Snn adaptation with transient synchrony

Abstract

To provide a method for configuring a spiking neural network.SOLUTION: The spiking neural network comprises a plurality of spiking neurons 1 and a plurality of synaptic elements 2 interconnecting the spiking neurons to form a hardware-implemented network. A response local cluster in the network comprises a set of spiking neurons and a plurality of synaptic elements interconnecting the set of spiking neurons, wherein a network state in the response local cluster is cyclostationary when an input signal 11 to the response local cluster comprises a predetermined oscillation frequency when represented in the frequency domain, by setting weights of the synaptic elements and spiking behavior of the spiking neurons in the response local cluster such that the network state in the response local cluster is periodic with the predetermined oscillation frequency.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates generally to automatic signal recognition techniques, and in particular to temporal control and adaptation that enable the implementation of transient synchrony mechanisms and homeostatic regulation within networks of spiking neurons. [Background technology]

[0002] In biological neural network models, individual neurons communicate asynchronously through sparse events or spikes. Within such event-based spiking neural networks (SNNs), only neurons that change state can generate spikes and trigger signal processing in subsequent layers, thereby saving computational resources. Spiking neural networks (SNNs) are promising tools for many different applications, including automatic signal recognition (ASR), which recognizes signals by identifying their component features.

[0003] SNNs encode information in the form of one or more precisely timed (voltage) spikes, rather than as integer or real-valued vectors. The computation for inference (i.e., inferring the presence of a certain feature in an input signal) can be efficiently performed in the analog and time domains. As a result, SNNs are typically implemented in hardware as fully custom mixed-signal integrated circuits, which allows them to perform inference functions with orders of magnitude lower energy consumption than their counterparts, deep neural networks (DNNs), in addition to having smaller network sizes.

[0004] An SNN consists of a network of spiking neurons interconnected by synapses, which determine the strength of the connections between the connected neurons. This strength is represented as a weight, which moderates the effect of the output of a presynaptic neuron on the input to a post-synaptic neuron. Typically, these weights are set during a training process, which involves exposing the network to large amounts of labeled input data and gradually adjusting the synaptic weights until the desired network output is achieved.

[0005] Relying on the principle that amplitude-, time-, and frequency-domain features in an input signal can be encoded into unique spatially and temporally encoded spike sequences, SNNs can be directly applied to pattern recognition and sensor data fusion.

[0006] Neuromorphic SNN emulators, e.g., systems including electronic analog / mixed-signal circuits that mimic the neurobiological architecture present in the nervous system, form distributed (in the non-von Neumann sense, i.e., where computational elements and memory are co-localized, resulting in memory storage and complex nonlinear behavior, which are executed simultaneously by neurons in the network) parallel event-driven systems that offer capabilities such as adaptation (including adaptation of physical properties, firing rates, homeostatic (behavior) regulation, etc.), self-organization, and learning.

[0007] A distinctive feature of the output of a robust neuromorphic system is the combination of coordinated excitability of designated neurons with circuit-level interactions that maintain the correct (i.e., correct within the boundaries or definitions imposed by the learning rules or frequency of operation) temporal coordination / phasing of these neurons. An example of circuit-level interactions is a circuit that mimics / models brain biochemical (Na-, K-, Ca-) mechanisms to influence neuronal firing and frequency adaptation.

[0008] The complexity and computational power of an SNN are evaluated by its dynamics, in particular the spatial localization and temporal variation of its activity. The dynamics (and architectural formulation) of an SNN can be adapted to traditional (ANN) connectionist models, such as multi-layer feedforward networks or recurrent networks, with time coding.

[0009] However, the inherent characteristics of SNNs are distinctly different from traditional networks, and SNNs do not need to follow the rigid schemes associated with these traditional networks.

[0010] However, there is currently no practical mapping method that presents users of an SNN with an adaptively trained SNN or a portion of an adaptively trained SNN that can be used in classification or further processing of specific predetermined frequency ranges that the user may be interested in. Conventional SNN mappings do not provide an efficient mechanism for control / synchronization of network transient dynamics, and as a result, limit the network's ability to acquire, be sensitive to, or filter out selected frequencies of the input signal, effectively preventing the network from fully utilizing the frequency content within the spatiotemporal spike train.

[0011] For example, when processing radar, lidar, image, or audio signals, or any other application where signals can be represented in the frequency domain, the signals consist of different frequency ranges that encode certain information. Using an SNN to process specific frequency ranges in the frequency domain and enabling a high level of network modularity and granularity with synchronous behavior of network subgroups greatly improves the practical implementability of the SNN in terms of cost-effectiveness, adaptability, scalability, and reliability. Summary of the Invention

[0012] To address the above-mentioned shortcomings of the prior art, according to a first aspect of the present disclosure, there is provided a method for configuring a spiking neural network, the spiking neural network comprising: a plurality of spiking neurons; and a plurality of synaptic elements interconnecting the spiking neurons to form a network at least partially implemented in hardware, each of the synaptic elements adapted to receive a synaptic input signal and to apply a weight to the synaptic input signal to generate a synaptic output signal, the synaptic elements being configurable to adjust the weight applied by each of the synaptic elements; and each of the spiking neurons receiving one or more of the synaptic output signals from one or more of the synaptic elements. a network adapted to receive one or more received synaptic output signals and adapted to generate a spatiotemporal spike train output signal in response to the received one or more synaptic output signals, wherein a response local cluster in the network comprises a set of spiking neurons and a plurality of synaptic elements interconnecting the set of spiking neurons, the method comprising setting weights of the synaptic elements and spiking behavior of the spiking neurons in the response local cluster so that when an input signal to the response local cluster includes a predetermined oscillation frequency when expressed in the frequency domain, the network state in the response local cluster is a periodic steady state, whereby the network state in the response local cluster is periodic at the predetermined oscillation frequency.

[0013] In this way, a spiking neural network can be configured so that when an input signal to a local cluster contains a predetermined oscillation frequency, the local cluster contained within the spiking neural network can be in a periodic steady state. This means that portions of the input signal that are close to and contain the predetermined oscillation frequency have a greater effect on the local cluster than other portions of the input signal that are close to and contain the predetermined oscillation frequency. This allows a spiking neural network configured in this way to process a specific (predetermined) frequency range in the frequency domain.

[0014] In one embodiment, setting the weights of the synaptic elements and the spiking behavior of the spiking neurons in the response local cluster includes iteratively training the response local cluster by optimizing the weights of the synaptic elements and the spiking behavior of the spiking neurons so that a required periodic steady-state behavior is achieved.

[0015] In this way, the response of a local cluster to a particular input signal can be efficiently configured and the desired behavior can be reliably achieved.

[0016] In a further embodiment, when an input signal to a response local cluster includes a predetermined oscillation frequency, the probability distribution activity or statistical parameter of a set of neurons in the response local cluster is cyclostationary at the predetermined oscillation frequency.

[0017] In a further embodiment, the periodic steady state is represented by the equation

[0018]

number

[0019] is the solution of where T is a predetermined time period, Φ(t,τ) is the state transition matrix of the synaptic drives Γ(t) of all neurons in the response local cluster, and F(t) is the function

[0020]

number

[0021] is a deterministic function of the stochastic part of the synaptic drive Γ(t) as given by K(t) is the autocorrelation function of the synaptic drive Γ(t), where K0 is the initial condition.

[0022] In this way, an effective, scalable and reliable method is obtained for configuring spiking neural networks, in particular local clusters, to obtain periodic steady states.

[0023] In a further embodiment, the spiking neural network comprises a drive local cluster, the drive local cluster comprising a set of spiking neurons and a plurality of synaptic elements interconnecting the set of spiking neurons, whereby an output signal of the drive local cluster serves as an input signal to the response local cluster, whereby the drive local cluster and the response local cluster are coupled with a particular coupling strength, and the method further comprises setting the network state in the response local cluster to have a steady state and / or a time-varying state when the input signal from the drive local cluster to the response local cluster does not include a predetermined oscillation frequency when expressed in the frequency domain or when the particular coupling strength is less than the predetermined coupling strength.

[0024] In a further embodiment, setting the weights of the synaptic elements and the spiking behavior of the spiking neurons in the response local cluster comprises iteratively training the response local cluster by optimizing the weights of the synaptic elements and the spiking behavior of the spiking neurons so as to achieve required steady-state and / or time-varying behavior.

[0025] In this way, the response of a local cluster to a particular input signal can be efficiently configured and the desired behavior can be reliably achieved.

[0026] In a further embodiment, the probability distribution activity or statistical parameters of the set of neurons in the response local cluster are stationary or non-stationary when the response local cluster receives an input signal from the drive local cluster that does not include a predetermined vibration frequency when expressed in the frequency domain, or when the specified connection strength is less than a predetermined connection strength.

[0027] In a further embodiment, the steady state is represented by the equation

[0028]

number

[0029] is the solution of In the formula, K(t) ∝ is the steady-state value of the autocorrelation function K(t) of the synaptic drive Γ(t), and F(t) is expressed by the formula

[0030]

number

[0031] is a deterministic function of the stochastic part of the synaptic drive Γ(t) as given by where Φ(t,τ) is the state transition matrix of the synaptic drive Γ(t) of all neurons in the response local cluster, dω is the infinitesimal stochastic change, and E(t) is

[0032]

number

[0033] is a deterministic function defined by The time-varying state is given by the matrix equation

[0034]

number

[0035] is the solution of The matrix equation is the differential Lyapunov matrix equation

[0036]

number

[0037] is a continuous-time algebraic Lyapunov matrix equation forr and Q r are discretized versions of E and F, and t r means the time point of the numerical integration.

[0038] In this way, an effective, scalable and reliable method is obtained for configuring spiking neural networks, in particular local clusters, to obtain steady-state and / or time-varying states.

[0039] In a further embodiment, the increase in the structural dimensionality of the response local cluster is achieved by ensuring generalized external synchronization between the drive local cluster and the response local cluster, where the generalized external synchronization is the coupling of the drive local cluster to the response local cluster with a specific coupling strength that is equal to or greater than a predetermined coupling strength.

[0040] In this way, a high level of network modularity and granularity is obtained, e.g., with synchronization behavior of network subgroups, which improves the practical implementability of SNNs in terms of adaptability and scalability.

[0041] In a further embodiment, the generalized exosynchronization is the average autocorrelation function of the synaptic drive Γ(t)

[0042]

number

[0043] where τ is the delay, N is the number of neurons in the local cluster, and the average is over the neuronal population.

[0044] In further embodiments, the steady-state numerical solution, the time-variable value solution, and / or the periodic steady-state solution are obtained by using feedback coupling between neurons in a response local cluster that results in synchronization of the neuronal activity of the neurons.

[0045] According to a second aspect of the present disclosure, there is provided a spiking neural network for processing an input signal representable in the frequency domain, comprising a plurality of spiking neurons and a plurality of synaptic elements interconnecting the spiking neurons to form a network at least partially implemented in hardware, each of the synaptic elements adapted to receive a synaptic input signal and to apply a weight to the synaptic input signal to generate a synaptic output signal, the synaptic elements being configurable to adjust the weight applied by each of the synaptic elements, and are adapted to receive one or more synaptic output signals from one or more of the synaptic elements and to generate a spatiotemporal spike train output signal in response to the received one or more synaptic input signals, and a responsive local cluster in the network comprises a set of spiking neurons and a plurality of synaptic elements interconnecting the set of neurons, and when an input signal to the responsive local cluster includes a predetermined first frequency of oscillation when expressed in the frequency domain, a probability distribution activity or statistical parameter of the set of neurons in the local cluster is cyclostationary at the predetermined first frequency of oscillation.

[0046] In this way, a spiking neural network can be configured so that when an input signal to a local cluster contains a predetermined oscillation frequency, the local cluster contained within the spiking neural network can be in a periodic steady state. This means that portions of the input signal that are close to or contain the predetermined oscillation frequency have a greater effect on the local cluster than other portions of the input signal. This allows a spiking neural network configured in this way to process a specific (predetermined) frequency range in the frequency domain.

[0047] In one embodiment, the probability distribution activity or statistical parameters of a set of neurons in a local cluster that is cyclostationary can be calculated using the equation

[0048]

number

[0049] is described by the periodic steady-state solution of where T is a predetermined time period, Φ(t,τ) is the state transition matrix of the synaptic drives Γ(t) of all neurons in the response local cluster, and F(t) is the function

[0050]

number

[0051] is a deterministic function of the stochastic part of the synaptic drive Γ(t) as given by K(t) is the autocorrelation function of the synaptic drive Γ(t), where K0 is the initial condition.

[0052] In a further embodiment, the spiking neural network comprises a drive local cluster, the drive local cluster comprising a set of spiking neurons and a plurality of synaptic elements interconnecting the set of spiking neurons, whereby an output signal of the drive local cluster serves as an input signal to the response local cluster, whereby the drive local cluster and the response local cluster are coupled with a particular coupling strength, and whereby the probability distribution activity or statistical parameter of the set of neurons in the response local cluster is stationary or non-stationary when the response local cluster receives an input signal from the drive local cluster that does not include a predetermined oscillation frequency when expressed in the frequency domain, or when the particular coupling strength is less than the predetermined coupling strength.

[0053] In a further embodiment, the probability distribution activity or statistical parameters of a set of neurons in a local cluster that is stationary is given by the equation

[0054]

number

[0055] is described by the steady-state numerical solution of In the formula, K(t) ∝ is the steady-state value of the autocorrelation function K(t) of the synaptic drive Γ(t), and F(t) is expressed by the formula

[0056]

number

[0057] is a deterministic function of the stochastic part of the synaptic drive Γ(t) as given by where Φ(t,τ) is the state transition matrix of the synaptic drive Γ(t) of all neurons in the response local cluster, dω is the infinitesimal stochastic change, and E(t) is

[0058]

number

[0059] The probability distribution of activity or statistical parameters of a set of neurons in a local cluster is a deterministic function defined by the matrix equation

[0060]

number

[0061] is described by the time-variable solution of The matrix equation is the differential Lyapunov matrix equation

[0062]

number

[0063] is a continuous-time algebraic Lyapunov matrix equation for r and Q r are discretized versions of E and F, and t r means the time point of the numerical integration.

[0064] In a further embodiment, the driving local cluster is an input encoder of a spiking neural network, which converts sampled input signals into spatiotemporal spike trains that are then subsequently processed by the responding local cluster.

[0065] According to a third aspect of the present disclosure, there is provided a method for processing specific frequency portions of an input signal representable in the frequency domain using a spiking neural network, the method comprising using a spiking neural network according to the second aspect or an embodiment of the second aspect of the present disclosure, or a spiking neural network obtained according to the first aspect or an embodiment of the first aspect of the present disclosure; A method is disclosed that includes providing an input signal in the form of a spatiotemporal spike train to a response local cluster of a spiking neural network, the input signal including one or more frequency portions, and processing the input signal using the response local cluster such that certain frequency portions of the input signal that include a predetermined oscillation frequency have a greater effect on neurons of the response local cluster than other frequency portions of the input signal.

[0066] In this way, when processing any application where the input signal to a spiking neural network can be represented in the frequency domain, the signal consists of different frequency ranges encoding certain information, and specific frequency ranges can be processed in specific ways, thus enabling a high level of network modularity and granularity with synchronization behavior of network subgroups, thus improving the practical implementability of SNNs in terms of cost-effectiveness, adaptability, scalability, and reliability.

[0067] According to a fourth aspect of the present disclosure, there is disclosed a physical signal to inference processor for adaptively processing physical signals, the processor comprising a selector and extractor for selecting and extracting specific signal features from the physical signals, and a spiking neural network for performing classification and processing of the physical signals based on the specific signal features extracted from the physical signals, wherein the processor further comprises an operational block for establishing a current operating context and an optimal feature set, the processor comprising a feedback loop to the selector and extractor, and the selector and extractor are adaptive in the sense that different signal features can be selected and extracted based on a specific processing task.

[0068] In this way, an optimal feature set can be selected for each operational context by the physical signal inference processor, improving the adaptability, cost-effectiveness, and optimization of the use of spiking neural networks.

[0069] According to a fifth aspect of the present disclosure, a method for adaptively processing a physical signal is disclosed, the method including receiving the physical signal in a physical signal inference processor according to the fourth aspect of the present disclosure, selecting and extracting specific signal features using a selector and extractor, processing the specific signal features using a spiking neural network, determining a current operational context and an optimal feature set using an operational block, and sending a feedback signal to the selector and extractor to adaptively change the selected and extracted signal features as needed.

[0070] In this way, optimal feature sets can be selected via this method, thus improving the adaptability, cost-effectiveness, and optimization of the use of spiking neural networks to adaptively process physical signals. [Brief explanation of the drawings]

[0071] Embodiments will now be described, by way of example only, with reference to the accompanying schematic drawings in which corresponding reference symbols indicate corresponding parts, and in which: [Figure 1] FIG. 1 illustrates an exemplary neural network consisting of neurons and synaptic elements. [Figure 2] FIG. 1 is a schematic diagram of a spiking neural network within a microcontroller integrated circuit. [Figure 3] FIG. 1 is a schematic diagram of an exemplary run-time physical signal to inference signal processor with an adaptive spiking neural network. [Figure 4] The nominal values ​​of the spike adaptation mechanism are shown in the two leftmost graphs, the reset and refractory period mechanisms for modeling sodium and potassium conductance activation and inactivation dynamics are shown in the two middle graphs, and the inserted variability in spike generation is shown in the rightmost graph. [Figure 5] FIG. 1 illustrates a network with different synchronous cell assemblies. [Figure 6] Analog neuronal voltage traces are shown in the top graph, synaptic activation in the middle graph, and synchronization and homeostatic regulation in the bottom graph. [Figure 7] The figure shows the non-modulated state at the top and the modulated state at the bottom, with the activity of individual neurons over time shown in the two graphs on the left and the network activity of the (non-)modulated network shown in the two graphs on the right.

[0072] The figures are intended for illustrative purposes only and do not serve as a limitation on the scope or protection defined by the claims. DETAILED DESCRIPTION OF THE INVENTION

[0073] Certain embodiments are described in more detail below, however, it should be understood that these embodiments are merely examples and should not be construed as limiting the scope of protection of the present disclosure.

[0074] FIG. 1 is a simplified diagram of a neural network 100. Neurons 1 are connected to each other through synaptic elements 2. To avoid obscuring the drawing, only a few neurons and synaptic elements are shown (only some are numbered). The connection topology shown in FIG. 1 , i.e., the manner in which synaptic elements 2 connect neurons 1 to each other, is only an example; many other topologies may be employed. Each synaptic element 2 can send a signal to an input of a neuron 1, and each neuron 1 that receives the signal can process the signal and then generate an output, which is transmitted to another neuron 1 via further synaptic elements 2. Each synaptic element 2 is assigned a weight, which is applied to each of the synaptic input signals it receives and transmits to generate a weighted synaptic output signal. Thus, the weight of a synaptic element is a measure of the type of causal relationship between two neurons 1 connected by a synaptic element 2. The relationship can be causal (positive weight), anticausal (negative weight), or nonexistent (zero weight).

[0075] Neurons 1 and synaptic elements 2 may be implemented in hardware, e.g., using analog circuits or circuit elements, digital hardwired logic circuits or circuit elements, or a combination thereof. For example, each neuron and each synaptic element may be implemented as a hardware circuit or circuit element. This type of hardware implementation may also include functions performed using software; thus, neurons and synaptic elements may be implemented partially in hardware and partially in software, i.e., with hardware circuits executing software to perform the functions of individual neurons and synaptic elements. This contrasts with designs using large processors running software, where the software mimics individual neurons and synaptic elements. These (partial) hardware implementations can achieve much faster processing of input signals, e.g., they enable much faster pattern recognition and much faster event-driven processing, where blocks of neurons and synaptic elements are activated only as needed.

[0076] Neural network 100 may be a spiking neural network. Neuron 1, then, is a spiking neuron, which generates a neuron output signal in the form of one or more spikes or neuron-generated events. Spiking neuron 1 may be configured to fire (i.e., generate an output spike) only when the membrane potential (e.g., energy potential, or voltage or current level) within the neuron reaches a predetermined threshold. The membrane potential of a spiking neuron changes as a result of received input signals; i.e., synaptic output signals received by the neuron from synaptic elements are accumulated, integrated, or otherwise processed to modify the membrane potential. When the weight of synaptic element 2 is positive, the synaptic output signal received from the synaptic element excites spiking neuron 1 receiving the signal, increasing the membrane potential of spiking neuron 1. When the weight of synaptic element 2 is negative, the synaptic output signal received from the synaptic element inhibits spiking neuron 1 receiving the signal, decreasing the membrane potential of spiking neuron 1. When the weight of a synaptic element 2 is 0, the synaptic output signal received from that synaptic element has no effect on the membrane potential of the spiking neuron 1 that receives the signal.

[0077] When the membrane potential of a spiking neuron 1 reaches a threshold, the neuron fires, generating a spike upon firing, and the membrane potential is reduced as a result of the firing. Subsequently, when the membrane potential reaches the threshold again, the neuron fires again, generating a second spike. Thus, each spiking neuron 1 is configured to generate one or more spikes forming a spatiotemporal spike train in response to input signals received from connected synaptic elements 2. Because spiking neurons 1 fire only when their membrane potential reaches a predetermined threshold, coding and processing of temporal information is incorporated into neural network 100. In this manner, a spatiotemporal spike train is generated in spiking neural network 100, which is a temporal sequence of spikes generated by spiking neurons 1 of network 100.

[0078] The temporal characteristics of a spike train encode the amplitude and frequency features of the input signal, including the latency between the onset of a stimulus (e.g., an input signal from a synaptic element) and the generation of a spike at the neuron's output, the latency between successive spikes from the same neuron, and the number of spikes generated by the neuron firing within the duration that the input stimulus is applied.

[0079] Synaptic elements 2 may be configurable, for example, so that each of the weights of the synaptic elements can be varied, for example, by training neural network 100. Neuron 1 may be configurable so that neuron 1 responds to signals from the synaptic elements. For example, in the case of a spiking neural network, neuron 1 may be configured so that a signal increases or decreases the membrane potential, the time it takes for the membrane potential to naturally decay toward the resting potential, the value of the resting potential, and / or the threshold that triggers a spike in spiking neuron 1. The configuration of neuron 1 may be held constant or variable during training, for example, and may be set by training neural network 100 using a particular training set.

[0080] The input signal 11 may be, for example, a number of different sampled input signals or a spatiotemporal spike train. The input may be analog-to-digital converted values ​​of the signal samples, or for example digital values ​​of the samples in the case of an analog or digital integrator, or analog values ​​of the samples in the case of an analog integrator.

[0081] The output signal 12 of the neural network 100 is, for example, a spatiotemporal spike train, which can be read out from the output neuron 1, further classified, and converted by an output conversion stage into a set of digital values ​​corresponding to the type of output code selected by the user.

[0082] Biological SNNs can be characterized by sparse and irregular inter-neuronal connections with low average activity, where only active neurons contribute to information processing. Examples of sparse inter-neuronal connections include the formation of local clusters of neurons, short or long path loops, and synchronous cell assemblies.

[0083] A local cluster, also referred to as a cell assembly, represents a group of neurons and their interconnecting synaptic elements with strong shared excitatory inputs. Neurons within such a local cluster tend to be operationally activated as a whole group when a sufficient subset of the neurons in the local cluster is stimulated. In this way, a local cluster can be thought of as a processing unit. The associations between neurons in a local cluster, i.e., the possible mechanisms by which a cluster is defined, can take different forms.

[0084] The synchronized response of a group of neurons is one example of defining a local cluster of neurons. One example of interest is the time-varying (cyclo) periodic activation of neurons within a local cluster in response to a stimulus. Other possible mechanisms by which local clusters are defined can be built-in redundancy (i.e., additional circuits / neurons to ensure fail-safe operation), averaging (i.e., the end result is the average of multiple neurons, not a single neuron), layer definition (layers can be defined based on, among other things, the network concept and layer size, and the network depth), etc. SNNs behave as complex systems defined by complex dynamic interactions between neurons, for example, by communication between neurons within a local cluster or between different local clusters.

[0085] When synchronous activation of neurons is used as a mechanism for forming local clusters, short-term memory (STM) may consequently be represented as persistent activity of neurons within a local cluster, sustained by reverberation (i.e., periodic activation of neurons), and long-term memory (LTM) corresponds to the formation of new local clusters, e.g., by plasticity mechanisms or other learning methods. Within this framework, firing time synchronization for a group of neurons is examined as a series of synchronized firings by a specific subset of neurons, as collective synchronization of a specific subset of neurons that provides spatiotemporal integration, and as an extended notion of polychronization within a subset of neurons.

[0086] A synchronized neuronal response may be defined herein as the correlated occurrence in time of two or more events associated with various aspects of neuronal activity. This synchronization may take the form of synchronized firing rates and / or firing times of subsets of neurons at or near a particular frequency. Multiple subsets of neurons may each synchronize at different frequencies, reflecting their role for learning / memory formation or for some other predefined signal processing function, such as filtering. Local clusters may be formed by polychronization, the self-organization of groups of neurons within an SNN into local clusters that exhibit reproducible time-locked patterns of spiking activity as a result of synaptic strengthening.

[0087] In one embodiment, neurons in a local cluster have a synchronous neuronal response, such that the neurons fire (i.e., generate output spikes) in sync. This synchronous response can take different forms: Neurons in a local cluster may fire periodically at the same time and rate in repeating cycles at an oscillatory frequency.

[0088] There may be some variation in output spike timing and rate over time; for example, the oscillation frequency may be substantially constant or may decrease or increase over time. There may be some variation in output spike timing and rate among neurons in a local cluster, such that the firing time of a neuron falls within a time window of the peak of neurons firing at the oscillation frequency, i.e., the peak of the oscillation cycle when the greatest number of neurons fire. This distribution of spike firing times for neurons in a local cluster may follow a Gaussian curve or an equivalent statistical distribution (e.g., log-normal, Poisson, exponential, etc.), with the greatest number of output spikes occurring at the oscillation frequency and spike occurrences becoming progressively less frequent the further from the peak. A specified time window may be understood as a time range within an oscillation period during which an input signal has the greatest effect on a local cluster oscillating at a given frequency.

[0089] The local cluster as a whole may exhibit synchronous periodic neuronal firing as described herein. This may occur because all neurons in the local cluster fire periodically, or because fewer than all neurons fire periodically, but the average spiking activity of the local cluster follows the synchronous periodic behavior described herein, for example. Here, the average spiking activity is the average activity of the local cluster within a particular time range. The average spiking activity may increase or decrease over multiple predetermined oscillation periods. This is an example of frequency- or rate-based coding. It is assumed that most, if not all, information about the stimulus is contained in the neuronal firing rates. Because the sequence of action potentials generated by a given stimulus varies from trial to trial, neuronal responses are typically processed statistically or probabilistically.

[0090] When a local cluster is not oscillating synchronously but is, for example, waiting for an input signal, neurons in the local cluster will be more responsive to the input signal if these input signals arrive sequentially at a given frequency to which the local cluster has high sensitivity.

[0091] Other examples of suitable coding are temporal or latency or time-to-first-spike coding, and alternatively, rank-order coding. Temporal or latency or time-to-first-spike coding relates to the relative timing of spikes, interaural delays, or timing differences between spikes. Rank-order coding examines only the order in which spikes arrive, rather than precise timing information.

[0092] In general, one can look at the activity of a statistical parameter or probability distribution associated with, for example, neurons in a local cluster. Thus, one example of a statistical parameter is the average spiking activity. In the case of a probability distribution, one can look at, for example, the first moment (expectation), the second central moment (variance), the third standardized moment (skewness), and the fourth standardized moment (kurtosis) when, for example, neurons in a local cluster spike.

[0093] Synaptic delay, i.e., the time required for signal propagation across synaptic elements, is the quantity required for transitions in oscillatory / transient dynamics and, consequently, for coherent and incoherent states of an SNN. Noise, either in the form of stochastic inputs or nonuniformity in intrinsic firing rates, reduces the coherence of oscillations between neurons. Here, oscillations are rhythmic or repetitive patterns of neural activity within an SNN. Neurons can generate oscillatory activity in many ways, driven by mechanisms within individual neurons or by interactions between neurons. Within individual neurons, oscillations can manifest as either oscillations in membrane potential or rhythmic patterns of action potentials, which then generate oscillatory activation of postsynaptic neurons. At the level of local clusters, synchronized activity of multiple neurons can give rise to oscillations within the local cluster. Oscillatory activity within a group of neurons can arise from feedback connections between neurons, which result in the synchronization of neuronal firing patterns. Oscillations and associated oscillatory activity are among the concepts used to analyze spike train properties such as rate, spike coincidence, and the presence of other temporal patterns.

[0094] The role of synchronization depends on oscillation frequency. For example, synchronization of lower-frequency oscillations (used at the local level) can establish transient networks implementing specific processes (e.g., learning, adaptation, STM / LTM, etc.), while synchronization of higher-frequency oscillations (used at the global level) can communicate information about process outcomes, such as cross-frequency modulation (CFM)—either power or phase locking. In such CFM, a first cell assembly modulates a second cell assembly through some modulation method, such as cross-coupling two oscillators through a feedback loop or modulating high-frequency amplitude with low-frequency phase. The overall coupling strength defines whether the network operates in a stable or chaotic activity regime. Stronger heterogeneity always leads to a state that is very different from synchrony and can be completely asynchronous without obvious coordination of spike times. Generally, higher coupling strength leads to more synchrony. Information transmission and signal integration through synchronization provide spiking neural networks with a high level of cost-effectiveness, adaptability, scalability, and reliability.

[0095] Neurons communicate within a network primarily through fast all-or-none events, i.e., spikes in the neuronal membrane potential. The relative spike firing times within a neuronal population can be understood as the information code carried by the network, and synchronization between neuronal populations can be understood as the signal for encoding and decoding information. Neuronal properties and associated variables describe neuronal states (e.g., describing the time course of the membrane potential). The encoded information can be represented, for example, by these neuronal states. When a neuron or a population of neurons oscillates and synchronizes, each of these variables can also oscillate at the same frequency and return to the same value once per oscillation period, with one such oscillation (e.g., membrane potential) corresponding to a spike with a subsequent interspike interval.

[0096] We then describe the implementation of a mapping method that optimizes the temporal dynamics of stochastic cell assemblies and enables synchronization of cell assemblies for improved information transmission with spiking neural networks.

[0097] A function f (which may be a discontinuous function) j (.) (This is a nonlinear transformation, and for mathematical convenience we use the function f j (.) is modeled as φ(x)=tanh(x), but other choices of φ are possible) is used to specify the input-output transfer function, which is the excitation ( / inhibitory) voltage v j E (.)(respectively, v j I (.)) through the membrane of the jth neuron to the firing rate (in Hz) of the jth neuron. The superscripts E and I denote excitatory and inhibitory, respectively, and relate to excitatory or inhibitory neurons, respectively. As a result, the receiving or postsynaptic voltage v in the ith neuron i (t) is

[0098]

number

[0099]

number

[0100] is given by During the ceremony,

[0101]

number

[0102] and

[0103]

number

[0104] is the firing time t j and firing time t j are indices that indicate the firing of excitatory and inhibitory transmitting (i.e., presynaptic) neurons in the E and n I and α j E (.) and α j I (.) are functions that describe the evolution of excitatory and inhibitory postsynaptic potentials, respectively. Furthermore, v th,i E (.) and v th,i E (.) is the continuous threshold input voltage, which specifies a constant input to the membrane potential, i.e., an extrinsic contribution to the membrane potential, which can be an external stimulus voltage. The threshold voltage is the voltage that the membrane potential needs to reach so that the corresponding neuron can spike; when the threshold is met, the neuron fires in the case of an excitatory neuron, or the neuron causes inhibition of action potential generation in the post-synaptic (inhibitory) neuron cell. The neuron connectivity matrix w ij The entry of represents the connection strength of the jth neuron to the ith neuron.

[0105] The interaction between excitatory and inhibitory populations can result in a fully synchronized state or a partially synchronized structure, depending on the cross terms in equations (1) and (2) above, which can be improved by increasing inhibitory dynamics. In a heterogeneous inhibitory network with sparse random connectivity, inhibitory dynamics provides a dual role in both suppressing population activity and generating neural reactivation. When the network becomes saturated, it is unable to properly process input signals. Neural reactivation allows the network to become responsive to input signals again. The concept of inhibition involves, among other things, the interruption or blocking of activity and the restriction of activity patterns in both space and time within the network. However, inhibitory interneurons do not only provide a stop signal for excitation; the operational dynamics within a neuronal network can only be maintained if excitatory forces are counteracted by effective inhibitory forces.

[0106] Synchronization of spike transmission within a circuit occurs when the input from an excitatory neuron to an adjacent inhibitory neuron is sufficiently strong. Connection strength is controlled by inhibitory innervation, and when inhibitory input is present, i.e., at intermediate connections, neurons spike with low rhythmicity / synchrony. The inhibitory spike-time dependent plasticity (STDP) function, shown below in equation (6), is symmetric and favors a decrease in synaptic conductance; in contrast, the excitatory STDP function is antisymmetric and biased toward potentiation.

[0107] The function τ(t) in the above equation is a bounded differential function of time t. For a bounded differential function of time t, the following condition is satisfied:

[0108]

number

[0109] is fulfilled, where r and h are positive constants. i(E,I) is the gain,

[0110]

number

[0111] modulates the exponential decay of synaptic voltage, mimicking the spike-time-dependent scaling of input conductance, and α i (E,I) (t)=B (E,I) e -t / λi(E,I) Assuming that, the synaptic drive of the i-th (excitatory or inhibitory) postsynaptic neuron is

[0112]

number

[0113] It can be defined by:

[0114] The synaptic drive essentially indicates the synaptic strength. The synaptic drive Γ of neuron i i can be understood as the total transfer function of all synaptic elements driving neuron i, which can be a function of, for example, the exponential decay values, weights, and gains of the synaptic elements driving neuron i. From (1), (2), and (4), this can be expressed as

[0115]

number

[0116] This means that...

[0117] Considering the case of spike-timing-dependent plasticity (STDP) learning and long-term plasticity (LTP), the following boundaries and changes are considered:

[0118]

number

[0119] Obtained, In the formula, η + and η - is a soft bound, i.e., for large weights, synaptic depression dominates over potentiation.

[0120] Neurons are noisy in both spike generation and synaptic signal transmission. Noise arises from the quantal discharge of neural transmitters, the random opening of ion channels, the coupling of background neural activity, etc. Noise then induces neuronal variability, increasing neuronal sensitivity to environmental stimuli, affecting interneuron synchronization, and facilitating probabilistic inference.

[0121] element w of the connectivity matrix ij is the correlation [w ij w ji ] J = ρ, where the square brackets [ ] J denotes the average over realizations of the random connection. The parameter ρ quantifies the symmetry of the connection, i.e., for ρ = 0, the element w ij and w ji are independent, the connectivity matrix is ​​completely asymmetric, for ρ=1 the connectivity matrix is ​​completely symmetric, and for ρ=-1 the connectivity matrix is ​​completely antisymmetric. The degree of symmetry can be employed to observe the evolution of a particular network configuration, i.e., how likely it is that the configuration is the result of chance.

[0122] Then, the definition of generalized exosynchronization for stochastic neural networks with time-varying delays is

[0123]

number

[0124]

number

[0125] can be expanded with where x and y represent unique cell assemblies or local clusters, i.e., the x network forms the driving network and the y network forms the response network. Thus, the driving network x drives the response in the response network y. The cell assembly y experiences the perturbation. The cell assemblies can be located directly adjacent to each other, i.e., at least one output neuron of the driving network x can be directly connected to at least one input neuron of the response network y. The cell assemblies x and y can be located, for example, in separate layers or in the same layer of the network, or the cell assemblies x and y can be separated by some hardware that couples the two spiking neural networks to effectively form a single spiking neural network. The time-dependent function ω(t) describes the noise in the input voltage and is represented by Brownian motion, i.e., an n-dimensional standard Wiener process. We consider the noise intensity function σ i (x,y,t) satisfies the Lipschitz condition, and the positive constants p and q satisfy the trace (σ i T σ i )≦px T x+qy T y. With selective stimulation of neurons, we can target several specific neural network features, such as excitation / inhibition balance, network dynamics, and performed computations. Here, excitation / inhibition balance refers to the fact that neurons can operate optimally in a (sparse) coincidence-detection regime, i.e., most of the synapses are excitatory and have low firing rates, and that neurons can operate optimally in a balanced regime with nearly balanced excitation and inhibition and high output firing rates.

[0126] When a perturbation occurs, the neuron computation element creates a spike based on the perturbation sign, magnitude, and phase, i.e., the cluster (cell assembly) of connected neurons can maintain the phase difference resulting from the cluster correlation input. Note that each neuron adjusts its spiking phase when the frequency response is changed. The adaptive controller u for node i i (t) is

[0127]

number

[0128] is expressed as In the formula, J(B i (.)) is a vector function B i (Γ x,i ) and ε i =y i (t)-B i (Γ x,i (t)) is the synchronization error, and k is a sufficiently large positive constant. Therefore, an adaptive controller is used to control the synaptic drive of neurons in the response network y so that the synchronization error between the drive network x and the response network y is small. In this way, synchronization occurs between the drive network x and the response network y. As mentioned above, the overall connection strength u defines whether the network operates in a stable or chaotic activity regime. Stronger heterogeneity, for example, cell assembly x and cell assembly y not oscillating at the same frequency, always leads to a state that is very different from synchrony and can be completely asynchronous without obvious coordination of spike times. Generally, higher connection strength leads to more synchrony.

[0129] In this way, the increase in the structural dimensionality of cell assembly y (the increase in the amount of synchronized neurons within cell assembly y) is described by ensuring generalized external synchronization of probabilistic neural networks based on the average autocorrelation function. This description provides a mechanism that is robust to heterogeneity, sparsity of connectivity, and noise.

[0130] Next, (8) is

[0131]

number

[0132] can be formulated as a system of linear stochastic differential equations in the standard form as

[0133] The solution of (10) is

[0134]

number

[0135] is given by In the formula, Φ(t,τ) is a state transition matrix, and the state transition matrix is ​​dΦ(t,τ) / dt=E(t)Φ(t,τ),Φ(τ,τ)=I m describes the time evolution of the dynamical system between time τ and time t, as a function of t as a solution of equation (11).

[0136]

number

[0137] If the deterministic matrix functions E(t) and F(t) are bounded within the time interval of interest, then there exists a unique solution for every initial value vector Γ(t0).

[0138] In addition to the timing of neuronal spikes, e.g., the temporal coordination of neuronal output, neurons that spike in unison with well-defined time lags between neuronal spikes convey information about sensory input. As a result, with coincident spiking, the internal state of the network can be correlated; specifically, synchronized groups of neuronal subpopulations that represent perceptual content that form a coherent entity, such as a visual object in a scene, can be correlated.

[0139] Network activity is determined by the average autocorrelation function Γ(t) (the autocorrelation function is the correlation between a signal and a delayed copy of itself as a function of delay) Γ(t) = 1 / N × {ΣE[Γ(t+τ / 2)Γ(t-τ / 2) T ]}, where τ is the delay and the average is over the neuronal population. Let K(t) be the autocorrelation matrix of Γ, i.e., E[Γ(t+τ / 2)Γ(t-τ / 2) T ] and the probability differential

[0140]

number

[0141] Using the theorem of and then taking the expectation of both sides, we obtain the differential Lyapunov matrix equation

[0142]

number

[0143] Obtained, The analytical solution for the differential Lyapunov matrix equation is

[0144]

number

[0145] It has the form:

[0146] At steady state, equation (12) becomes

[0147]

number

[0148] is expressed as where the subscript α denotes the steady-state value, and E and F denote the deterministic matrix functions E(t) and F(t). Steady state refers to the state of a network with stable behavior, i.e., the recently observed behavior of the system continues into the future. In a stochastic system at steady state, the probability of various states repeating remains constant. For example, steady state occurs when the network is in an idle state using only direct current. The steady-state covariance function is given below by, for t≦t′:

[0149]

number

[0150] and If t>t', then

[0151]

number

[0152] As stated above, it can be obtained.

[0153] To obtain a numerical solution, equation (12) needs to be discretized in time using a suitable scheme, such as any linear multi-step formula (e.g., trapezoidal method, backward differentiation formula, or Runge-Kutta method). Using the backward Euler (first-order backward differentiation formula), the differential Lyapunov matrix equation (12) can be reduced to the continuous-time algebraic Lyapunov matrix equation

[0154]

number

[0155] It can be written in a special form called

[0156] where t r =t (r-1) +h r where h r is the numerical integration time step. r For each of these, numerical methods can compute an approximation to the exact solution. r and matrix P r are the discretized (in time) versions (numerical integration methods) of the deterministic matrix function E and the deterministic matrix function F, respectively.

[0157] P r When is sparse and large-scale, and Q r If , is of low rank, we solve equation (17) with a Krylov-type method.

[0158]

number

[0159] and

[0160]

number

[0161] This allows us to calculate a low-rank solution by solving for K(t r )=UOU T where U is an orthogonal matrix.

[0162]

number

[0163] Starting from This can be solved by using the orthogonality of U to obtain the reduced Lyapunov equation

[0164]

number

[0165] Leading to The approximation is calculated by applying the Galerkin condition, i.e., given U, we apply the Galerkin orthogonal condition U T R(UOU T )U=0, where R(UOU T ):=P r (UOU T )+(UOU T )P r T +Q r Q r T UOU T is the residual associated with

[0166] This is a small dense Lyapunov equation that can be solved directly, for example with the Bartels-Stewart method. r For i = 1, 2, ..., we solve equation (17) with a low-rank version of an iterative method related to rational matrix functions. The assumed iteration for the Lyapunov equation (17) for i = 1, 2, ... is

[0167]

number

[0168] and is given by In the formula, γ i is the iteration shift parameter. For efficient implementation, we divide the iterations by the Cholesky factor of the iteration, i.e., K i =L i L i H can be replaced by the coefficient L i can be reformulated in terms of

[0169] This therefore provides a mathematical description of time-varying solutions to equation (12). These are, for example, networks that do not exhibit stable behavior, for which the activity of neurons fluctuates over time.

[0170] For the periodic steady-state solution (if there is one), E(t+kT)=E(t) and F(t+kT)=F(t) for all t,

[0171]

number

[0172] and for some time period T>0. The initial condition K0 for equation (12) is

[0173]

number

[0174] Equation (13) is satisfied by setting K(T)=K0, where

[0175]

number

[0176] and This can be calculated by numerically integrating equation (12) with the initial condition K(0) = 0. Equation (21) is a system of algebraic linear equations for the entries of matrix K, and equations of this form are commonly referred to as discrete-time algebraic Lyapunov matrix equations due to the equations' origin from Lyapunov stability theory for discrete-time systems. Equation (21) can be solved numerically with the Bartels-Stewart algorithm for continuous-time algebraic Lyapunov matrix equations.

[0177] Thus, equations (14)-(16) above describe the steady-state numerical solution, equations (17)-(20) describe the time-variable solution, and finally, equations (21)-(22) describe the periodic steady-state numerical solution. These solutions describe the spiking neural network in terms of its network activity, i.e., they describe the linearized solution, which in turn describes the synaptic drive for each of the neurons in a cell assembly or other local cluster of the spiking neural network. A local cluster can follow a steady-state solution, a time-variable solution, and / or a periodic steady-state numerical solution, depending on the current state of the network.

[0178] Thus, to obtain synchronized oscillations within a spiking neural network, or within a subset of neurons within a spiking neural network, e.g., within a cell assembly or other local cluster within a spiking neural network, the network drive of different neurons with synchronized oscillatory behavior must obey Equations (21)-(22). The synaptic drive of a particular neuron within a spiking neural network can be tuned, for example, by varying the weights of each of the synaptic elements connected to the neuron, by configuring the manner in which a signal increases or decreases the neuron's membrane potential, by configuring the time it takes for the neuron's membrane potential to naturally decay toward its resting potential, by changing the value of the neuron's resting potential, and / or by changing the threshold that triggers the neuron to spike.

[0179] The parameters that define the value of synaptic drive at a particular instance in time can be optimized through an iterative process, e.g., using software. This can be done through hardware simulation, or directly through hardware implementation and output study. Thus, the network can be trained so that a particular cell assembly follows the equations that describe synchronous oscillatory behavior.

[0180] In other words, when drive network x is oscillating at a predetermined frequency, and response network y is mapped to allow response network y to obey equations (21)-(22) that govern the periodic steady-state solution, and the connection strength between drive network x and response network y is strong enough, response network y will oscillate in a manner governed by equations (21)-(22). Then, the oscillation period is a predetermined oscillation period T. This means that the probability distribution activity or statistical parameters of the set of neurons in response network y are cyclostationary at the frequency at which drive network x is oscillating.

[0181] If the connection strength is not strong enough, the probability distribution activity or statistical parameters of the set of neurons in the response network y may be stationary or non-stationary. The value of the connection strength that makes the connection between the driving network x and the response network y sufficiently strong may be a predetermined value that results in a particular response to an input signal. This may be determined on an application-by-application basis.

[0182] In general, driving network x can be thought of as closer to the input signals coming from sensors or other devices that serve as inputs to the spiking neural network than response network y, because the spike trains are processed by driving network x to drive response network y before they are processed by response network y.

[0183] The input encoder converts a sampled input signal into a spatiotemporal spike train. Each output neuron of the input encoder generates an output signal that is a spatiotemporal spike train. The input encoder is configured in a specific configuration, e.g., by setting the weights of the synaptic elements of the input encoder in a specific configuration. Multiple input encoders may be used, split across multiple frequency ranges of the input signal frequency domain, for multiple input encoder implementations.

[0184] The driving network x can be an input encoder. The driving network x can be connected to the input encoder, such that the output signal of the input encoder, a spatiotemporal spike train, is processed by the driving network to generate spatiotemporal spike trains in the driving network x. These spatiotemporal spike trains encode specific frequencies in the sampled input signal to the spiking neural network, so that the spatiotemporal spike trains occur at specific frequencies in specific neurons of the driving network x. This means that the driving network x can also oscillate at this specific frequency, but this is not required.

[0185] Because drive network x and response network y are coupled, the spatiotemporal spike trains processed by drive network x are then processed by response network y. As a result, response network y can be stationary, non-stationary, or cyclostationary, i.e., the probability distribution activity or statistical parameters of the set of neurons in response network y are stationary, non-stationary, or cyclostationary.

[0186] The two, stationary and non-stationary, can be the result of either weak coupling strength between the driving network x and the response network y, or the driving network x not providing an input signal in the form of a spatiotemporal spike train at a given frequency.

[0187] Thus, the cyclostationary outcome may be the result of a sufficiently strong coupling strength between the driving network x and the response network y, and the driving network providing an input signal in the form of a spatiotemporal spike train at a given frequency.

[0188] The frequency at which the synchronously vibrating cell assemblies vibrate is a positive input constant and can be selected based on which frequencies the cell assemblies need to be sensitive to. In principle, any frequency at which the local clusters are capable of vibrating can be the predetermined frequency.

[0189] All neurons within a local cluster or cell assembly need not spike at exactly the same moment, i.e., they need not be perfectly synchronized. Spiking probability may be rhythmically modulated so that neurons are more likely to fire at the same time, which causes oscillations in the average activity of neurons. Thus, the frequency of large-scale oscillations within a local cluster or cell assembly need not match the firing patterns of each of the individual neurons within the local cluster or cell assembly.

[0190] Neurons within a synchronously oscillating local cluster or cell assembly may also spike within a time range of each other, for example, all neurons spike within a time range that is 10 percent of the oscillation period, or within a time range that is within 1 percent of the oscillation period, or within a time range that is within 0.1 percent of the oscillation period, or within a time range that is within 0.01 percent of the oscillation period.

[0191] Generally, the spiking neural network is first modified to meet the timing requirements of the spiking neural network. Local clusters can operate within a frequency range centered on a predetermined frequency, and the distance from the center of the range can be a 3-sigma variation. Other distances from the center of the range can be a 1-sigma or 2-sigma variation. Next, synchronization within a certain time range can be defined by a probability density function as described above. Then, selecting a specific sigma can result in a certain time range.

[0192] The number of neurons in the local cluster that spike within this time range may be 50 percent or more, more preferably 75 percent or more, more preferably 90 percent or more, more preferably 95 percent or more, more preferably 97.5 percent or more, more preferably 99 percent or more of the total number of neurons in the local cluster.

[0193] 2 illustrates one embodiment of a high-level architecture of a microcontroller integrated circuit 100 that includes a learning spiking neural network 110. In this context, the microcontroller 110 is an economical means of data collection, sensing, pattern recognition, and actuation of physical signals. Within a spiking neural network, individual neurons communicate asynchronously through sparse events or spikes. A spiking neural network system is implemented as an array of spiking integrators that include neurosynaptic cores. One possible implementation of a learning spiking neural network is described in WO 2020 / 099680 A2.

[0194] The spiking neural network 110 is connected to one or more streaming input data ports 111, which provide the spiking neural network 110 with inputs that are converted into spatiotemporal spike trains. The spiking neural network 110 is connected to one or more output ports 112. The memory-mapped control and configuration interface 113 controls the configuration parameters of the spiking neural network 110, such as synaptic weights and / or neuron configurations, and may further include peripherals (e.g., A / D converters, D / A converters, bandgaps, PLLs) and circuits for control and adaptation of neurons, synapses, and plasticity (learning) circuits, among others. The interface 113 reads the memory device 102, where the configuration for the spiking neural network 110 is stored, and sends signals to the spiking neural network 110 to configure the hardware accordingly. The interface 113 may send analog signals to the spiking neural network 110. The settings may include configuration parameters of each of the neurons 1 or synaptic elements 2 of the spiking neural network 110, or the network topology.

[0195] Each neuron 1 can have a set of configuration parameters that control the precise firing behavior of the neuron 1. For example, a neuron can be designed with a firing threshold, which represents a threshold of voltage, energy, or other variable that accumulates within the neuron as a result of receiving an input; when the accumulated variable meets or exceeds the firing threshold, the neuron generates an output spike (such as a voltage, current, or energy spike). The neuron can implement an integral function, which integrates the input to the neuron to determine an adjustment to the accumulated variable. In addition, the neuron can also be designed with (a) a leakage rate, which represents the rate at which the accumulated variable within the neuron decays over time; (b) a resting value of the accumulated variable, which represents the value to which the accumulated variable decays over time in the absence of an input signal to the neuron; (c) an integration time constant, which represents the time over which the input signal is integrated over time to determine an increase in the accumulated variable within the neuron; (d) a refractory level, which represents the value of the accumulated variable within the neuron immediately after the neuron fires; and (e) a refractory period, which represents the period required for the accumulated variable within the neuron to rise to its resting value after the neuron fires. These parameters may be predetermined and / or configurable and / or adjustable for each neuron. For example, by adjusting a neuron's firing threshold, leak rate, integration time constant, and refractory period to match the energy capacity of an important input signal feature, neuron 1 will generate one or more precisely timed spikes when stimulated with an input signal containing that feature.

[0196] The configuration parameters of synaptic element 2 include the weight of synaptic element 2 and the gain of synaptic element 2. The weight of synaptic element 2 is typically used to tune synaptic element 2, and the gain of synaptic element 2 is used for signal amplification in hardware, typically related to a low-pass filter implementation. Typically, the gain is fixed at network 110 initialization, but the weight can change based on the evolution / training of spiking neural network 110.

[0197] Microcontroller integrated circuit 100 further comprises a microprocessor core 101 for performing computations and control of integrated circuit 100. For example, microprocessor core 101 may oversee communications between memory-mapped control and configuration interface 113 and memory device 102.

[0198] The memory device 102 can be any computer-readable storage medium. The memory device 102 can be a non-transitory storage medium. Exemplary computer-readable storage media include, but are not limited to, (i) non-writable storage media in which information can be permanently stored (e.g., a read-only memory device within a computer, such as a CD-ROM disk readable by a CD-ROM drive, a ROM chip, or any type of solid-state non-volatile semiconductor memory), and (ii) writable storage media in which changeable information can be stored, such as a hard disk drive or any type of solid-state random access semiconductor memory, flash memory.

[0199] An external bus 104 is connected to one or more sensors or other data sources 103. The microcontroller integrated circuit 100 may also be directly attached to sensors 105. The sensors may first pass through an analog-to-digital converter 106. One or more serial input / output ports 107 and a general-purpose input / output port 108 may be present on the microcontroller integrated circuit 100. Direct access of external devices may be placed into the memory of the microcontroller integrated circuit 100 by direct memory access (DMA) 109.

[0200] The SNN core may employ a mixed analog-digital computing platform, i.e., spike trains incorporate analog information in the timing of events, which are then converted back to analog representation at the input of the synaptic matrix.

[0201] The neurosynaptic cores disclosed in this invention can be organized as a repeating array of synaptic circuits and neuronal units, where each unit can form a cell assembly. The system incorporates the presence of electronic synapses at the junctions of the array. The periphery of the array contains rows of synaptic circuits that mimic the function of the soma and axon hillocks of biological neurons. Furthermore, each neurosynaptic core in the array can have a local router that communicates with routers in other cores within the dedicated real-time reconfigurable network-on-chip.

[0202] A classifier in the present invention can have a set of output neurons (one for each classification class), each of which fires to generate events (spikes) according to the output neuron's respective firing probability distribution. The neurosynaptic computational elements can generate complex spatiotemporal dynamics that can be scaled toward specific features that can support target signal processing functions.

[0203] Neuronal spiking characteristics are controlled by a specific set of parameters. Neurons function as (frequency) band-pass devices, where phase-locked states occur at frequencies associated with the time constants of both the neuron frequency adaptation dynamics and the spike frequency adaptation dynamics. Here, a band-pass device can be understood as a band-pass filter, i.e., processing only signals within a specified (designed) frequency band (range). Spike frequency adaptation mechanisms produce specific results for sensory signal processing, such as highlighting changes in stimulus parameters, preventing spiking frequency saturation, or adjusting the frequency response to specific stimulus features. The topology of neurosynaptic elements controls the regularity of spontaneous neuronal spiking, such as the firing rate scaling factor and the strength of intrinsic noise, resulting in coherence / frequency evolution.

[0204] The present invention encompasses the ability of neurosynaptic cores to apply transient synchronization, homeostatic regulation, non-uniform learning rules, weight memory types, and communication protocols to the synaptic matrix.

[0205] In one embodiment, a neurosynaptic core can be organized as a single core implementing a uniform learning rule. In another embodiment, different regions of the synaptic matrix can be configured with non-uniform learning rules, depending on transient synchronization. The size of this region of the synaptic matrix can be configurable at design time to create specialized circuits, or dynamically configurable at runtime, where the array has a non-uniform clustering of synaptic circuits, each of the clusters implementing, for example, a different learning rule.

[0206] This non-uniform clustering may require a synchronization mapping and / or synthesis algorithm to map each of the neuronal units to a determined cluster of synaptic matrices based on transient synchrony. Without loss of generality, adaptive spiking neural networks enable the implementation of run-time physical signals into inference signal processors, e.g., for radar, lidar, and biophysical signal inference.

[0207] 3 schematically illustrates an exemplary runtime physical signal inference signal processor with an adaptive spiking neural network 200. As shown, the system 200 can include a set of parallel adaptive channel selectors and extractors 203 in which each of the channels or each of a subset of the channels can be activated at any given moment in time based on a particular signal processing task.

[0208] A physical signal 201 enters the processor 200, and based on the specific signal processing task, a selector and extractor 203 selects and extracts specific signal features of the physical signal 201 to serve as input to the processor 200. The adaptability of the feature selection and extraction module 202, which comprises the selector and extractor, means that different signal features can be selected and extracted based on the specific processing task. This is important because the system thus becomes more versatile. Different applications often require completely different signal features; for example, analysis of an image is done differently than analysis of an electrocardiogram, and even images can be analyzed differently based on the information one wants to obtain from the image. Because the processor 200 is adaptable, the processor 200 can be used for many applications.

[0209] The different signals coming from the feature selection and extraction module 202 can then pass through a multiplexer and pre-processing stage 204, where the feature set is pre-processed (e.g., filtered, amplified, etc.) and forwarded on a single output line. In some cases, signal pre-processing may not be necessary.

[0210] The feature set from the adaptive channel selector and extractor is provided to the SNN information processor 205, which performs classification and processing of the preprocessed signals and performs application-specific signal information estimation, which eliminates the need for a complete reconstruction of the sensor waveform, i.e., the sensor-specific signal. This provides a significant improvement in energy efficiency. Features can be activated and configured at runtime along with several parameters, such as gain, bandwidth, noise filtering, and variability bounds. The information or inference metrics 206 coming as output from the SNN can then be used by a specific application.

[0211] The inference accuracy depends on environmental conditions such as signal dynamics or signal interference. At runtime, the operational settings and conditions block 207 establishes the current operational context and, consequently, the optimal feature set. A feedback loop dynamically adapts the channel specifications within the power-performance design space. The runtime configuration execution unit then activates and configures only the relevant settings.

[0212] Based on training and test data across processing tasks 209 and parameterized performance settings 210, a subset of learning and signal processing tasks may be selected 208, for example, noise filtering, variability bounds, channel synchronization.

[0213] Neuromorphic systems display a wide range of pattern activities, e.g., complete synchrony, clusters, or asynchronous states, depending on the excitatory / inhibitory network interaction conditions, heterogeneity in input patterns, and spatiotemporal dynamics of neurosynaptic elements, see Figures 4 and 5.

[0214] Figure 4 shows nominal values ​​of the spike adaptation mechanism in the two leftmost graphs, variability in the signal generation and refractory period mechanisms modeling sodium (Na) and potassium (K) conductance activation and inactivation dynamics, respectively, shown in the two middle graphs, and neuronal spikes with similar levels of interpolated variability shown in the rightmost graph.

[0215] FIG. 5 illustrates a network with different synchronous cell assemblies 501-512. Several local clusters 501-512 are illustrated at different locations within the network, and some of the local clusters 501-512 exhibit synchronous behavior, with their neurons oscillating synchronously, e.g., spiking at the same frequency. The behavior of different local clusters 501-512 may differ based on the requirements for each of the local clusters 501-512 within the network. SNNs may be adaptive, such that the local clusters 501-512 may change, for example, their oscillation frequency, or even the neurons contained within a single local cluster may change.

[0216] Neurons generate action potentials with a wide range of spike dynamics. The synchrony of the reconfigurable network over a wide time range is shown in Figures 6 and 7.

[0217] FIG. 6 shows an increase in oscillatory synchronicity of spiking neurons within a cell assembly over a particular time range. As can be seen in the top graph, the spikes become increasingly localized in time. In the bottom graph, it can be seen that neurons increasingly spike within the same time window within an oscillatory period. As can be seen in the middle graph, the cell assembly begins with an amount of synchronicity approaching 0 percent, but at the end of the particular time range, the synchronicity approaches approximately 80 percent. Depending on the application, the amount of synchronicity can be increased even further.

[0218] Figure 7 shows the unmodulated state at the top and the modulated state at the bottom, with the two graphs on the left showing the activity of individual neurons over time and the two graphs on the right showing the network activity of the (un)modulated network.

[0219] Regulated networks exhibit synchronous oscillatory behavior. Thus, network activity is much more localized, with virtually all neurons spiking within specific time windows within the oscillatory period. This contrasts with the unregulated network state in the graph above, where network activity is largely random.

[0220] The duration selectivity curve is asymmetric to combat non-uniformity in spike latency and the resulting latency curve. The strength of phase locking to the target frequency is controlled by changing the detuning parameters of the curve, e.g., the initial frequency. The initial transient dynamics set by synaptic connections guides the network into a steady-state dynamic regime. With increasing coupling, a transition occurs from an ensemble of individually firing neurons to a coherent, synchronized network.

[0221] In addition to information transmission, the extent to which transient synchrony captures the (instantaneous) phase relationships between (controlled) target frequencies was examined. Neurons generate time-locked patterns, and due to the interplay between conductance delays and plasticity rules, networks can form sets of neurons with reproducible and precise firing sequences, which are conditioned by activation patterns. As conductance increases (resulting in net excitation to the network), firing rates can increase and become more uniform with a lower coefficient of variation. Entrainment of low-frequency synchronous behavior involves phase reorganization, so that the optimal phase, i.e., the most excitatory phase, aligns with the temporal characteristics of events in the ongoing input stimulus. Thus, phase differences between the signal from the input signal features and the firing of local clusters disappear due to phase locking. The sequencing of synaptic currents, i.e., outward and inward currents, reduces time jitter in the generation of action potentials within individual neurons, resulting in networks with synchronous activity and increased control of homeostatic regulation.

[0222] In this way, by using the synchronized oscillations of cell assemblies in a spiking neural network, the network becomes highly sensitive to specific frequencies in the (transformed) input signal.

[0223] The present invention relates to the implementation of a mapping method that optimizes the time dynamics of stochastic cell assemblies and allows for synchronization of cell assemblies to improve information transmission in spiking neural networks.

[0224] The present invention further relates to a method for synchronization of stochastic synaptic drive firing rates for excitatory and inhibitory spiking networks with system time delays and stochastic input disturbances.

[0225] The present invention further achieves an increase in cell assembly structural dimensionality by ensuring generalized external synchronization of probabilistic neural networks based on the average autocorrelation function, providing a mechanism that is robust to heterogeneity, connectivity sparsity, and noise.

[0226] The present invention further implements a synchrony mechanism that is compatible with hedonistic approaches such as Hebbian learning, i.e., correlated inputs tend to be strengthened (spike-timing dependent plasticity).

[0227] The present invention further implements a spiking neural network information processor system, which performs application-specific signal information estimation, which eliminates the need for a full reconstruction of the sensor waveform.

[0228] The present invention further relates to a procedure for a system-specific transition from spatiotemporally chaotic activity patterns to pulsed activity, which represents content that forms a coherent entity, such as structure or sensory invariants in a stimulus, or visual objects in a scene, allowing neural circuits to extract relevant information from realistic sensory stimuli in the presence of distractors.

[0229] The present invention further implements computational elements that are capable of generating complex spatiotemporal dynamics that can be extended for specific properties such as changes in frequency response and consequent adjustments in the phase of spiking to perform time-dependent computations in addition to phase-locking of network activity.

[0230] One or more embodiments may be implemented as a computer program product for use with a computer system. The program(s) of the program product may define functions of the embodiments (including the methods described herein) and may be included on various computer-readable storage media. The computer-readable storage media may be non-transitory storage media. Exemplary computer-readable storage media include, but are not limited to, (i) non-writable storage media in which information may be permanently stored (e.g., a read-only memory device in a computer, such as a CD-ROM disk readable by a CD-ROM drive, a ROM chip, or any type of solid-state non-volatile semiconductor memory), and (ii) writable storage media in which changeable information may be stored, such as a hard disk drive or any type of solid-state random-access semiconductor memory, flash memory.

[0231] Two or more of the above embodiments may be combined in any suitable manner.

Claims

1. 1. A method for constructing a spiking neural network, comprising: the spiking neural network comprises a plurality of spiking neurons and a plurality of synaptic elements interconnecting the spiking neurons to form the network, the network being at least partially implemented in hardware; each of the synaptic elements is adapted to receive a synaptic input signal and to apply a weight to the synaptic input signal to generate a synaptic output signal, the synaptic elements being configurable to adjust the weight applied by each of the synaptic elements; each of the spiking neurons is adapted to receive one or more of the synaptic output signals from one or more of the synaptic elements and to generate a spatiotemporal spike train output signal in response to the received one or more synaptic output signals; a responsive local cluster in the network comprising the set of spiking neurons and a plurality of synaptic elements interconnecting the set of spiking neurons; The method comprises: setting the weights of the synaptic elements and the spiking behavior of the spiking neurons in the response local cluster, whereby when an input signal to the response local cluster includes a predetermined oscillation frequency when expressed in the frequency domain, the network state in the response local cluster is a periodic steady state, whereby the network state in the response local cluster is periodic at the predetermined oscillation frequency. A method comprising:

2. 2. The method for configuring a spiking neural network of claim 1, wherein setting the weights of the synaptic elements and the spiking behavior of the spiking neurons in the response local cluster comprises iteratively training the response local cluster by optimizing the weights of the synaptic elements and the spiking behavior of the spiking neurons so that the required periodic steady-state behavior is achieved.

3. 3. A method for constructing a spiking neural network as described in claim 1 or 2, wherein when an input signal to the response local cluster includes the predetermined oscillation frequency, a probability distribution activity or a statistical parameter of the set of neurons in the response local cluster is cyclostationary at the predetermined oscillation frequency.

4. The periodic steady state is expressed by the equation [Equation 1] is the solution of where T is the predetermined time period, Φ(t,τ) is the state transition matrix of the synaptic drive Γ(t) of all neurons in the response local cluster, and F(t) is the state transition matrix of the synaptic drive Γ(t) of all neurons in the response local cluster. [Equation 2] is the deterministic function of the stochastic part of the synaptic drive Γ(t) as given by K(t) is the autocorrelation function of the synaptic drive Γ(t), where K ０ The method for configuring a spiking neural network according to any one of claims 1 to 3, wherein: is the initial condition.

5. the spiking neural network comprises a driver local cluster, the driver local cluster comprising a set of spiking neurons and a plurality of synaptic elements interconnecting the set of spiking neurons, whereby output signals of the driver local cluster serve as input signals to the response local cluster, whereby the driver local cluster and the response local cluster are coupled with a particular coupling strength; and the method comprises:

5. The method for configuring a spiking neural network according to claim 1, further comprising setting the network state in the response local cluster to have a steady state and / or a time-varying state when an input signal from the drive local cluster to the response local cluster does not include the predetermined oscillation frequency when expressed in the frequency domain, or when the specific coupling strength is smaller than a predetermined coupling strength.

6. 6. The method for configuring a spiking neural network of claim 5, wherein setting the weights of the synaptic elements and the spiking behavior of the spiking neurons in the response local cluster comprises iteratively training the response local cluster by optimizing the weights of the synaptic elements and the spiking behavior of the spiking neurons so that the required steady-state behavior and / or time-varying behavior is achieved.

7. 7. A method for configuring a spiking neural network as described in claim 5 or 6, wherein the probability distribution activity or statistical parameter of the set of neurons in the response local cluster is stationary or non-stationary when the response local cluster receives an input signal from the drive local cluster that does not include the predetermined oscillation frequency when expressed in the frequency domain, or when the specific connection strength is smaller than the predetermined connection strength.

8. The steady state is expressed by the equation [Equation 3] is the solution of In the formula, K(t) ∝ is the steady-state value of the autocorrelation function K(t) of the synaptic drive Γ(t), and F(t) is the [Equation 4] is the deterministic function of the stochastic part of the synaptic drive Γ(t) as given by where Φ(t,τ) is the state transition matrix of the synaptic drive Γ(t) of all neurons in the response local cluster, dω is the infinitesimal stochastic change, and E(t) is [Equation 5] the deterministic function defined by The time-varying state is expressed by the matrix equation [Equation 6] is the solution of The matrix equation is the differential Lyapunov matrix equation [Equation 7] where P ｒ and Q ｒ are discretized versions of E and F, and t ｒ The method for configuring a spiking neural network according to any one of claims 5 to 7, wherein ∑ denotes a numerical integration time point.

9. 9. A method for configuring a spiking neural network as claimed in any one of claims 5 to 8, wherein the increase in structural dimensionality of the response local cluster is achieved by ensuring generalized external synchronization between the drive local cluster and the response local cluster, the generalized external synchronization being the coupling of the drive local cluster to the response local cluster with the specific coupling strength that is equal to or greater than the predetermined coupling strength.

10. The generalized external synchronization is the average autocorrelation function 1 / N×{Σ E[Γ(t+τ / 2)Γ(t−τ / 2) Ｔ 10. The method for constructing a spiking neural network of claim 9, wherein τ is the delay, N is the number of neurons in the response local cluster, and the average is over the neuron population.

11. 11. The method for configuring a spiking neural network according to claim 1, wherein the steady-state numerical solution, the time-variable value solution, and / or the periodic steady-state solution is obtained by using feedback connections between the neurons in the response local cluster that result in the synchronization of neuronal activities of the neurons.

12. 1. A spiking neural network for processing an input signal representable in the frequency domain, comprising a plurality of spiking neurons and a plurality of synaptic elements interconnecting the spiking neurons to form the network, the network being at least partially implemented in hardware, each of the synaptic elements is adapted to receive a synaptic input signal and to apply a weight to the synaptic input signal to generate a synaptic output signal, the synaptic elements being configurable to adjust the weight applied by each of the synaptic elements; each of the spiking neurons is adapted to receive one or more of the synaptic output signals from one or more of the synaptic elements and to generate a spatiotemporal spike train output signal in response to the received one or more synaptic input signals; a responsive local cluster in the network comprising the set of spiking neurons and a plurality of synaptic elements interconnecting the set of neurons; A spiking neural network, wherein when an input signal to the response local cluster includes a predetermined first oscillation frequency when expressed in the frequency domain, the probability distribution activity or statistical parameter of the set of neurons in the local cluster is cyclostationary at the predetermined first oscillation frequency.

13. The probability distribution activity or statistical parameters of the set of neurons in the local cluster that are cyclostationary are calculated by the equation [Equation 8] and where T is the predetermined time period, Φ(t,τ) is the state transition matrix of the synaptic drive Γ(t) of all neurons in the response local cluster, and F(t) is the state transition matrix of the synaptic drive Γ(t) of all neurons in the response local cluster. [Equation 9] is the deterministic function of the stochastic part of the synaptic drive Γ(t) as given by K(t) is the autocorrelation function of the synaptic drive Γ(t), where K ０ 13. The spiking neural network of claim 12, wherein: is the initial condition.

14. the spiking neural network comprises a driver local cluster, the driver local cluster comprising a set of spiking neurons and a plurality of synaptic elements interconnecting the set of spiking neurons, whereby output signals of the driver local cluster serve as input signals to the response local cluster, whereby the driver local cluster and the response local cluster are coupled with a particular coupling strength; 14. The spiking neural network of claim 12 or 13, wherein a probability distribution activity or a statistical parameter of the set of neurons in the response local cluster is stationary or non-stationary when the response local cluster receives an input signal from the drive local cluster that does not include the predetermined oscillation frequency when expressed in the frequency domain, or when the particular connection strength is less than a predetermined connection strength.

15. The probability distribution activity or statistical parameters of the set of neurons in the local cluster that are stationary are calculated by the equation [Equation 10] is described by the steady-state numerical solution of In the formula, K(t) ∝ is the steady-state value of the autocorrelation function K(t) of the synaptic drive Γ(t), and F(t) is the [0011] is the deterministic function of the stochastic part of the synaptic drive Γ(t) as given by where Φ(t,τ) is the state transition matrix of the synaptic drive Γ(t) of all neurons in the response local cluster, dω is the infinitesimal stochastic change, and E(t) is [0012] the deterministic function defined by The probability distribution activity or statistical parameters of the set of neurons in the local cluster, which are non-stationary, are calculated by the matrix equation [0013] is described by the time-variable solution of The matrix equation is the differential Lyapunov matrix equation [0014] where P ｒ and Q ｒ are discretized versions of E and F, and t ｒ 15. The spiking neural network of claim 14, wherein ∑ denotes a numerical integration time point.

16. 16. A spiking neural network according to claim 12, wherein the driving local cluster is an input encoder of the spiking neural network, the input encoder converting sampled input signals into spatiotemporal spike trains that are then subsequently processed by the responding local cluster.

17. 1. A method for processing specific frequency portions of an input signal representable in the frequency domain using a spiking neural network, comprising: Use of a spiking neural network according to any one of claims 12 to 16 or a spiking neural network obtained by the method according to any one of claims 1 to 11, providing an input signal in the form of a spatiotemporal spike train to the response local cluster of the spiking neural network, the input signal including one or more frequency portions; processing the input signal using the response local cluster such that the particular frequency portion of the input signal that includes the predetermined vibration frequency has a greater effect on the neurons of the response local cluster than other frequency portions of the input signal; A method comprising:

18. 1. A physical signal inference processor for adaptively processing a physical signal, comprising: a selector and extractor for selecting and extracting specific signal features from said physical signal; a spiking neural network that performs the classification and processing of the physical signal based on the specific signal features extracted from the physical signal; 1. A processor comprising: the processor further comprises an operational block that establishes a current operational context and the optimal feature set; The processor includes a feedback loop to the selector and extractor, the selector and extractor being adaptive in the sense that different signal features can be selected and extracted based on the particular processing task. A processor characterized by:

19. 1. A method for adaptively processing a physical signal, comprising: Receiving a physical signal in a physical signal inference processor according to claim 18; selecting and extracting specific signal features using said selector and extractor; processing the particular signal features using the spiking neural network; determining a current operational context and the optimal feature set using the operational block; A method comprising sending a feedback signal to the selector and extractor to adaptively change the signal features selected and extracted as needed.

Images