any one or more processors of any kind of digital computer.
• a processorwill receive instructions and data from a read-only memory or a random access memory or both.
• the essential elements of a computerare a processor for performing actions in accordance with instructions and one or more memory devices for storing instructions and data.
• a computerwill also include, or be operatively coupled to receive data from or transfer data to, or both, one or more mass storage devices for storing data, e.g., magnetic, magneto-optical disks, or optical disks.
• mass storage devicesfor storing data, e.g., magnetic, magneto-optical disks, or optical disks.
• a computerneed not have such devices.
• a computercan be embedded in another device, e.g., a mobile telephone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a Global Positioning System (GPS) receiver, or a portable storage device (e.g., a universal serial bus (USB) flash drive), to name just a few.
• PDApersonal digital assistant
• GPSGlobal Positioning System
• USBuniversal serial bus
• Devices suitable for storing computer program instructions and datainclude all forms of non-volatile memory, media and memory devices, including by way of example semiconductor memory devices, e.g., EPROM, EEPROM, and flash memory devices;
• magnetic diskse.g., internal hard disks or removable disks; magneto-optical disks; and CD-ROM and DVD-ROM disks.
• the processor and the memorycan be supplemented by, or incorporated in, special purpose logic circuitry.
• embodiments of the subject matter described in this specificationcan be implemented on a computer having a display device, e.g., a CRT
• a computercan interact with a user by sending documents to and receiving documents from a device that is used by the user; for example, by sending web pages to a web browser on a user’s client device in response to requests received from the web browser.
• representation 1200is a binary representation in which each bit individually represents the presence or absence of a feature in a graph
• representations of informationare possible.
• a vector or matrix of multi-valued, non-binary digitscan be used to represent, e.g., the presence or absence of features and possibly other characteristics of those features.
• An example of such a characteristicis the weight of edges with the activity that constitutes the features.

Landscapes

• Engineering & Computer Science (AREA)
• Theoretical Computer Science (AREA)
• Physics & Mathematics (AREA)
• Software Systems (AREA)
• Evolutionary Computation (AREA)
• Data Mining & Analysis (AREA)
• Computing Systems (AREA)
• General Engineering & Computer Science (AREA)
• General Physics & Mathematics (AREA)
• Mathematical Physics (AREA)
• Computational Linguistics (AREA)
• Artificial Intelligence (AREA)
• Health & Medical Sciences (AREA)
• Life Sciences & Earth Sciences (AREA)
• Biomedical Technology (AREA)
• Biophysics (AREA)
• General Health & Medical Sciences (AREA)
• Molecular Biology (AREA)
• Robotics (AREA)
• Image Analysis (AREA)
• Mobile Radio Communication Systems (AREA)
• Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
• Multi Processors (AREA)

Applications Claiming Priority (6)

Publications (1)

Family

ID=66776339

Family Applications (5)

Family Applications Before (2)

Family Applications After (2)

Country Status (5)

Cited By (10)

Families Citing this family (10)

Citations (1)

Family Cites Families (13)

• 2019
  • 2019-06-05KRKR1020207035843Apatent/KR102497238B1/koactiveActive
  • 2019-06-05CNCN201980053465.4Apatent/CN112567387B/zhactiveActive
  • 2019-06-05WOPCT/EP2019/064593patent/WO2019238483A1/ennot_activeCeased
  • 2019-06-05EPEP19728962.2Apatent/EP3803699A1/deactivePending
  • 2019-06-06KRKR1020207035847Apatent/KR102526132B1/koactiveActive
  • 2019-06-06CNCN201980053141.0Apatent/CN112567389B/zhactiveActive
  • 2019-06-06CNCN201980054063.6Apatent/CN112585621A/zhactivePending
  • 2019-06-06CNCN201980053463.5Apatent/CN112567390A/zhactivePending
  • 2019-06-06WOPCT/EP2019/064776patent/WO2019238523A1/ennot_activeCeased
  • 2019-06-06KRKR1020207035845Apatent/KR102475411B1/koactiveActive
  • 2019-06-06KRKR1020207035844Apatent/KR102488042B1/koactiveActive
  • 2019-06-06TWTW108119813Apatent/TWI822792B/zhactive
  • 2019-06-06KRKR1020207035846Apatent/KR102465409B1/koactiveActive
  • 2019-06-06WOPCT/EP2019/064741patent/WO2019238513A1/ennot_activeCeased
  • 2019-06-06CNCN201980053140.6Apatent/CN112567388B/zhactiveActive
  • 2019-06-06EPEP19728992.9Apatent/EP3803707A1/deactivePending
  • 2019-06-06EPEP19728989.5Apatent/EP3803705A1/deactivePending
  • 2019-06-06EPEP19728990.3Apatent/EP3803706A1/deactivePending
  • 2019-06-06WOPCT/EP2019/064740patent/WO2019238512A1/ennot_activeCeased
  • 2019-06-06WOPCT/EP2019/064773patent/WO2019238522A1/ennot_activeCeased
  • 2019-06-06EPEP19728993.7Apatent/EP3803708A1/deactivePending

Patent Citations (1)

Non-Patent Citations (3)

Also Published As

Similar Documents

Legal Events