System, apparatus and method for spiking neural network

Abstract

Systems and associated methods for constructing neural networks (NN) without the benefit of predefined output spike times or predefined network architecture. Functions are constructed from other functions determined from an input construction training set of output variable type. Method aspects of the neural network modeler may comprise partitioning the construction training set, representing the partitions, restricting the representations, constructing subfunctions from the restrictions, and combining the subfunctions to model the target function. Specifically, subfunctions represented from partition-specific neural networks of output value type are created using two fundamental composition operations: same-constant composition and different-constants composition. Different choices of weights and delays lead to different NNs with different output spikes that implement the same function.

Description

BACKGROUND

Supervised learning is an area of the machine learning field directed to construction of a function that maps an input to an output based on example input-output pairs. The function of interest may be inferred from labeled training data consisting of a set of training examples. An artificial neural network (ANN) is a computing system architecture loosely modeled after biological neural networks (e.g., the human brain). The neural network itself is not an algorithm, but rather a framework through which certain machine-learning algorithms may cooperate to process complex data inputs. Much of the work on neural networks (NNs) concerns training a neural network with a training set to compute some function of interest.

A Spiking Neural Network (SNN) is a network of artificial neurons and synapses that computes using as values the times at which neurons fire their spikes. The model of a neuron determines that time. In operation, an SNN is a function that maps input spike times to output spike times. The training set is a subset of the function that has been encoded in spike times. In typical SNN implementations, the output spike times in the training set are chosen from estimates that consider the architecture proposed for the SNN. Error backpropagation algorithms are then used to adjust the weights of the synapses, which entails making repeated passes through the entire network until the errors with respect to the outputs of the training set become acceptably small. Common error backpropagation algorithms frequently employ a large number of sum and sigmoid calculations, which may result in low efficiency and/or high computing overhead when dealing with a large volume of data.

BRIEF DESCRIPTION OF THE DRAWINGS

The accompanying drawings provide visual representations which will be used to more fully describe various representative embodiments and can be used by those skilled in the art to better understand the representative embodiments disclosed and their inherent advantages. The drawings are not necessarily to scale, emphasis instead being placed upon illustrating the principles of the devices, systems, and methods described herein. In these drawings, like reference numerals may identify corresponding elements.

FIG. 1 is a schematic block diagram of a neural network design system, in accordance with embodiments of the disclosure;

FIG. 2 is a schematic diagram of single-valued relation, in accordance with embodiments of the disclosure;

FIG. 3 is a schematic diagram of a function containing a training set, in accordance with embodiments of the disclosure;

FIG. 4 is a schematic diagram and associated equation for same-constant composition, in accordance with embodiments of the disclosure;

FIG. 5 is a schematic diagram and associated equation for different-constant composition, in accordance with embodiments of the disclosure;

FIG. 6A is a schematic diagram of a neural network (NN);

FIGS. 6B and 6C are schematic diagrams of spiking neural network (SNN) configurations illustrating single and multiple synapses, respectively, between presynaptic and postsynaptic neurons;

FIG. 6D is a table illustrating spike times for unary basic SNNs (bSNN);

FIG. 6E is a definition of the XOR and associated constant functions;

FIG. 6F is schematic diagrams of implementations of spike delay in a bSNN;

FIG. 6G is tables illustrating spike times for bSNN;

FIG. 6H is schematic diagrams of bSNNs implementing the constant functions of FIG. 6E;

FIG. 7 is a schematic diagram illustrating designating and interfering outputs for neural network components, in accordance with embodiments of the disclosure;

FIG. 8 is a schematic diagram for a same-constant composition construct, in accordance with embodiments of the disclosure;

FIGS. 9A and 9B are schematic diagrams for different-constant composition constructs, in accordance with embodiments of the disclosure;

FIG. 10 is a flowchart for a method of compositional construction of a SNN, in accordance with embodiments of the disclosure;

FIG. 11 is a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure;

FIG. 12 is a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure;

FIG. 13 is a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure;

FIG. 14 is a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure;

FIG. 15 is a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure;

FIG. 16 is a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure;

FIG. 17 is a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure;

FIG. 18 is a table illustrating exemplary parameter values for neural network composition constructs according to embodiments of the disclosure;

FIG. 19 is a table illustrating exemplary parameter values for neural network composition constructs according to embodiments of the disclosure;

FIG. 20 is a table illustrating exemplary parameter values for neural network composition constructs according to embodiments of the disclosure;

FIGS. 21A and 21B are a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure;

FIG. 22 is a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure;

FIG. 23 is a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure;

FIG. 24 is a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure;

FIG. 25 is a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure;

FIG. 26 is a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure;

FIG. 27 is a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure;

FIG. 28 is a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure;

FIG. 29 is a table illustrating exemplary parameter values for neural network composition constructs according to embodiments of the disclosure;

FIG. 30 is a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure;

FIG. 31 is a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure;

FIG. 32 is a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure;

FIG. 33 is a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure; and

FIG. 34 is a schematic diagram for exemplary neural network composition constructs in accordance with embodiments of the disclosure.

DETAILED DESCRIPTION

Specific embodiments of the disclosure will now be described in detail with reference to the accompanying figures. While this disclosure is susceptible of being embodied in many different forms, there is shown in the drawings and will herein be described in detail specific embodiments, with the understanding that the present disclosure is to be considered as an example of the principles and not intended to be limited to the to the specific embodiments shown and described. In the description below, like reference numerals may be used to describe the same, similar or corresponding parts in the several views of the drawings.

Generally, the systems and methods disclosed herein may be characterized as a supervised learning solution, because the disclosed designs may utilize an input training set. As described above, traditional supervised learning approaches typically involve a) choosing an architecture of the neural network, b) providing the entire training set to the chosen architecture, and c) repeatedly adjusting parameters until errors become acceptably small. The disclosed method first partitions an input training set and then constructs neural networks for these smaller subfunctions (i.e. subsets of the training set). The component neural networks may then be combined into larger neural networks.

More specifically, the present disclosure relates to systems and associated methods of employing compositional construction of a neural network to yield a parametric network model architecture. The design systems and methods described herein advantageously exploit properties of the Spike-Response Model, taking as a guiding principle the property that defines the notion of a function, single-valued. NNs that compute larger functions are constructed from NNs that compute smaller functions, yielding parametric architectures. As such, the disclosed design supports transformations of an NN that preserve the function it computes: certain transformations may map to different output spike times, while others may preserve the output spike times. The present design method is scalable.

For example, and without limitation, embodiments of present disclosure may comprise a software-implemented process, stored on a computer-readable medium, such as a tangible computer-readable medium and executed by one or more computing devices, or processors. Referring to FIG. 1, in more detail, a neural network design system 200 (also referred to as a neural network modeling system or simply as a modeling system, in accordance with embodiments of the disclosure) may comprise a computing device 202 having a processor 212 that may be operable to access, accept and/or execute computerized instructions, and also a data store 213 which may access and/or store data and instructions used by the processor 212. More specifically, the processor 212 may be positioned in data communication with some number of the networked devices and may be configured to direct input from such networked devices to the data store 213 for storage and subsequent retrieval. For example, and without limitation, the processor 212 may be in data communication with external computing resources, such as the Internet 210, through a network interface 219. User interface 230 is operatively coupled to processor 212.

The computerized instructions of the neural network design system 200 may be configured to implement a Modeling Subsystem 220 that may be stored in the data store 213 and accessed and/or retrieved by the processor 212 for execution. The Modeling Subsystem 220 may be operable to access or receive and/or partition a training set, and then construct neural networks for these smaller subfunctions (i.e. subsets of the training set) as described in more detail hereinafter.

Those skilled in the art will appreciate that the present disclosure contemplates the use of computer instructions and/or systems configurations that may perform any or all of the operations involved in neural network design. The disclosure of computer instructions that include Modeling Subsystem 220 instructions is not meant to be limiting in any way. Those skilled in the art will readily appreciate that stored computer instructions and/or systems configurations may be configured in any suitable manner. Those skilled in the art also will understand that the principles of the present disclosure may be implemented on or in data communication with any type of suitably arranged device or system configured to perform neural network construction operations, in any combination.

Neural network composition may exploit certain features of the Spike-Response Model class of supervised learning solutions. For example, and without limitation, an SNN is a function that maps input spike times to output spike times, as described above. The concept of a function or map is defined by the single-valued property. A function is a single-valued relation.

Referring now to FIG. 2, for example, and without limitation, a partial function ƒ 300 maps each element of the domain A to at most one element of the codomain B. An element a∈A that ƒ does not map to an element in B is said to be undefined, which may be written as ƒ(a)↑. If mapped to some b∈B, a is said to be defined, written ƒ(a)↓. A partial function is said to be total if, for each a∈A, ƒ(a)↓.

To implement a function ƒ: A→B as an SNN the elements of set A and set B must be encoded as spike times. Thus, an SNN S that implements fencodes the distinct v, v′∈B with different spike times. In certain embodiments of the present design, this encoding is determined during the construction of S.

Referring now to FIG. 4, for example, and without limitation, a construction of S may employ two subfunctions of ƒ, each a respective constant function that evaluates to v. These two functions are shown as SNNs S_v^t1 and S_v^t2. One encodes v by t₁, and the other by t₂. Same-constant composition 500 may construct an SNN S_v that implements the union of the two (disjoint) constant functions, ensuring that there is a unique spike time to represent v. Referring now to FIG. 5, for example, and without limitation, different-constants composition 600 may construct from given SNNs S_v^t1 and S_v′^t2 which implement constant functions that evaluate to v and v′, where v≠v′, an SNN S_vv′, ensuring that distinct spike times encode v and v′. Methods of exploiting the properties of a common model of a neuron, and the properties of some foundational SNNs, to define these composition operations are described in detail hereinbelow.

As described generally above, spiking neural networks (SNN) consist of neurons that fire spikes, and of synapses that connect neurons and transmit spikes. More specifically, a neuron has a membrane potential that may be affected by the spikes it receives from neighboring neurons. A neuron is also characterized by a threshold, which may be tunable. A neuron fires a spike when its membrane potential crosses the threshold from below. Each synapse may also be characterized by a weight w and/or a delay d. Parameters of a neuron and its incoming synapses, the threshold of the neuron and the weights and delays of the synapses, may be chosen either to ensure that the neuron will spike or to prevent it from spiking.

As described generally above, the times at which neurons fire spikes are the values that SNNs use in computations. For example, a neuron may be defined as having a membrane potential x, which varies over time, and a threshold θ. When the membrane potential crosses the threshold from below, the neuron fires a spike. FIG. 6B and FIG. 6C illustrate SNNs 110 and 120, respectively, as collections of neurons connected by synapses, represented by circles and arrows, respectively. The neuron before the synapse is presynaptic 112 to the one after the synapse, and the one after, is postsynaptic 115 to the one before. When a presynaptic neuron 112 fires, a spike is transmitted through synapses to its postsynaptic neurons 115. Neurons may be connected by zero, one or more synapses.

As spikes are transmitted through synapses, these spikes affect the membrane potential of the postsynaptic neuron 115. For example, and without limitation, in the Spike-Response Model, a spike-response function E captures the basic effect that a spike fired by a presynaptic neuron 112 causes on the membrane potential of the postsynaptic neuron 115. This model defines the membrane potential of a neuron as follows:

$x\left(t\right)=\sum _{i=1}^m\sum _{k=1}^{K_i}{w}_{ik}\epsilon \left(t-t_i-d_{ik}\right)$ Presynaptic neuron n_i fires a spike at time t_i, and its K_i synapses transmit it to neuron n, each according to its weight w_ik and delay d_ik.

The membrane potential x and the spike-response E functions have as domain the set of nonnegative real numbers ≥0, which is used to represent time, and as codomain the set of real numbers, used to represent the membrane potential of the neuron:x: ^≥0→, ε: ^≥0→

Simulations of real-time systems often discretize time and examine the model of the system at times determined by the length of a step A, the time step. Thus, the membrane potential also can be expressed as a function of discretized time, where the time at step n∈ is t_n=nΔ, and delay d_ik=d_ikΔ, for some d_ik ∈.

The membrane potential of a neuron as a function of discretized time is defined as follows:

${\hat{x}}^{-}:\mathbb{N}\to ℝ$ $x^{-}\left(n\right)=\sum _{i=1}^m\sum _{k=1}^{K_i}{w}_{ik}\epsilon \left(\left(n-n_i-d_{ik}^{-}\right)\Delta \right)$

The Shift Law may contribute substantively to the compositional method of constructing SNNs described herein and may be stated as follows: Shifting all inputs by n^I time steps, shifts the membrane potential by n^I time steps, also. Mathematically, the Shift Law may be described as follows:

$\begin{array}{c}{x}^{-}\left(n\right)=\sum _{i=1}^m\sum _{k=1}^{K_i}{w}_{ik}\epsilon \left(\left(n-n_i-d_{ik}^{-}\right)\Delta \right)\\ \Rightarrow \sum _{i=1}^m\sum _{k=1}^{K_i}{w}_{ik}\epsilon \left(\left(\left(n+n^{\prime }\right)-\left(n_i+n^{\prime }\right)-d_{ik}^{-}\right)\Delta \right)\\ =x^{-}\left(n+n^{\prime }\right)\end{array}$

A condition in ensuring the properties necessary for composition of SNNs is that a neuron spike at most once. The SNNs described hereinbelow possess this and other properties that describe how spikes propagate through an SNN.

A unary basic spiking neural network (bSNN) consists of one input neuron, one output neuron, and one synapse between them. The synapse may be characterized by a weight w and a delay d. As described generally above, choice of the synaptic parameters w, d, and threshold θ may make it possible to ensure that an input spike i will trigger a spike in the output neuron at some time o. When the transmission of a spike through a synapse is not immediate, a delay d≥1. Given a spike-response function ε, a time step of length Δ, a synaptic delay d, and a synaptic weight w, the SNN just described is denoted by w^ε,Δ,d; when ε and Δ are known, by w^d; and when d=1, simply by w.

For a unary bSNN, only one presynaptic neuron is present for purposes of the membrane potential defined above. Some weight values might not produce a spike on the output neuron. For example, if the maximum value in the range of E is 1.0 and θ=1.0, any weight w<1.0 would prevent a spike in the output neuron. Thus, in this case w is not a total function. For weights that allow spikes, such as w≥1.0 for the example above, bSNN w maps 0 to the time step at which the membrane potential of the output neuron crosses the threshold θ. This time step is denoted by sw and called the spike time of w. Thus, for a unary bSNN w that is total: x(sw)=θ.

The spike time sw^d of a unary bSNN w^d is the time step at which the postsynaptic neuron fires when the presynaptic neuron fires at time step 0. For non-total bSNN w^d, sw^d↑.

FIG. 6D illustrates a table 130 showing spike times 134 for a few unary bSNNs 132 exhibiting the following parameters:ε(t)=t/δe^{1-t/δ, Δ=}0.03, d=1, and approximation error 0.05.

From the definition of membrane potential, and the Shift Law, it follows that the time at which the output neuron of a unary bSNN fires a spike may be delayed in several ways (as illustrated in FIG. 6F, basic firing 142 may be shifted by delaying 144 the input, by increasing the delay 146 of the synapse, or by vertically composing 148 the bSNN with another unary bSNN, whose output becomes the input of the one whose output is to be delayed). Thus, the spike time of a total unary bSNN is a time length that characterizes that bSNN. Because a given time length may be achieved in different ways, replacing a sequence of unary bSNNs by another one with the same time length transforms an SNN while preserving its output spike times.

For a unary bSNN w the input value for a complete description of the graph of the bSNN is 0. For any other input, the output may be obtained by the Shift Law. For a pair of inputs (x, y), the Shift Law suggests that it be reduced to one of the two forms below:c(x,y)=(0,y−x)[x] if x≤y c(x,y)=(x−y,0)[y] if y≤x

If an SNN maps (0, y−x) to n, then it maps (x, y) to n+x, and if it maps (x−y, 0) to n, then it maps (x, y) to n+y. The pair in the canonical representation of (x, y) may be referred to as its base, and the natural number in square brackets, its shift.

A binary basic spiking neural network (bSNN) consists of two input neurons and one output neuron, with each input neuron connected by a respective, or associated, single synapse to the output neuron. Similar to the unary bSNN described above, choices of synaptic parameters cause input spikes to trigger an output spike. For example, a binary bSNN consists of two input neurons n₁ and n₂; one output neuron n; and a single synapse of weight w_i connecting n_i to n, for i=1, 2. A binary bSNN w₁^d1, w₂^d2 maps the pair of time steps at which the input neurons fire their spikes to the time step at which the output neuron fires its spike. The delays are omitted below, unless these delays are needed for some result. Thus, bSNN w₁, w₂ has the following type:w₁,w₂: ×→.

To determine the behavior of w₁, w₂ (i.e., its graph), a simulation need only consider bases of the canonical representations of pairs of natural numbers. A pair (0, n) indicates that n₁ fires its spike at (time step) 0, and n₂ fires its spike later at (time step) n. This representation may be written as (0, n)∈[0]×, where [0] is the singleton set containing 0. A pair (n, 0)∈×[0], in turn, indicates that n₂ fires its spike at 0, and n₁ fires its spike at n. The graph of a binary bSNN w₁, w₂ is represented canonically by two functions, which have the following types:w₁,w₂⁰′: [0]×→w₁,w₂′⁰: ×[0]→

FIG. 6G shows the graph, as represented in 152 and 154 of bSNN 3.5, 4.5. The earliest spiking time occurs when both n₁ and n₂ fire their spikes at 0. When one fires its spike at 0 and the other is delayed, the output spike time is generally later. If the delay is long enough it allows the earlier input spike to be the sole cause of the output spike. The synaptic connection of n₁ has weight 3.5, and s3.5=23. If n₂ fires its spike at 23 or later, the output neuron either will be firing its spike concurrently or will have fired it already, and the spike from n₂ will have no effect on the output spike time. So, the output spike time remains constant when n₂ fires its spike at s3.5 or later. For the same reason, if n₂ fires its spike at 0, and n₁ at s(4.5)=17 or later, n will fire at 17, and n₁ has no effect on the spike time of the output neuron.

Letting sw₁, w₂ denote the output spike time for inputs (0, 0), and [n, m], the interval of naturals between n and m, inclusive, the pattern of graph of total binary bSNNs is as follows:

Let Δ be sufficiently small not to miss spikes, and unary bSNNs w₁ and w₂ be total. The graph of binary bSNN w₁, w₂ has the following pattern:

$\left[0\right]\times \left[0,s\langle w_1\rangle \right]\stackrel{{\langle w_1,w_2\rangle }^{,0}}{\to }\left[s\langle w_1,w_2\rangle ,s\langle w_1\rangle \right]$ $\forall \left(0,n\right)·n\ge s\langle w_1\rangle \Rightarrow \left(0,n\right)\mapsto s\langle w_1\rangle$ $\left[0,s\langle w_2\rangle \right]\times \left[0\right]\stackrel{{\langle w_1,w_2\rangle }^{,0}}{\to }\left[s\langle w_1,w_2\rangle ,s\langle w_2\rangle \right]$ $\forall \left(m,0\right)·n\ge s\langle w_2\rangle \Rightarrow \left(n,0\right)\mapsto s\langle w_2\rangle$

Referring now to FIG. 6A, a neural network such as an SNN learns to compute a function ƒ: I→O. In general, inputs and outputs of ƒ may be vectors or tuples:(i₁, . . . ,i_n)∈I, for some n≥1(o₁, . . . ,o_m)∈O, for some m≥1

A first element in this kind of training, a training set, is a finite subfunction of ƒ: ƒ_T. A finite function may be expressed as a set of pairs, each mapping an input to its corresponding output. The training set, it follows, is a finite function:ƒ_T: I_T→O I_T⊆I ƒ_T(i₁, . . . ,i_n)=f(i₁, . . . ,i_n) for (i₁, . . . ,i_n)∈I_T If |I_T|=N, then ƒ_T may be written as a set of input-output pairs of size N as follows:(i₁₍₁₎, . . . ,i_n(1))(o₁₍₁₎, . . . ,o_m(1))(i₁₍₂₎, . . . ,i_n(2))(o₁₍₂₎, . . . ,o_m(2))(i_1(N), . . . ,i_n(N))(o_1(N), . . . ,o_m(N))

As exemplified in FIG. 6A, the neural network 100 may receive the inputs 102 from the training set and, for each input, may produce an output 105. As shown, the neural network computes the following:(i₁₍₁₎, . . . ,i_n(1))(u₁₍₁₎, . . . ,u_m(1))(i₁₍₂₎, . . . ,i_n(2))(u₁₍₂₎, . . . ,u_m(2))(i_1(N), . . . ,i_n(N))(u_1(N), . . . ,u_m(N))

From the discrepancies between the target outputs to be learned, the o's, and the actual outputs 105 of the neural network, the u's, some error measure is defined. A common error measure is

$E=\frac{1}{2}{\sum }_{k=1}^m{\left(u_k-o_k\right)}^2.$ To reduce this error, adjustments may be made to parameters of the neural network being trained. For example, error backpropagation is commonly used to modify parameters of the neural network that determine the outputs that the network produces. In this error-correction method, certain input neurons receive some inputs 102 which may be modified by some parameters according to some activation function. The u's 105 that the neural network produces are a complex function that combines the activation functions of the constituents of the network. Backpropagation involves derivatives and activation functions are differentiable. The term backpropagation refers to the order in which this technique adjusts the parameters; beginning from the output neurons and propagating the adjustments through the entire network until they reach the input neurons.

An initial benchmark when evaluating learning algorithms is the Boolean XOR function, denoted herein with the symbol ⊕, such that⊕: {T,F}→{T,F}

• • T⊕T=F
  • T⊕F=T
  • F⊕T=T
  • F⊕F=F

A neural network may represent the values T and F in some way. For example, to represent input Boolean values, T may be represented by 0 milliseconds (ms) and F by 6 ms; and to represent output Boolean values, T may be represented by 10 ms and F by 16 ms. One may apply the notation above: ƒ: I→O as follows:I={0;6}O={10;16}

To distinguish between the function eventually implemented in (i.e., learned by) the neural network versus the machine-independent function, different fonts may be applied herein, as follows:

• • function learned by neural net: ƒ: I→O⊕: {0;6}→{10;16}
  • machine-independent function: f: I→O⊕: {T,F}→{T,F}

The Boolean XOR function, with binary operator ⊕, is used herein to illustrate an approach for training and constructing SNNs by operating upon input and variable output, referred to herein as a construction training set. The spike times that encode Boolean values as inputs are chosen arbitrarily at the beginning of the process. The encoding of the Boolean values as outputs is determined by the construction of the SNN. The domain of XOR is of size four.

One way of designing an SNN that computes the XOR function is to compose four SNNs that compute constant functions: two that evaluate to T, and two that evaluate F, according to the definition of XOR (see FIG. 6E). FIG. 6E shows XOR function 162 and four constant function 164.

For example, and without limitation, let an input T be encoded by spike time 0, and input F, by spike time 20. FIG. 6H illustrates four binary bSNNs 172, 174, 176 and 178 arbitrarily chosen to implement the constant functions 164 introduced in FIG. 6E. Each is a total function and determines the encoding of the output Boolean value for the function. Binary bSNN 1.0, 5.5 implements function ⊕_F¹, and encodes output F by spike time 33; while 1.5, 3.5 implements ⊕_F², and encodes output F by spike time 36. Constant functions ⊕_T¹, and ⊕_T², in turn, are implemented by 2.0, 2.5 and 1.5, 3.5 and encode output T by 29 and 22, respectively.

Building larger neural nets from smaller ones will now be described in more detail. For example, and without limitation, presume a training set of input-output pairs of size N is defined as described above (repeated as follows):(i₁₍₁₎, . . . ,i_n(1))(o₁₍₁₎, . . . ,o_m(1))(i₁₍₂₎, . . . ,i_n(2))(o₁₍₂₎, . . . ,o_m(2))(i_1(N), . . . ,i_n(N))(o_1(N), . . . ,o_m(N))

This training set is partitioned into two training sets, thereby leading to construction of two neural networks that implement these training sets: NN₁ and NN₂. NN₁ may produce the correct result from the inputs in its training set, and similarly for NN₂. However, if these two networks are together in some fashion, it may be possible for an input to progress through both smaller neural networks, resulting in one having a correct output and the other (potentially) not. Consequently, restricted forms of composition are needed that guarantee a correct output.

Continuing as described above, consider the following function:ƒ: I→O, and also some subfunction off that may serve as the abstract function from which the training set may be obtained:

ƒ_T: I_T→O.

FIG. 3 illustrates how this finite function can be represented pictorially 400. In the XOR example (as described above), o and o′ were T and F (e.g., to obtain the training set from which to train the neural network, T was represented by 10 ms and F was represented by 16 ms). Representations of abstract values must satisfy the following conditions: 1) each abstract value must have a unique representation, and 2) different abstract values must have different corresponding representations.

In same-constant composition, the given SNNs implement functions that evaluate to the same constant, which each SNN may be encoding differently. For example, and without limitation, consider the bSNNs that implement XOR functions ⊕_T¹ and ⊕_T². These functions evaluate to the same constant, T Let C₁=2.0, 2.5, and C₂=1.5, 3.5. The conditions to construct the composition SNN C₁∥C₂ are as follows:

• • The designated domains of C₁ and C₂ are disjoint;
  • The designated domain of C₁∥C₂ is the union of the designated domains of C₁ and C₂; and
  • For every element in the designated domain of C₁∥C₂, the output neuron of the composition fires at the same time step.

Component C₁ evaluates input pair {(0, 20)}. The only outcome expected of C₁ concerns this designated domain, {(0, 20)}. Whether outputs for other inputs are defined or what values they have is not critical. Similarly, for component C₂, the designated domain is {(20, 0)}. The composition C₁∥C₂ has the set {(0, 20), (20, 0)} as its designated domain.

These sample components are considered total functions. Each one triggers an output spike for the designated input of the other, its interfering output, as well as for its own designated input, its designated output.

FIG. 7 shows the designated (702, 708) and interfering (704, 706) outputs for both components. A condition to be satisfied in constructing the composition SNN C₁∥C₂ is that its designated outputs be determined by the designated outputs of its components.

In the composition of two SNNs C₁ and C₂ that implement constant functions, each component may get inputs outside its designated domain, and produce interfering outputs which may be recognized and blocked by the neuron components of the modeling system 200 (as shown in FIG. 1). Because a component is implementing a constant, regardless of the size of the designated domain, the designated output is a constant spike time. The goal is to construct from a given component C, which implements a constant function with value c and has spike time s, a component C* which is a partial function, blocks all interfering outputs, and exhibits a spike time s*>s. Component C* will not spike for interfering outputs, and whenever spikes with designated output s, C* will spike at s*>s, the new designated output for the extended component.

Referring now to FIG. 8, constructs for same-constant composition 800 will now be described in detail. For example, and without limitation, same-constant composition may be characterized by the following:

• • Components C₁ and C₂ have disjoint designated domains, and designated spike times s₁ and s₂;
  • From each C_i, for i=1, 2, a partial SNN C_i* (802 and 805, respectively) that blocks interfering outputs is constructed with a new designated spike time of s_i*>sⁱ; and
  • To define a new spike time to represent the constant, two unary bSNNs, w₁^d1 and w₂^d2 may be introduced such that:s=s₁*+sw₁^d1=s₂*+sw₂^d2

Under the assumption that C_i* spikes only for designated outputs, and the disjointness of the designated domains of C₁* and C₂*, only one these components may spike for a given designated input of the composition. This ensures that for all designated inputs the designated output is s.

Different-constants composition takes component SNNs that implement different constant functions, ƒ^c1 and ƒ^c2 with c₁≠c₂, and may construct an SNN that implements a function with a range of size 2: ƒ^c1,c2, Let C_i be an SNN that implements constant function ƒ^ci, which evaluates to c_i, and which encodes c_i by spike s_i, time for i=1, 2. For composition, the spike times that encode the distinct constants are required to be distinct also. If they are not, simply adding a unary bSNN to the output of one will ensure they are (see for example, and without limitation, composition 910 at FIG. 9A).

For example, and without limitation, different-constants composition may be characterized as follows:

• • The designated domains of the components are disjoint;
  • The component SNNs are partial functions that spike only for designated outputs; and
  • The components have different designated outputs.

FIG. 9B (composition 900) shows the structure of C₁C₂. Given an input from its designated domain, only one of its components will fire a spike. If C₁ spikes at s₁ then C₁C₂ spikes at s₁+sw, and if C₂ spikes at s₂ then C₁C₂ spikes at s₂+sw. If the components are single-valued, and their output spike times are different, this composition is single-valued.

Referring now to FIG. 10, an embodiment 1000 of compositional construction of a neural network is described in more detail. For example, and without limitation, a compositional construction of the XOR SNN, will be used to describe the disclosed method. Using the symbol ⊕ to signify the XOR function, this function is defined as follows:⊕: {T,F}×{T,F}→{T,F}

• • T⊕T=F
  • T⊕F=T
  • F⊕T=T
  • F⊕F=F

From the beginning at Block 1001, the present method may partition (Block 1002) the set of inputs of I={T, F}×{T, F} of ⊕ such that I=I₁∪ . . . ∪I_p. These sets in the partition, by definition, are mutually disjoint (e.g., this representation 1004 may be a one-to-one function r: I→I). The representation may obtain a partition of I corresponding to that of I, as follows:I=I₁∪ . . . ∪I_p

Each set I_k, for k>{1, . . . , p} may be used by the present method to obtain (at Block 1006) the restriction of ⊕ to I_k, written ⊕|I_k, and defined as follows:⊕|I_k: I_k→O where O={T,F}⊕|I_k(i)def=⊕(i) for i∈k

At Block 1008, for each k∈{1, . . . , p}, the present method may construct an SNN _k to compute function ⊕|I_k. _k may take inputs in I_k. The set O_k of outputs produced may be determined during the construction of _k.

At Block 1010, the present method may construct an SNN from ₁, . . . , _p to compute the function of interest (e.g., the function ⊕), at which point the present method may end at Block 1099.

Taking certain of these process tasks in more detail in the context of the XOR benchmark, Block 1002 may comprise methods to partition |={T, F}×{T, F}, for example, and without limitation, as follows:{T,F}×{T,F}={(T,T)}∪{(T,F)}∪{(F,T)}∪{(F,F)}

At Block 1004, the present method may choose a representation for |={T, F}×{T, F}. For example, and without limitation, the present method may choose to represent each Boolean value the same regardless of whether it occurs in the first or the second element of an input pair, as follows:

• • Let r₁: {T, F}→
  • r₁(T)=0
  • r₁(F)=20Δ; where Δ is some time-step parameter
  • (note: Δ is omitted below, as it will be implicit in all inputs and outputs)
  • Then r: {T, F}×{T, F}→×
  • r(B₁, B₂)=(r₁(B₁); r₁(B₂)) for B₁, B₂∈{T, F}

The partition of the representation of inputs is as follows:{0,20}×{0,20}={(0,0)}∪{(0,20)}∪{(20,0)}∪{(20,20)}

At Block 1006, the present method may obtain the restrictions of ⊕ to the sets in the partition of |, as follows:⊕|{(T,T)}=⊕1/F={(T,T)F}⊕|{(T,F)}=⊕1/T={(T,F)T}⊕|{(F,T)}=⊕2/T={(F,T)T}⊕|{(F,F)}=⊕2/F={(F,F)F}

At Block 1008, the present method may construct SNNs for ⊕|{(T, T)}; ⊕|{(T, F)}; ⊕|{(F, T)}; ⊕|{(F, F)}.

A binary SNN (bSNN) has two input neurons, one output neuron, and one synaptic connection between each input neuron and the output neuron (see, for example, FIG. 11). As shown, a binary bSNN 1100 is a function comprising inputs that are pairs of spike times (i₁; i₂) (i.e., the times at which neurons n_i1 and n_i2 fire spikes). A binary bSNN is a function that either produces no output or a single output o (e.g., no output when no doesn't fire a spike; o, the time at which n_o fires a spike). A neuron fires a spike when its membrane potential x crosses its threshold θ from below. According to the Spike-Response model of the neuron, the membrane potential of the output neuron no at time t is given by the following formula:x(t)=w₁ε(t−i₁−d₁)+w₂ε(t−i₂−d₂)where w₁ and d₁ are the weight and delay parameters of the synapse connecting n_i1 to the output neuron no. The basic spike response or effect caused by a spike received by a neuron is given by some function ε. Inputs i₁ and i₂ are the spike times of n_i1 and n_i2, respectively. Neuron n_o fires a spike at o when θ=x(o)=w₁ε(o−i₁−d₁)+w₂ ε(o−i₂−d₂).

As described above, in some neural network design approaches, the outputs are chosen first and then the arbitrarily chosen original parameters of the network are readjusted by backpropagation until the network produces an output sufficiently close to the one chosen in advance. In the disclosed method, the output neuron fires a spike (that is, merely that there be an o that satisfies the above equation). Parameters θ, w₁, w₂, d₁, d₂ and even the function ε may be chosen so as to obtain this equation. Different choices may lead to different values of o, which means that these choices may yield different SNNs that compute the same abstract function. Exemplary choices yielding such SNNs will now be discussed in detail.

An SNN to compute ⊕|_{{(T, T)}}: Consider the simple abstract function ⊕|_{{(T, T)}}. After having chosen 0 to represent an input T, and instantiating the parameters as follows,

• • θ=1.0
  • d₁=d₂=1
  • w₁=1.0 and w₂=1.5
  • Δ=0.03
  • δ=0.02, where real numbers x, y are considered equal if |x−y|<=δ and the spike response function with τ=6; applying the formula:

$\epsilon \left(t\right)=\{\begin{array}{cc}\frac{t}{\tau }{e}^{1-\frac{t}{\tau }}& t\ge 0\\ 0& t\le 0\end{array}$ yields an output spike time of o=35.

Since the output of the abstract function ⊕|_{{(T, T)}} is F, the choice of parameters has determined that this SNN represents output F by 35. Without loss of generality, in constructing the other subfunctions below, the same instantiations of parameters, except the weights, are kept (the constant parameters are omitted in the diagrams below). For example, and without limitation, function ⊕|{(T, T)} is computed by the binary bSNN 1200 illustrated in FIG. 12. More specifically, SNN 1.0, 1.5 maps (0, 0)→35 (that is, it represents output F by 35. At this stage implementation of ⊕|{(T, T)}, whose input is (0, 0), is of primary significance. Using the notation for restriction of function used above, this result may be written as follows:

• • 1.0, 1.5|_{(0,0)}: {(0, 0)}→{35} computes ⊕|_{{T, T)}}: {(T, T)}→{F}

An SNN to compute ⊕|_{{(T, F)}}: This abstract function may be defined as ⊕|_{{(T, F)}}: {(T, F)}→{T}. Some pair of weights that allows the output neuron to fire a spike may be found: pair w₁=1.5, w₂=1.5 does so (see, for example, SNN 1300 at FIG. 13). Therefore, SNN 1.5, 1.5 maps (0, 20)→39 (that is, it represents T by 39). Expressed in notation for restriction, this result is as follows:1.5,1.5|_{(0,20)}: {(0,20)}→{39} computes ⊕|_{(T,F)}: {(T,F)}→{T}

An SNN to compute ⊕|_{{(F, T)}}: This abstract function may be defined as ⊕|_{{(F, T)}}: {(F, T)}→{T}. One possible SNN 1400 that may compute this function is shown at FIG. 14. Therefore, SNN 2.0, 2.5 maps (20, 0)→27 (that is, it represents output T by 27), which may be expressed in the following notation for restriction:

• • 2.0, 2.5|_{(20,0)}: {(20, 0)}→{27} computes ⊕|_{{(F, T)}}: {(F, T)}→{T}

An SNN to compute ⊕|_{{(F, F)}}: This abstraction function may be defined as ⊕|_{{(F, F)}}: {(F, F)}→{F}. One possible SNN 1500 that may compute this function is shown at FIG. 15. Therefore, SNN 1.5, 3.5 maps (20, 20)→36 (that is, it represents output F by 36, which may be expressed in the following notation for restriction:

• • 1.5, 3.5|_{(20,20)}: {(20, 20)}→{36} computes ⊕|_{{(F, F)}}: {(F, F)}→{F}

The abstract functions determined above comprise the following component SNNs:

SNN computes function
<1.0, 1.5> | {(0,0)}: {(0, 0)}     {35} ⊕| {(T,T)}: {(T, T)}     {F}
<1.5, 1.5> | {(0,20)}: {(0, 20)}     {39} ⊕| {(T,F)}: {(T, F)}     {T}
<2.0, 2.5> | {(20,0)}: {(20, 0)}     {27} ⊕| {(F, T)}: {(F, T)}     {T}
<1.5, 3.5> | {(20,20)}: {(20, 20)}     {36} ⊕| {(F,F)}: {(F, F)}     {F}

As described above, one building block for neural network composition is a unary basic SNN (bSNN) 1600 (see FIG. 16), which comprises one input neuron, one output neuron, and one synaptic connection between them. The membrane potential of the output neuron no at time t is given by the formula x(t)=wε(t−i−d), where w and d are the weight and delay parameters of the synapse connecting input neuron n_i to the output neuron no. Neuron n_o fires a spike at o when θ=x(o)=wε(o−i−d). By choice of parameters w, d, θ, it is possible to either ensure that input i will trigger a spike at o, or to prevent there being an o that will satisfy this equation, thus, blocking a spike. The spike time of neuron no may be denoted by sw^d, when the input (spike) is (at) 0, for synaptic parameters weight w and delay d, and threshold θ is known. More generally and explicitly, the time between the input spike time i and the output spike time o, may be denoted sw_θ^d such that o=i+sw_θ^d.

Referring now to unary bSNN 1700 at FIG. 17, for input i, n_o may spike at i+sw^d for the following choice of parameters: θ=1.0; d=1; Δ=0.03; and the spike response function S1.0¹=143 and s3.0¹=27. Table 1800 at FIG. 18 illustrates some exemplary spike times for these parameters.

A filter fc|d|th may be an instance of a unary bSNN. For example, and without limitation a filter may allow input c to trigger a spike on the output of the neuron. That is, for any other input the output neuron will not spike. Parameters of a filter may include a synaptic weight w=1.0, a delay d, and an output neuron having a threshold th. One way of defining a filter is using the following spike-response function:

$\epsilon \left(t\right)=\{\begin{array}{cc}\mathrm{th}& \mathrm{if}t=c+d\\ 0& \mathrm{if}t\ne c+d\end{array}$ Thus, at t=c+d:x(c+d)=w*ε(c+d)=1.0*th=th and the output neuron of the filter fires a spike.

Components may be employed with or without filters. Exemplary results of component bSNNs for XOR without filters is illustrated in table 1900 at FIG. 19. By contrast, results of a binary bSNN 2100 (see FIG. 21A) vertically composed with filter 2110 (see FIG. 21B) are illustrated in table 2000 at FIG. 20. Filter operation defined: if i=c there is output o; if i≠c, there is no output. FIG. 22 illustrates binary bSNN 2200 composed with filter (and associated mathematical representation 2205). Operation defined: if u=c there is output o; if u≠c, there is no output.

Returning to an SNN that is constructed to compute XOR (from the method 1000 of FIG. 10), the remaining processes may be summarized as follows:

Compose f₃₉[C₂] and f₂₇[C₃] to compute ⊕|_{{(T, F), (F, T)}}:

• • f₃₉[C₂]: (0, 20)→40
  • computes ⊕|_{{(T, F)}}: {(T, F)}→T
  • f₂₇[C₃]: (20, 0)→28
  • computes ⊕|_{{(F, T)}}: {(F, T)}→TBoth components implement constant functions that evaluate to the same constant, TThis composition is denoted by:
  • f₃₉[C₂]∥f₂₇[C₃]

Compose f₃₅[C₁] and f₃₆[C₄] to compute ⊕|_{{(T, T), (F, F)}}:

• • f₃₅[C₁]: (0, 0)→36
  • computes ⊕|_{{(T, T)}}: {(T, T)}→F
  • f₃₆[C₄]: (20, 20)→37
  • computes ⊕|_{{(F, F)}}: {(F, F)}→FBoth components implement constant functions that evaluate to the same constant, FThis composition is denoted by
  • f₃₅[C1]∥f₃₆[C₄]

Compose the two resulting SNNs: f₃₉[C₂]∥f₂₇[C₃] and f₃₅[C1]∥f₃₆[C₄]

Subsequent (same-constant) construction of f₃₉[C₂]∥f₂₇[C₃], may be summarized as follows:

• • SNN f₃₉[C₂] computes function ⊕|_{{(T, F)}}, which evaluates to T
    • For input (0, 20) it outputs 40
    • Thus, it represents T as an output by 40
    • For inputs (0, 0), (20, 0), (20, 20) it produces no output
  • SNN f₂₇[C₃] computes function ⊕|_{{(F, T)}}, which evaluates to T
    • For input (20, 0) it outputs 28
    • Thus, it represents T as an output by 28
    • For inputs (0, 0), (0, 20), (20, 20) it produces no output.

The structure of the composition of these two components, which implements ⊕|_{{(T, F), (F, T)}} is as follows (see graphic 2300 at FIG. 23):

The inputs this structure may accept are

• • (0, 20) to represent (T, F)
  • (20, 0) to represent (F, T)

A single value o may represent the output T

For inputs (0, 0) and (20, 20) it produces no output, since neither component produces an output.

Consider input (0, 20) (see component 2400 at FIG. 24)

Since only the first component fires a spike, the output spike time o satisfies the equationo=40+s[w^d1₁]

Consider input (20, 0) (see component 2500 at FIG. 25)

Since only the second component fires a spike, the output spike time o satisfies the equationo=28+s[w^d2₂]

Output o may satisfy the two constraints above.40+s[w₁^d1]=28+s[w₂^d2]40−28=12=s[w₂^d2]−s[w₁^d1]

Two possible instantiations: (see the spike times for unary bSNNs described above)12=47−35=(46+1)−35=(s[2.0¹]+1)−s[2.5¹]=s[2.0²]=s[2.5¹]  1.

• • w₁=2.5 d₁=1
  • w₂=2.0 d₂=2
  • o=75

This composition of f₃₉[C₂] and f₂₇[C₃] is denoted by2.5¹|f₃₉[C₂]∥f₂₇[C₃]|2.0²12=35−23=35−(20+3)=s[2.5¹]−(s[4.0¹]÷3)=s[2.5¹]−s[4.0⁴]  2.

• • w₁=4.0 d₁=4
  • w₂=2.5 d₂=1
  • o=63

This other composition of f₃₉[C₂] and f₂₇[C₃] with different parameters is denoted by4.0⁴Sf₃₉[C₂]∥f₂₇[C₃]S2.5¹

For the ongoing example scenario, assume the second composition is chosen.

Subsequent (same-constant) construction of f₃₅[C₁]∥f₃₆[C₄] may be summarized as follows:

• • SNN f₃₅[C₁] computes function ⊕|_{{(T, T)}}, which evaluates to F
    • For input (0, 0) it outputs 36
    • Thus, it represents F as an output by 36
    • For inputs (0, 20), (20, 0), (20, 20) it produces no output.
  • SNN f₃₆[C₄] computes function ⊕|_{{(F, F)}}, which evaluates to F
    • For input (20, 20) it outputs 37
    • Thus, it represents F as an output by 37
    • For inputs (0, 0), (0, 20), (20, 0) it produces no output

The structure of the composition of these two components, which implements ⊕|_{{(T, T), (F, F)}} is as shown in graphic 2600 at FIG. 26:

• • The inputs this structure may accept are—
    • (0, 0) to represent (T, T)—
    • (20, 20) to represent (F, F)
  • A single value o may represent the output F

For inputs (0, 20) and (20, 0) it produces no output, since neither component produces an output

Consider input (0, 0) (see component 2700 at FIG. 27)

• • Since only the first component fires a spike, the output spike time o satisfies the equationo=36+s[w^d1₁]

Consider input (20, 20) (see component 2800 at FIG. 28)

• • Since only the second component fires a spike, the output spike time o satisfies the equationo=37+s[w^d2₂]

Output o may satisfy the two constraints above.36+s[w₁^d1]=37+s[w₂^d2]37−36=1=s[w₁^d1]−s[w₂^d2]

One possible instantiation: (See the spike times for unary bSNNs as described above):1=28−27=28−(24+3)=s[3.0¹]−(s[3.5¹]+3)=s[3.0¹]=s[3.5¹]  1.

• • w₁=3.0 d₁=1
  • w₂=3.5 d₂=2
  • o=64

Construction of an SNN to compute ⊕, the XOR function, may be summarized as follows (more specifically, (composition of 4.0⁴|f₃₉[C₂]∥f₂₇[C₃]|2.5¹ and 3.0¹|f₃₅[C₁]∥f₃₆[C₄]|3.5⁴):4.0⁴|f₃₉[C₂]∥f₂₇[C₃]|2.5¹

• • 4.0⁴|f₃₉[C₂]∥f₂₇[C₃]|2.5¹
  • maps (0, 20)63
  • (20, 0)63
  • no output for (0, 0), (20, 20)
  • computes ⊕|_{{(T, F), (F, T)}}: {(T, F), (F, T)}T

Let C_T=4.0⁴|f₃₉[C₂]∥f₂₇[C₃]|2.5¹

SNN CT represents output T by 633.0¹|f₃₅[C₁]∥f₃₆[C₄]|3.5⁴

• • 3.0¹|f₃₅[C₁]∥f₃₆[C₄]|3.5⁴
  • maps (0, 0)64
  • (20, 20)64
  • no output for (0, 20), (20, 0)
  • computes ⊕|_{{(T, T), (F, F)}}: {(T, T), (F, F)}F

Let C_F=3.0¹|f₃₅[C₁]∥f₃₆[C₄]|3.5⁴

SNN C_F represents output F by 64

Each component may produce extraneous outputs when given inputs for which it was not designed

• • Thus, before composition with another component, each component is composed with a filter resulting inf₆₃[C_T]=f₆₃[4.04|f39[C2]∥f27[C3]|2.51]f₆₄[C_F]=f₆₄[3.01|f35[C1]∥f36[C4]|3.54]Table 2900 at FIG. 29 illustrates exemplary SNNs already constructed, such that f₆₃[C_T] represents output T by 64, and f₆₄[C_F] represents output T by 65.

Different-constant composition of f₆₃[C_T] and f₆₄[C_F] may be summarized as follows (see also component 3000 at FIG. 30):

• • These two components implement ⊕|_{{(T, F), (F, T)}} and ⊕|_{{(T, T), (F, F)}}
  • Their composition implements ⊕|_{{(T, F), (F, T), (T, T), (F, F)}}=⊕
  • The structure of their composition is as follows:
    • Consider inputs (0, 20) and (20, 0) (see components 3100, 3200 at FIG. 31 and FIG. 32, respectively)
    • Consider inputs (0, 0) and (20, 20) (see components 3300, 3400 at FIG. 33 and FIG. 34, respectively)o_F=65+s[w^d]
    • Since T≠F, the values representing them are necessarily different.
      • Component f₆₃[C_T] represents T as an output by 64, and component f₆₄[C_F] represents F as an output by 65
      • Since 64≠65, then o_T=64+s[w^d]≠65+s[w^d]=o_F, and this composition has distinct representations for the two Boolean values
    • Any choice of synaptic parameters w and d that produces a spike, that is for which s[w^d] is defined, is acceptable.
      • For example, from the table above, arbitrarily chosen is s[3.0¹]=28o_T=64+28=92o_F=65+28=93
      • The resulting composition may be denoted by

$f_{63}\left[{𝒞}_T\right]\stackrel{{3.}^1}{{\rm Y}}{f}_{64}\left[{𝒞}_F\right]$

It will be appreciated that the systems and methods described above are set forth by way of example and not of limitation. Absent an explicit indication to the contrary, the disclosed steps may be modified, supplemented, omitted, and/or re-ordered without departing from the scope of this disclosure. Numerous variations, additions, omissions, and other modifications will be apparent to one of ordinary skill in the art. In addition, the order or presentation of method steps in the description and drawings above is not intended to require this order of performing the recited steps unless a particular order is expressly required or otherwise clear from the context.

Embodiments disclosed herein may include computer program products comprising computer-executable code or computer-usable code that, when executing on one or more computing devices, performs any and/or all of the steps thereof. The code may be stored in a non-transitory fashion in a computer memory, which may be a memory from which the program executes (such as random-access memory associated with a processor), or a storage device such as a disk drive, flash memory or any other optical, electromagnetic, magnetic, infrared or other device or combination of devices. In another implementation, any of the systems and methods described above may be embodied in any suitable transmission or propagation medium carrying computer-executable code and/or any inputs or outputs from same.

DRAWING NUMBER KEY

• • 100 Neural Network
  • 102 Input to Neural Network
  • 105 Output of Neural Network
  • 110 Spiking Neural Network (SNN)/single synapse
  • 112 Presynaptic neuron
  • 115 Postsynaptic neuron
  • 120 Spiking Neural Network (SNN)/multiple synapse
  • 130 Sample spike times table
  • 132 Unary basic SNN (bSNN)
  • 134 Spike time of bSNN
  • 142 Firing shift scenario: Basic
  • 144 Firing shift scenario: Input delay
  • 146 Firing shift scenario: Synapse delay increase
  • 148 Firing shift scenario: Vertical composition of bSNNs
  • 152 Table of bSNN (3.5, 4.5) for n₁=0
  • 154 Table of bSNN (3.5, 4.5) for n₂=0
  • 162 XOR definition
  • 164 Constant functions that compute XOR
  • 172 Constant function implemented as bSNN (1.0, 1.5)
  • 174 Constant function implemented as bSNN (2.0, 2.5)
  • 176 Constant function implemented as bSNN (1.5, 3.5)
  • 178 Constant function implemented as bSNN (1.5, 3.5)
  • 200 Neural Network Design System (NNDS)
  • 202 Computing device
  • 210 Wide network
  • 212 Processor
  • 213 Data Store
  • 219 Network Interface
  • 220 Modeling Subsystem
  • 230 User Interface
  • 300 Single-valued relation representation: ƒ: A→B
  • 400 Finite function representation: f_T: I_T→O
  • 500 Same-constant composition representation
  • 600 Different-constant composition representation
  • 702 Designated output of C₁
  • 704 Interfering output of C₂
  • 706 Interfering output of C₁
  • 708 Designated output of C₂
  • 800 Same-constant composition construct
  • 802 partial SNN C₁*
  • 805 partial SNN C₂*
  • 900 Different-constant composition construct
  • 910 Different-constant composition with added unary bSNN
  • 1000 Neural Network Compositional Construction method
  • 1001 Method start
  • 1002 Partition training input
  • 1004 Determine representation for partition
  • 1006 Obtain restrictions for partitions
  • 1008 Construct subfunctions (SNN _k)
  • 1010 Combine subfunctions into modeled function of interest
  • 1099 Method end
  • 1100 Sample Binary bSNN
  • 1200 Binary bSNN computing function ⊕|{(T, T)}
  • 1300 Binary bSNN computing function ⊕|{(T, F)}
  • 1400 Binary bSNN computing function ⊕|{(F, T)}
  • 1500 Binary bSNN computing function ⊕|{(F, F)}
  • 1600 Sample Unary bSNN
  • 1700 Unary bSNN spiking at i+sw^d
  • 1800 Sample spike times table for θ=1.0; d=1; Δ=0.03
  • 1900 Sample spike times table for XOR without filters
  • 2000 Sample spike times table for XOR with filter
  • 2100 Binary bSNN
  • 2110 Filter
  • 2200 Binary bSNN with filter
  • 2205 Mathematical representation for binary bSNN with filter
  • 2300 Composition implementing ⊕|_{{(T, F), (F, T)}}
  • 2400 Composition implementing ⊕|_{{(T, F), (F, T)}} for input (0,20)
  • 2500 Composition implementing ⊕|_{{(T, F), (F, T)}} for input (20,0)
  • 2600 Composition implementing ⊕|_{{(T, T), (F, F)}}
  • 2700 Composition implementing ⊕|_{{(T, T), (F, F)}} for input (0,0)
  • 2800 Composition implementing ⊕|_{{(T, T), (F, F)}} for input (20,20)
  • 2900 Sample output of constructed SNNs
  • 3000 Composition implementing ⊕|_{{(T, F), (F, T), (T, T), (F, F)}}=⊕
  • 3100 Composition implementing ⊕ for input (0,20)
  • 3200 Composition implementing ⊕ for input (20,0)
  • 3300 Composition implementing ⊕ for input (0,0)
  • 3400 Composition implementing ⊕ for input (20,20)

Claims

1. A method of neural network (NN) construction comprising: receiving a construction training set of an output variable type and characterized by a construction domain, wherein said training set comprisesan n×N array of inputs i;an m×N array of target outputs to be learned o; andan m×N array of actual outputs u containing error E

2. The method according to claim 1, where each of the plurality of component neural networks is of a single-valued type.

3. The method according to claim 1, where each of the plurality of partition domains further comprises a respective designated input domain such that the plurality of designated input domains is disjoint.

4. The method according to claim 1, where each of the plurality of component neural networks further comprises: a first input artificial neuron CJ;a second input artificial neuron C2;andan output artificial neuron.

5. The method according to claim 4, where at least one of the first input artificial neuron C1, the second input artificial neuron C2, and the output artificial neuron comprise at least one of a membrane potential, a threshold, and a spike time.

6. The method according to claim 4, further comprising: a first artificial synapse connecting the first input artificial neuron C1 and the output artificial neuron; anda second artificial synapse connecting the second input artificial neuron C2 and the output artificial neuron.

7. The method according to claim 6, where at least one of the first synapse and the second synapse comprise at least one of a weight and a delay.

8. The method according to claim 1, further comprising: determining a respective representation for each of the plurality of partition training sets; anddetermining a respective restriction for each of the plurality of representations; andwhere creating the plurality of component neural networks further comprises constructing for each of the plurality of restrictions a first input artificial neuron C1, a second input artificial neuron C2, and an output artificial neuron.

9. A neural network (NN) model construction system utilizing a computer processor and a non-transitory computer-readable storage medium comprising a plurality of instructions which, when executed by the computer processor, is configured to: receive a construction training set of an output variable type and characterized by a construction domain, wherein said training set comprisesan n×N array of inputs i;an m×N array of target outputs to be learned o; andan m×N array of actual outputs u containing error E

10. The system according to claim 9, where each of the plurality of component neural networks is of a single-valued type.

11. The system according to claim 9, where each of the plurality of partition domains further comprises a respective designated input domain such that the plurality of designated input domains is disjoint.

12. The system according to claim 9, where each of the plurality of component neural networks further comprises: a first input artificial neuron CJ;a second input artificial neuron C2;andan output artificial neuron.

13. The system according to claim 12, where at least one of the first input artificial neuron C1, the second input artificial neuron C2, and the output artificial neuron comprise at least one of a membrane potential, a threshold, and a spike time.

14. The system according to claim 12, further comprising: a first artificial synapse connecting the first input artificial neuron C1 and the output artificial neuron; anda second artificial synapse connecting the second input artificial neuron C2 and the output artificial neuron.

15. The system according to claim 14, where at least one of the first synapse and the second synapse comprise at least one of a weight and a delay.

16. The system according to claim 9, where the plurality of instructions which, when executed by the computer processor, is further configured to: determine a respective representation for each of the plurality of partition training sets;determine a respective restriction for each of the plurality of representations; andconstruct for each of the plurality of restrictions a first input artificial neuron C1, a second input artificial neuron C2, and an output artificial neuron.

17. A computer-implemented method of neural network (NN) model construction comprising: accessing a construction training set of an output variable type and characterized by a construction domain, wherein said training set comprisesan n×N array of inputs i;an m×N array of target outputs to be learned o; andan m×N array of actual outputs u containing error E

18. The computer-implemented method according to claim 17, where creating each of the plurality of component neural networks further comprises: creating a first input artificial neuron CJ;creating a second input artificial neuron C2; andcreating an output artificial neuron.

19. The computer-implemented method according to claim 18, where creating each of the plurality of component neural networks further comprises creating at least one of the first input artificial neuron C1, the second input artificial neuron C2, and the output artificial neuron to comprise at least one of a membrane potential, a threshold, and a spike time.

20. The computer-implemented method according to claim 18, where creating each of the plurality of component neural networks further comprises: creating a first artificial synapse connecting the first input artificial neuron C1 and the output artificial neuron; andcreating a second artificial synapse connecting the second input artificial neuron C2 and the output artificial neuron.