GENERATING AND IDENTIFYING FUNCTIONAL SUBNETWORKS WITHIN STRUCTURAL NETWORKS

Abstract

In one aspect, a method includes generating a functional subgraph of a network from a structural graph of the network. The structural graph comprises a set of vertices and structural connections between the vertices. Generating the functional subgraph includes identifying a directed functional edge of the functional subgraph based on presence of structural connection and directional communication of information across the same structural connection.

Description

RELATED APPLICATIONS

This application is a continuation of U.S. patent application Ser. No. 15/864,146, now allowed, filed 8 Jan. 2018, which claims the benefit of U.S. Provisional Application 62/443,071 filed 6 Jan. 2017, the entire contents of which are incorporated herein by reference.

TECHNICAL FIELD

This invention relates to neural networks, and more particularly to methods and tools for generating and identifying functional subnetworks within structural networks.

BACKGROUND

In simple neural networks that have a limited number of vertices and structural connections between the vertices, the function of the neural network can be understood by a determined analysis of the topology and the weights in the neural networks.

In neural networks of increased complexity, such analyses become untenable. Even if the complete network topology is known, human oversight is lost and the function of subnetworks inscrutable.

A neural network device is a device that mimics the information encoding and other processing capabilities of networks of biological neurons using a system of interconnected nodes. A neural network device can be implemented in hardware, in software, or in combinations thereof.

A neural network device includes a plurality of nodes that are interconnected by a plurality of structural links. Nodes are discrete information processing constructs that are analogous to neurons in biological networks. Nodes generally process one or more input signals received over one or more of links to produce one or more output signals that are output over one or more of links. For example, in some implementations, nodes can 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. In some implementations, nodes can operate as accumulators, e.g., in accordance with an integrate-and-fire model.

Structural links are connections that are capable of transmitting signals between nodes. In some implementations, structural links are bidirectional links that convey a signal from every first node to a second node in the same manner as a signal is conveyed from the second to the first. However, this is not necessarily the case. For example, in a neural network, some portion or all of structural links can be unidirectional links that convey a signal from a first of nodes to a second of nodes without conveying signals from the second to the first. As another example, in some implementations, structural links can have diverse properties other than or in addition to directionality. For example, in some implementations, different structural links can carry signals of different magnitudes—resulting in a different strengths of interconnection between respective of nodes. As another example, different structural links can carry different types of signal (e.g., inhibitory and/or excitatory signals). Indeed, in some implementations, structural links can 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.

SUMMARY

Methods and tools for generating and characterizing functional subnetworks of a neural network are described. The tools include a definition of functional edges that—at a particular period of time—constitute a functional subnetwork of a neural network. The definition is directional in that it specifies a direction of information propagation between vertices. The subnetworks are temporal constructs in that different functional subnetworks exit at different times during the operation of the neural network. In effect, the definition is a composite definition that it defines both the structural characteristics and functional characteristics of subnetworks within a neural network at different periods in time.

In a first aspect, a method includes providing a plurality of structural connections between vertices in a neural network, assigning a direction of information flow to respective of the structural connections, generating a first functional subgraph of the neural network, and generating a second functional subgraph of the neural network. The first functional subgraph includes a first proper subset of the structural connections, wherein vertices connected by the first proper subset of the structural connections are active during a first period of time. The second functional subgraph includes a second proper subset of the structural connections, wherein vertices connected by the second proper subset are active during a second period of time.

In a second aspect, a method includes generating a functional subgraph of a network from a structural graph of the network, wherein the structural graph comprises a set of vertices and structural connections between the vertices. Generating the functional subgraph comprises identifying a directed functional edge of the functional subgraph based on presence of structural connection and directional communication of information across the same structural connection.

In a third aspect, a method is suitable for characterizing a neural network that comprises a set of vertices and structural connections between the vertices. The method includes defining a functional subgraph in the neural network using a definition to identify a plurality of functional edges and analyzing the functional subgraph using one or more topological analyses. The definition can include a definition of a class of structural connection between two vertices in the set of vertices, and a definition of a directional response to a first input elicitable from the two vertices that satisfy the definition of the structural connection class.

In a fourth aspect, a method is suitable for distinguishing between inputs to a neural network that comprises a set of vertices and structural connections between the vertices. The method includes inputting a first input to the neural network; characterizing the response of the neural network to the first input using the method of the third aspect; inputting a second input to the neural network; characterizing the response of the neural network to the second input using the method of the third aspect; and distinguishing between the first input and the second input based on results of the respective topological analyses.

In a fifth aspect, for a network that comprises a set of vertices and structural connections between the vertices, a method comprising demarking a proper subset of the vertices and functional edges between the vertices in the subset as a directed subgraph, wherein the functional edges are defined based on a presence of structural connections between the vertices in the subset and directional information communication across the structural connections.

In a sixth aspect, a method of manufacturing a neural network includes any one the first aspect, the second aspect, the third aspect, the fourth aspect, or the fifth aspect.

In a seventh aspect, a method of analyzing performance of a neural network includes any one the first aspect, the second aspect, the third aspect, the fourth aspect, or the fifth aspect.

In an eighth aspect, wherein the second aspect can be used to generate the first functional subgraph and to generate the second functional subgraph in the first aspect.

Each one of the first aspect, the second aspect, the third aspect, the fourth aspect, or the fifth aspect can include one or more of the following features. Only some of the vertices connected by the second proper subset can be included in the first proper subset. The activity during the first period of time and the activity during the second period of time are both responsive to a same input. A plurality of structural connections can be provided by identifying the plurality of structural connection in a pre-existing neural network.

The direction of information flow can be assigned by determining the direction of information in the pre-existing neural network and assigning the direction in accordance with the determined direction. First and the second functional subgraphs of the neural network can be generated by inputting an input into the pre-existing neural network and identifying the first and the second functional subgraph based on the response of the pre-existing neural network to the input.

A first functional subgraph of the neural network can be generated by weighting the structural connections of the first proper subset to achieve a desired activity during the first period of time. A first functional subgraph of the neural network can be generated by training the first proper subset of the structural connections. A plurality of structural connections can be provided by adding structural connections between at least two functional subgraphs of the neural network.

A structural graph can include undirected structural connections between the vertices. A directed functional edge can be identified by requiring a direct structural connection between two vertices and a directional response to a first input elicitable from the two vertices. A directed functional edge can be identified by identifying a directional communication of information by requiring a response to a first input by a first vertex and a subsequent response by a second vertex. A response to the first input can be required to have occurred during a first time period defined with respect to the first input. The first time period can, e.g., by defined as occurring after an initial propagation of the first input through the functional subgraph. The subsequent response can be required to have occurred during a second time period defined with respect to either the first time period or the response to the first input. The second time period can begin immediately after the first time period ends. The second time period can overlaps with the first time period.

The directed functional edge of the functional subgraph can also be identified based on presence of second structural connection and directional communication of information across the second structural connection. The directed functional edge can be identified by identifying a plurality of directed functional edges and classifying the plurality of directed functional edges as the functional subgraph.

A definition of the class of a structural connection can require a direct structural connection between two of the vertices, e.g., a undirected structural connection. A definition of a directional response can require a response to the first input by a first vertex of the two vertices, and a subsequent response by a second vertex of the two vertices. The response to the first input can be required to have occurred during a first time period defined with respect to the first input. For example, the first time period can be defined as occurring after an initial propagation of the first input through the functional subgraph. The subsequent response can be required to have occurred during a second time period defined with respect to either the first time period or the response to the first input. For example, the second time period begins immediately after the first time period ends or the second time period can overlap with the first time period.

A definition of a functional edge can include a definition of a second structural connection between two vertices in the set of vertices. A definition of a functional edge can include a definition of a second directional response to the first input elicitable from the two vertices that satisfy the definition of the second structural connection.

A functional subgraph can be analyzed by determining a simplex count based on the defined functional edges. A simplex count can be determined by determining at least one of a 1-dimensional simplex count and a 2-dimensional simplex count. A functional subgraph can be analyzed by determining Betti numbers for a network that consists of the defined functional edges. Betti numbers can be determined by determining at least one of the first Betti number and the second Betti number. A functional subgraph can be analyzed by determining an Euler characteristic for a network that consists of the defined functional edges.

A first input and second input can be distinguished based on the respective topological analyses by classifying results of the respective topological analyses, e.g., using a probabilistic classifier. The probabilistic classifier can be a Bayesian classifier.

An input to the network can be classified based on a topological analysis of the directed subgraph. For example, a classifier, e.g., a probabilistic classifier, can be applied to one or more topological metrics of the of the directed subgraph. The directional information communication across the structural connections can include information communication from a first vertex of the set to a second vertex of the set.

The details of one or more implementations are set forth in the accompanying drawings and the description below. Other features, objects, and advantages of the invention will be apparent from the description and drawings, and from the claims.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a flowchart of a process for characterizing parameters of a neural network device using topological methods.

FIG. 2 is a schematic representation of different oriented simplices.

FIG. 3 is a schematic representation of an example directed graph and its associated flag complex.

FIG. 4 is a flowchart of a process for distinguishing functional responses to different input patterns fed into a neural network.

DETAILED DESCRIPTION

Network topology is the arrangement of the various elements of a network, including neural network devices. In neural network devices, the elements include nodes and links between the nodes. In some neural network devices, the nodes can be modeled in one or more ways after biological neurons. Further, the links between the nodes can be modeled in one or more ways after the connections between biological neurons, including synapses, dendrites, and axons.

Even though the nodes and links of a neural network device can have a variety of different characteristics, their network topology can be characterized using various topological methods. Topological characterizations of a networks tend to focus on the connections between elements rather than the exact shape of the objects involved.

The topological methods described herein can be used to characterize structural connections between nodes in neural network, functional connections between nodes in neural network, or both. In a neural network, a structural connection between nodes may provide a link between two nodes over which a signal can be transmitted. In contrast, a functional connection reflects the actual transmission of information from one node to the other over the structural connection, i.e., is part of the “functioning” of the neural network. For example, the signal transmission can be part of active information processing and/or information storage by the neural network. In a particular example, a functional connection between two nodes may arise in response to an input and may indicate active participation of those nodes in processing the information content of the input.

Structural characterizations of neural network devices can be used, e.g., in the construction and/or reconstruction of neural networks. Reconstruction of a neural network can include, e.g., copying or mimicking at least some of the structure of a first neural network in a second neural network. For example, a simpler second neural network can recreate a portion of the structure of a more complex first neural network. In some implementations, the first neural network can be a biological neural network and the second neural network can be an artificial neural network, although this is not necessarily the case.

In some implementations, the neural network devices need not be mere reconstructions. Rather, neural network devices can also be constructed or, in effect, “manufactured” ab initio. In some implementations, the characterizations provided by topological methods can provide general characteristics of desirable structure in neural network devices. For example, the topological characterizations may define a desired level of “structuring” or “ordering” of the neural network device.

By way of example, topological characterizations can be used to construct and reconstruct neural networks in which the distribution of directed cliques (directed all-to-all connected subsets) of neurons by size differs significantly from both that in Erdos-Renyi random graphs with the same number of vertices and the same average connection probability and that in more sophisticated random graphs, constructed either by taking into account distance-dependent probabilities varying within and between cortical layers or morphological types of neurons, or according to Peters' Rule. In particular, the neural network devices can include highly prominent motifs of directed cliques of up to eight neurons. For example, in neural networks with approximately 3×10e4 vertices and 8×10e6 edges, the neural networks can incorporate approximately 10e8 3-cliques and 4-cliques, approximately 10e7 5-cliques, approximately 10e5 6-cliques, and approximately 10e3 7-cliques.

As another example, topological methods can be used to construct or reconstruct neural networks in which the Euler characteristic (EC) of the neural networks can be a value on the order of 10e7, indicating a preponderence of directed cliques consisting of an odd number of neurons.

As another example, topological methods can be used to construct or reconstruct neural networks in which the homological dimension of the neural networks is 5, which compares to the homological dimension of at most 4 for random graphs and hence indicates that the neural networks possess a higher degree of organizational complexity than random graphs. The homological dimension of a neural network is the maximum n such that β_n is not equal to zero, wherein β₀, β₁, β₂, . . . are Betti numbers that provide a measure the higher-order organizational complexity of the network by detecting “cyclic” chains of intersecting directed cliques.

The topological methods described herein can also use defined functional patterns of communication across multiple links to characterize the processing activity within neural network devices. The presence of the functional patterns indicates “structuring” or “ordering” of the information flow along the nodes and links of the neural network device under particular circumstances. Such functional characterizations thus characterize neural network devices in ways that are lost when mere structural information is considered.

In some implementations, characterizations of the patterns of communication across multiple links in a neural network device can be used in the construction and/or reconstruction of neural networks. Reconstruction of a neural network can include, e.g., copying or mimicking at least some of the function of a first neural network in a second neural network. For example, a simpler second neural network can recreate a portion of the functioning of a more complex first neural network. In some implementations, the first neural network can be a biological neural network and the second neural network can be an artificial neural network, although this is not necessarily the case.

In some implementations, neural network devices that are characterized based on defined patterns of communication across multiple links need not be mere reconstructions. Rather, the function of neural network devices can be constructed or, in effect, “manufactured” ab initio. In some implementations, the characterizations provided by topological methods can provide general characteristics of the functional behavior of a desirable neural network device. This can be beneficial to, e.g., reduce training time or even provide a partially- or fully-functional neural network device “out of the box.” In other words, the topological characterizations may define a desired level of “structuring” or “ordering” of the information flow within a functioning neural network device. In some implementations, functional sub-networks can be assembled like components, e.g., by adding structural links between different functional sub-networks to achieve desired processing results.

For example, in some implementations, the topological characterizations may define particular functional responses to particular input patterns. For example, a given stimulus can be applied repeatedly to a neural network device and the responsive activity within the neural network device can be measured. The activity can be binned into timesteps and characterized using topological techniques. For example, the topological techniques can be akin to those used to characterize the structure of a neural network device. In some implementations, a transmission-response graph can be generated for each timestep to represent the activity in the neural network device in a manner suited for analysis using topological methods. In particular, in a transmission-response graph, the vertices are the nodes of the neural network device and the edges are the links between the nodes that are active during the timestep leads. The activity can be, e.g., signal transmissions along the link that leads to firing of a connected node. In some cases, the duration of the timesteps and the precise rule for formation of the transmission-response graph for each timestep can be biologically motivated.

FIG. 1 is a flowchart of a process 100 for characterizing parameters of a neural network device using topological methods. In the flowchart, the topological parameters characterized using process 100 are structural and characterize nodes and their links, i.e., without encompassing the activity in the neural network and its function. Such structural parameters can be used, e.g., to construct or reconstruct a neural network device.

As discussed further below, in other implementations, functional rather than structural topological parameters can be characterized. For example, rather than determining a structural adjacency matrix, a functional connectivity matrix can be determined. Nevertheless, topological techniques can be applied to characterize the functional activity of nodes and their links. Such functional parameters can be used, e.g., to construct or reconstruct a neural network device that has desirable processing activity. As another example, such functional parameters can be used to distinguish different inputs into the neural network device.

In the illustrated implementation, process 100 includes computing binary adjacency matrices for a neural network device at 105. An adjacency matrix is a square matrix that can be used to represent a finite graph, e.g., such as a finite graph that itself represents a neural network device. The entries in an adjacency matrix—generally, a binary bit (i.e., a “1” or a “0”)—indicate whether pairs of nodes are structurally linked or not in the graph. Because each entry only requires one bit, a neural network device can be represented in a very compact way.

Process 100 also includes determining associated directed flag complexes at 110. Directed flag complexes are oriented simplicial complexes that encode the connectivity and the direction of orders of the underlying directed graph. In particular, each directed n-clique in the underlying graph corresponds to an oriented (n−1)-simplex in the flag complex and the faces of a simplex correspond to the directed subcliques of its associated directed clique. Associated directed flag complexes can be determined for structural links (e.g., the existence of directional links), for functional links (e.g., directed activity along links), or both.

FIG. 2 is a schematic representation of different oriented simplices 205, 210, 215, 220. Simplex 205 has a dimension 0, simplex 210 has a dimension 1, simplex 215 has a dimension 2, and simplex 220 has a dimension 3. Simplices 205, 210, 215, 220 are oriented simplices in that simplices 205, 210, 215, 220 have a fixed orientation (that is, there is a linear ordering of the nodes). The orientation can embody the structure if the links, the function of the links, or both.

In the figures, the nodes of the neural network device are represented as dots whereas the links between the nodes are represented as lines connecting these dots. The lines include arrows that denote either the direction of structural link or the direction of activity along the link. For the sake of convenience, all links herein are illustrated as unidirectional links, although this is not necessarily the case.

The nodes and links in a neural network device can be treated as vertices and edges in topological methods. The network or subnetwork of nodes and links can be treated as a graph or a subgraph in topological methods. For this reason, the terms are used interchangeably herein.

FIG. 3 is a schematic representation of an example directed graph 305 and its associated flag complex 310. In further detail, a “directed graph” consists of a pair of finite sets (V, E) and a function τ:E→V×V. The elements of the set V are the or “vertices” of , the elements of E are the “edges” of , and the function τ associates with each edge an ordered pair of vertices. The “direction” of a connection or edge e with τ (e)=(v1, v2) is taken to be from τ₁(e)=v₁, the source node or vertex, to τ₂(v)=v₂, the target vertex.

The function τ is required to satisfy the following two conditions.

• • (1) For each e∈E, if τ(e)=(v₁, v₂), then v₁≠v₂, i.e., there are no loops in the graph.
  • (2) The function τ is injective, i.e., for any pair of vertices (v₁, v₂), there is at most one edge directed from v₁ to v₂.

A vertex v∈ is said to be a “sink” if τ₁(e)≠v for all e∈E. A vertex v∈G is said to be a “source’ if τ₂(e)≠v for all e∈E.

A “morphism of directed graphs” from a directed graph =(V, E, τ) to a directed graph ′=(V′, E′, τ′) consists of a pair of set maps α: V→V′ and β: E→E′ such that β takes an edge in with source v₁ and target v₂ to an edge in ′ with source α(v₁) and target α(v₂), i.e., τ′ºβ=(α, α)ºτ. Two graphs and ′ are “isomorphic” if there is morphism of graphs (α, β):→′ such that both α and β are bijections, which can be called an “isomorphism of directed graphs.”

A “path” in a directed graph consists of a sequence of edges (e₁, . . . , e_n) such that for all 1≤k≤n, the target of e_k is the source of e_k+1, i.e., τ₂(e_k)=τ₁(e_k+1). The “length” of the path (e₁, . . . , e_n) is n, i.e., the number of edges of which the path is composed. If, in addition, target of e_n is the source of e₁, i.e., τ₂(e_n)=τ₁(e₁), then (e₁, . . . , e_n) is an “oriented cycle.”

An “abstract oriented simplicial complex” is a collection S of finite, ordered sets with the property that if σ∈S, then every subset τ of σ is also a member of S. A “subcomplex” of an abstract oriented simplicial complex is a sub-collection S′⊆S that is itself an abstract oriented simplicial complex. For the sake of convenience, abstract oriented simplicial complexes are referred to herein as “simplicial complexes.”

The elements of a simplicial complex S are called its “simplices.” A simplicial complex is said to be “finite” if it has only finitely many simplices. If σ∈S, we define the “dimension” of σ, denoted dim(σ), to be |σ|, i.e., the cardinality of the set σ minus one. If σ is a simplex of dimension n, then we refer to σ as an n-simplex of S. The set of all n-simplices of S is denoted S_n. A simplex τ is said to be a face of σ if π is a subset of σ of a strictly smaller cardinality. A “front face” of an n-simplex σ=(v₀, . . . , v_n) is a face τ=(v₀, . . . , v_m) for some m<n. Similarly, a “back face” of σ is a face τ′=(v_i, . . . , v_n) for some 0<i<n. If α=(v₀, . . . , v_n)∈S_n, then the i^th face of σ is the (n−1)-simplex σⁱ obtained from σ by removing the node or vertex v_i.

A simplicial complex gives rise to a topological space by means of the construction known as “geometric realization.” In brief, one associates a point (a standard geometric 0-simplex) with each 0-simplex, a line segment (a standard geometric 1-simplex) with each 1-simplex, a filled-in triangle (a standard geometric 2-simplex) with each 2-simplex, etc., glued together along common faces. The intersection of two simplices in S, neither of which is a face of the other, is a proper subset, and hence a face, of both of them. In the geometric realization this means that the geometric simplices that realize the abstract simplices intersect on common faces, and hence give rise to a well-defined geometric object. A geometric n-simplex is nothing but a (n+1)-clique, canonically realized as a geometric object. An n-simplex is said to be “oriented” if there is a linear ordering on its vertices. In this case the corresponding (n+1)-clique is said to be a “directed (n+1)-clique.”

If S is a simplicial complex, then the union S⁽ⁿ⁾=S_n∪ . . . ∪S₀, which is called the “n-skeleton” of S, is a subcomplex of S. We say that S is “n-dimensional” if S=S⁽ⁿ⁾, and n is minimal with this property. If S is n-dimensional, and k≤n, then the collection S_k∪ . . . ∪S_n is not a subcomplex of S because it is not closed under taking subsets. However if one adds to that collection all the faces of all simplices in S_k∪ . . . ∪S_n, one obtains a subcomplex of S called the “k-coskeleton” of S, which we will denote by S_(k).

Directed graphs such as directed graph 305 give rise to abstract oriented simplicial complexes. Let =(V, E, τ) be a directed graph. The “directed flag complex” associated with G is the abstract simplicial complex S=S(), with S₀=V and whose n-simplices S_n for n≥1 are (n+1)-tuples (v₀, . . . , v_n), of vertices such that for each 0≤i<j≤n, there is an edge (or connection) in from v_i to v_j. Notice that because of the assumptions on τ, an n-simplex in S is characterised by the (ordered) sequence (v₀, . . . , v_n), but not by the underlying set of nodes or vertices. For instance (v₁, v₂, v₃) and (v₂, v₁, v₃) are distinct 2-simplices with the same set of vertices.

Returning to process 100 and FIG. 1, for each node in the neural network device, there is a vertex in the underlying directed graph that is labelled with the unique global identification number (GID). The (j, k)-coefficient of the structural adjacency matrix is a binary “1” if and only if there is a directed connection in the neural network from the node/vertex with GID j to the node/vertex with GID k. The adjacency matrix can thus be referred to as the “structural matrix” of the neural network and the directed flag complex can be referred to as a “neocortical microcircuit complex” or “N-complex.”

Process 100 also includes determining a parameter of a neural network device based on the relevant matrix and/or the N-complex using topological methods at 115. There are several different parameters that can be determined.

For example, the simplices in the neural network in each dimension can simply be counted. Such simplex counts can indicate the degree of “structuring” or “ordering” of the nodes and connections (or the activity) within the neural network device. As another example, the Euler characteristic of all N-complexes can be computed. As yet another example, the Betti numbers of a simplicial complex can be computed. In particular, the n-th Betti number, β_n, counts the number of chains of simplices intersecting along faces to create an “n-dimensional hole” in the complex, which requires a certain degree of organization among the simplices.

FIG. 4 is a flowchart of a process 400 for distinguishing—using topological methods—functional responses to different input patterns fed into a neural network.

Process 400 includes receiving, at the neural network, one or more input patterns at 405. In some implementations, the one or more input patterns can correspond to a known input. For example, process 400 can be part of the training of a neural network, the design of a process for reading the processed output of a neural network, the testing of a neural network, and/or the analysis of an operational neural network. In these cases, known input patterns can be used to confirm that the functional response of the neural network is appropriate. In other implementations, the one or more input patterns can correspond to an unknown input(s). For example, process 400 can be part of the operation of a trained neural network and the functional response of the neural network can represent the processed output of the neural network.

Process 400 also includes dividing the activity in neural network into time bins at 410. A time bin is a duration of time. The total functional activity in a neural network responsive to an input pattern (e.g., signal transmission along edges and/or nodes) can be subdivided according the time in which the activity is observed.

In some implementations, the duration of the time bins can be chosen based on the extent to which activity in each bin is distinguishable when different input patterns are received. Such an ex post analysis can be used, e.g., when designing a process for reading the processed output of a neural network using input patterns that correspond to known inputs. In other words, the process for reading the output of a neural network can be adapted to the observed activity responsive to different input patterns. This can be done, e.g., to ensure that the process for reading the neural network appropriately captures the processing results when input patterns that correspond to unknown inputs are received.

In some implementations, the duration of the time bins can be constant. For example, in neural network devices that are modeled after biological neural networks, the duration of the time bins can be 5 ms.

In some implementations, the time bins can commence after the input pattern is received at the neural network. Such a delay can allow the input pattern to propagate to relevant portions of the neural network device prior to meaningful processing results are expected. Subsequent time bins can be defined with respect to an end of a preceding time bin or with respect to the time after the input pattern is received.

Process 400 also includes recording a functional connectivity matrix for each time bin at 415. A functional connectivity matrix is a measure of the response of each edge or link in a neural network device to a given input pattern during a time bin. In some implementations, a functional connectivity matrix is a binary matrix where active and inactive edges/links are denoted (e.g., with a binary “1” and “0”, respectively). A functional connectivity matrix is thus akin to a structural adjacency matrix except that the functional connectivity matrix captures activity.

In some implementations, an edge or connection can be denoted as “active” when a signal is transmitted along the edge or connection from a transmitting node to a receiving node during the relevant time bin and when the receiving node or vertex responds to the transmitted signal by subsequently transmitting a second signal.

In general, the responsive second signal need not be transmitted within the same time bin as the received signal. Rather, the responsive second signal can be transmitted within some duration after the received signal, e.g., in a subsequent time bin. In some implementations, responsive second signals can be identified by identifying signals that are transmitted by the receiving node within a fixed duration after the receiving node receives the first signal. For example, the duration can be about 1.5 times as long as the duration of a time bin, or about 7.5 ms in neural network devices that are modeled after biological neural networks.

In general, the responsive second signal need not be responsive solely to single signal received by the node. Rather, multiple signals can be received by the receiving node (e.g., along multiple edges or connections). The receiving node or vertex can “respond” to all or a portion of the received signals in accordance with the particularities of the information processing performed by that node. In other words, as discussed above, the receiving node or vertex can, e.g., weight and sum multiple input signals, pass the sum through one or more non-linear activation functions, and output one or more output signals, e.g., as an accumulator, e.g., in accordance with an integrate-and-fire model.

In some implementations, the transmitted signals are spikes and each (j, k)-coefficient in a functional connectivity matrix is denoted as “active” if and only if the following three conditions are satisfied, where s_i^j denotes the time of the i-th spike of node j:

• • (1) The (j, k)-coefficient of the structural matrix is 1, i.e., there is a structural connection from the node or vertex j to the node or vertex k;
  • (2) the node or vertex with GID j spikes in the n-th time bin; and
  • (3) the node or vertex with GID k spikes within an interval after the neuron with GID j.
    In effect, it is assumed that spiking of node or vertex k is influenced by the spiking of node or vertex j.

Process 400 also includes characterizing one or more parameters of the activity recorded in the functional connectivity matrix using topological methods at 420. For example, the activity recorded in the functional connectivity matrix can be characterized using one or more of the approaches used in process 100 (FIG. 1), substituting the functional connectivity matrix for the structural adjacency matrix. For example, a characterization of a topological parameter of the neural network device can be determined for different functional connectivity matrices from different time bins. In effect, the structuring or ordering of the activity in the neural network can be determined at different times.

Process 400 also includes distinguishing the functional response of the neural network device to a received input pattern from other functional responses of the neural network device to other received input patterns based on the characterized topological parameters at 425. As discussed above, in some instances, input patterns that correspond to known inputs can be used in a variety contexts. For example, during testing of a neural network, distinguishing the functional response of the neural network device to a received input pattern from functional responses of the neural network device to other patterns can indicate whether the neural network device is functioning properly. As another example, during training of a neural network, distinguishing the functional response of the neural network device to a received input pattern from functional responses of the neural network device to other patterns can provide an indication that training is complete. As another example, in analyzing an operational neural network, distinguishing the functional response of the neural network device to a received input pattern from functional responses of the neural network device to other patterns can provide an indication of the processing performed by the neural network device.

In instances wherein the input patterns can correspond to unknown inputs, distinguishing the functional response of the neural network device to a received input pattern from functional responses of the neural network device to other patterns can be used to read the output of the neural network device.

Distinguishing functional responses to different input patterns fed into a neural network using topological methods can also be performed in a variety of other contexts. For example, changes in functional responses over time can be used to identify the results of training and/or the structures associated with training.

A number of implementations have been described. Nevertheless, it will be understood that various modifications may be made. Accordingly, other embodiments are within the scope of the following claims.

Claims

1-20. (canceled)

21. A computer-implemented method, comprising:

receiving, at a neural network, a first input;

dividing functional activity in the neural network that is responsive to the first input into first time bins;

recording, for each of the first time bins, a first measure of the functional activity in the neural network device during that first time bin that is responsive to the first input;

characterizing one or more parameters of the recorded first measure of the functional activity using topological methods; and

receiving, at the neural network, a second input;

dividing functional activity in the neural network that is responsive to the second input into second time bins;

recording, for each of the second time bins, a measure of the functional activity in the neural network device during that second time bin that is responsive to the first input;

characterizing one or more parameters of the recorded second measure of the functional activity using topological methods; and

distinguishing the functional response of the neural network to the first input from the functional response of the neural network to the second input based on the characterized topological parameters.

22. The method of claim 21, wherein the topological methods comprise determining associated directed flag complexes.

23. The method of claim 21, wherein the functional activity in the neural network that is responsive to the first and second inputs includes signal transmission along edges of the neural network.

24. The method of claim 21, wherein the duration of the time bins is constant.

25. The method of claim 21, wherein the measure of the functional activity is a functional connectivity matrix.

26. The method of claim 21, wherein the first input and the second input are known inputs.

27. The method of claim 21, wherein the method further comprises determining that neural network device is functioning properly or trained based on the distinguishing of the functional response to the first input from the functional response to the second input.

28. A computer-implemented method, comprising:

receiving, at a first neural network, an input;

dividing functional activity in the neural network that is responsive to the input into time bins;

recording, for each of the time bins, a measure of the functional activity in the neural network device during that time bin that is responsive to the input;

characterizing one or more parameters of the recorded measure of the functional activity using topological methods; and

reconstructing at least some of functioning of the first neural network in a second neural network using the characterizations provided by the topological methods.

29. The method of claim 28, wherein the second neural network is simpler than the first neural network.

30. The method of claim 28, wherein the functioning of the first neural network is reconstructed in the second neural network without training the second neural network.

31. The method of claim 28, wherein the topological methods comprise determining associated directed flag complexes.

32. The method of claim 28, wherein the functional activity in the neural network that is responsive to the input includes signal transmission along edges of the neural network.

33. The method of claim 28, wherein the duration of the time bins is constant.

34. The method of claim 28, wherein the measure of the functional activity is a functional connectivity matrix.

35. The method of claim 28, wherein the method further comprises determining that second neural network is functioning properly or trained based on a functional response of the second neural network to the input pattern.