Resilient neural network

A spiking neural network with configurable synaptic elements and programmable interconnects addresses inefficiencies in pattern recognition and data fusion, enhancing power efficiency and adaptability in signal recognition systems.

JP2024160227APending Publication Date: 2024-11-13INNATERA NANOSYSTEMS BV
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
JP2024111113
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Priority Date
2018-11-18
Filing Date
2024-07-10
Publication Date
2024-11-13

AI Technical Summary

Technical Problem

Existing automatic signal recognition systems face inefficiencies in pattern recognition and data fusion due to the sequential execution paradigm of microprocessors, leading to high power consumption, latency, and noise sensitivity, especially in power-constrained devices, and traditional deep neural networks struggle with real-time adaptation to changing environments.

Method used

A spiking neural network architecture with configurable synaptic elements and programmable interconnect structures is implemented, partitioned into subnetworks and cores, utilizing asynchronous communication and heterogeneous learning rules to enhance efficiency and resilience.

Benefits of technology

The solution reduces power consumption and latency while improving noise resilience and adaptability, enabling efficient real-time signal recognition and fusion in resource-constrained devices.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024160227000001_ABST
    Figure 2024160227000001_ABST
Patent Text Reader

Abstract

To provide a spiking neural network for classifying input signals.SOLUTION: A spiking neural network 10 comprises a plurality of spiking neurons 1, and a plurality of synaptic elements 2 interconnecting the spiking neurons to form the network. Each synaptic element is adapted to receive a synaptic input signal 11 and apply a weight to the synaptic input signal to generate a synaptic output signal 12, the synaptic elements being configurable to adjust the weight applied by each synaptic element. Furthermore, 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 generate a spatiotemporal spike train output signal in response to the received one or more synaptic output signals.SELECTED DRAWING: Figure 1
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present disclosure relates generally to automatic signal recognition techniques, and more specifically to systems and methods for hardware-resilient deep learning inference accelerators using spiking neurons. [Background technology]

[0002]

[0002] Automatic signal recognition (ASR) refers to the recognition of a signal through the identification of its constituent features. ASR is utilized in a variety of applications, such as, for example, recognizing a speaker's voice and spoken words in voice / speech recognition systems, recognizing arrhythmias in electrocardiograms (ECGs), and determining gesture shapes in motion control systems, among others. ASR is typically performed by characterizing patterns present in short samples of the input signal, and therefore accurate pattern recognition capabilities are essential for an effective ASR system.

[0003]

[0003] Measuring some physical quantity to derive an input signal for an ASR system may require the fusion of data from multiple types of sensors. For example, recognizing hand gestures using a handheld input device may require the fusion of data from an accelerometer to measure movement and data from a gyroscope to measure orientation. Combining data from both sensors allows detection of gestures in three-dimensional space.

[0004]

[0004] In general, data fusion refers to the integration of data collected from different, potentially heterogeneous sources in order to reduce the uncertainty in interpreting the data from each individual source. During the fusion process, it is important that the essential features in the different input signals are adequately represented in the fused signal.

[0005]

[0005] Pattern recognition and fusion are typically performed using microprocessors and / or digital signal processors, both of which implement a stored program architecture. This architecture is inherently inefficient for the analysis of streaming data. On a single processor, pattern extraction and recognition are performed sequentially. This is because pattern extraction and identification are realized in terms of a common simplified instruction set (e.g., RISC or CISC instruction set), resulting in a long execution sequence for every pattern in the signal sample. Complex patterns in the input signal require the use of more complex signal processing algorithms, and in systems that require real-time response from the pattern recognition engine, further necessitate the use of higher clock frequencies for the processor. In power-constrained devices (e.g., portable electronic devices, wearable devices, etc.), this is not feasible. Furthermore, due to the sequential execution paradigm, the latency and power dissipation of pattern recognition operations on a microprocessor increases significantly as the complexity and repertoire of patterns increase. The presence of noise in the input signal further increases the complexity of the analysis, adversely affecting performance and efficiency.

[0006]

[0006] Data fusion is a non-native operation for microprocessors. This means that before input signals or data streams can be fused, their information content must be extracted and then combined with the corresponding content from other streams. As a result, in a microprocessor implementation, each input signal / data source for fusion is processed separately by an independent pattern recognition pipeline, and a rule-based framework subsequently combines the different pattern recognition results. This approach requires multiple invocations of the pattern recognition infrastructure for each input signal or data stream, leading to increased power dissipation. Furthermore, the limitations of microprocessors and digital signal processors (DSPs) of load-store architectures for pattern recognition mean that as the complexity of the patterns increases, so do the power and latency costs of recognizing these patterns. The latency costs can be mitigated by increasing the clock frequency, but this is achieved at the expense of further increased power dissipation. Similarly, the quality of fusion is limited by the complexity of the processing and ASR performed on the input signals, the number of signals to be fused, and the computational power of the microprocessor or DSP. The sequential nature of the processing reduces the throughput of fusion-based ASR, and thus, as the complexity of the patterns in the signals increases, the number of input signals that can be fused using a microprocessor or DSP decreases.

[0007]

[0007] Artificial neural networks in the form of deep neural networks (DNNs) have been proposed as an alternative to microprocessor implementations. DNNs form the basis of many machine learning applications, and the number of applications that utilize DNNs, including speech and image recognition, is growing exponentially, but they suffer from intrinsic limitations, mainly due to the processing of large amounts of data or rapid adaptation to changing environments.

[0008]

[0008] First, hardware deep network accelerators were implemented on standard synchronous digital logic. The high level of parallelism of neural networks is not replicated in the (typical) serial and time-multiplexed processing of digital systems; on the contrary, the computational primitives of hardware DNN emulators, which are realized as analog computing nodes with co-localized memory and processing elements, offer significant improvements in terms of speed, size, and power consumption.

[0009]

[0009] In biological neural network models, each individual neuron communicates asynchronously through sparse events or spikes. In such event-based spiking neural networks (SNNs), only neurons that change state can generate spikes and trigger signal processing in subsequent layers, thus saving computational resources. Spiking neural networks (SNNs) are a promising means of realizing ASR for many different applications.

[0010]

[0010] SNNs encode information in the form of one or more precisely timed (voltage) spikes, rather than as integer or real-valued vectors. The computations for inference (i.e., inferring that a certain feature is present in the input signal) are effectively performed in the analog and time domains. For this reason, SNNs are typically implemented in hardware as fully custom mixed-signal integrated circuits. In addition to having smaller network sizes, this allows them to perform inference functions with orders of magnitude lower energy consumption than their artificial neural network counterparts.

[0011]

[0011] An SNN consists of a network of spiking neurons interconnected by synapses that determine the strength of the connections between the spiking neurons. This strength is represented as a weight, which moderates the influence of the output of the presynaptic neurons on the input to the post-synaptic neurons. Typically, these weights are set in a training process that involves exposing the network to large amounts of labeled input data and gradually adjusting the synaptic weights until the desired network output is achieved.

[0012]

[0012] SNNs can be directly applied to pattern recognition and sensor data fusion, relying on the principle that amplitude-domain, time-domain, and frequency-domain features in an input signal can be encoded into unique spatially and temporally coded spike sequences.

[0013]

[0013] The generation of these sequences relies on the use of one or more ensembles of spiking neurons, an ensemble being a cooperating group of neurons. Each ensemble performs a specific signal processing function, i.e., for example, feature encoding, conditioning, filtering, data fusion, classification. Each ensemble consists of one or more interconnected layers of spiking neurons, and the connectivity within and between layers follows a specific topology. The size of each ensemble (number of neurons), their connectivity (topology and number of synapses), and their configuration (number of layers and weights) depend on the characteristics of the input signal, e.g., dynamic range, bandwidth, time scale, or complexity of the features in the input signal. As a result, the ensembles used in the case of a speech pattern matching system are likely to be different from those used in a handwriting recognition system.

[0014]

[0014] In general, as the complexity of the features to be recognized in the input signal increases, the size of the ensemble required to process them also increases. Spiking neural network hardware can utilize a configurable array of spiking neurons, synapses connected using a programmable interconnect structure that facilitates the implementation of any connection topology. However, to implement a large ensemble, it is necessary that the underlying SNN hardware has at least the required number of neurons and synapses.

[0015]

[0015] Although the basic operations required by SNNs are realized very efficiently by analog electronic circuits, inevitable variations in microelectronic circuits due to the manufacturing process induce distortions in their functional properties, such as resistance, capacitance, gain, time response, etc. In particular, at smaller manufacturing process geometries and lower operating currents, these circuits are increasingly susceptible to quantum effects and external noise, effectively reducing the signal-to-noise ratio and limiting processing performance. The impact of these nonidealities is magnified for large arrays, where driver, bias, and encoder / decoder circuits are shared by a larger number of devices over longer interconnections. Summary of the Invention

[0016]

[0016] To address the shortcomings of the prior art mentioned above, according to a first aspect of the present disclosure, a spiking neural network for classifying an input signal is proposed. The spiking neural network comprises a plurality of spiking neurons and a plurality of synaptic elements interconnecting the spiking neurons to form a network, each synaptic element adapted to receive a synaptic input signal and apply a weight to the synaptic input signal to generate a synaptic output signal, the synaptic element being configurable to adjust the weight applied by each synaptic element. 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 generate a spatio-temporal spike train output signal in response to the received one or more synaptic output signals. The network is partitioned into a plurality of sub-networks, each sub-network comprising a subset of spiking neurons connected to receive synaptic output signals from a subset of the synaptic elements. The sub-networks are adapted to generate a sub-network output pattern signal in response to a sub-network input pattern signal applied to the sub-network. Furthermore, each sub-network forms part of one or more cores in an array of cores, each core consisting of a programmable network of spiking neurons implemented in hardware or a combination of hardware and software, and communication between cores in the core array is arranged via a programmable interconnect structure.

[0017] In one embodiment, the programmable interconnect structure comprises switchable rows and columns.

[0018]

[0018] In one embodiment, the switchable matrix comprises a plurality of routers adapted to route a plurality of output signals from a first core in the core array to a plurality of inputs of a second core in the core array.

[0019] In one embodiment, the programmable interconnect structure forms a packet-switched network between the cores in the core array. In one embodiment, the programmable interconnect structure uses address-event representation. In one embodiment, the programmable interconnect structure uses synchronous or asynchronous communication. In one embodiment, the core array is implemented in an integrated circuit and the programmable interconnect structure comprises a network-on-chip.

[0020] In one embodiment, the network-on-chip is configurable in real time.

[0021]

[0021] In one embodiment, either the learning rules for the synaptic elements, the weight storage mechanism, or the communication protocol are applied heterogeneously in a single core of a core array. In one embodiment, either the learning rules for the synaptic elements, the weight storage mechanism, or the communication protocol are applied heterogeneously across multiple cores in a core array. In one embodiment, the learning rule configuration of one or more of the cores in a core array is implemented using a learning rule block. In one embodiment, cores using the same learning rule use a common learning rule block that implements the learning rule.

[0022] In one embodiment, the distribution of learning rules within and / or across each core in the core array is dynamically configurable at run-time.

[0023] In one embodiment, one of the sub-networks is part of a classifier for classifying the sub-network's input signal.

[0024] In one embodiment, one of the sub-networks is part of an ensemble of classifiers for classifying input signals to that sub-network. In one embodiment, the weights of the synaptic elements are configured using stochastic weight updates.

[0025] In one embodiment, the weights of the synaptic elements are bounded by bound values, which are stochastic values.

[0026]

[0026] According to a second aspect of the present disclosure, an integrated circuit is disclosed comprising a spiking neural network implemented in the core array of the first aspect of the present disclosure.

[0027]

[0027] According to a third aspect of the present disclosure, a method for partitioning a spiking neural network and mapping it onto a core array is disclosed, where the spiking neural network comprises a plurality of spiking neurons and a plurality of synaptic elements interconnecting the spiking neurons to form a network, each synaptic element adapted to receive a synaptic input signal and 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 synaptic element, and each of the spiking neurons adapted to receive one or more of the synaptic output signals from one or more of the synaptic elements and generate a spatiotemporal spike train output signal in response to the received one or more synaptic output signals. Further, the core array comprises a plurality of cores, each core being comprised of a programmable network of spiking neurons implemented in hardware or a combination of hardware and software, and communication between the cores in the core array is arranged via a programmable interconnect structure. Here, the method comprises partitioning the network into a plurality of sub-networks, each sub-network comprising a subset of spiking neurons connected to receive synaptic output signals from a subset of synaptic elements, the sub-networks adapted to generate sub-network output pattern signals from the subset of spiking neurons in response to sub-network input pattern signals applied to the subset of synaptic elements, each sub-network being mapped onto one or more cores.

[0028] In one embodiment, the partitioning of a spiking neural network into sub-networks implemented on one or more cores is determined by a mapping method comprising constraint-driven partitioning, where the constraints are performance metrics linked to a function of each respective sub-network.

[0029]

[0029] According to a fourth aspect of the present disclosure, a spiking neural network is disclosed. The spiking neural network comprises a plurality of spiking neurons implemented in hardware or a combination of hardware and software, and a plurality of synaptic elements interconnecting the spiking neurons to form a network, each synaptic element adapted to receive a synaptic input signal and apply a weight to the synaptic input signal to generate a synaptic output signal, the synaptic elements are configurable to adjust the weight applied by each synaptic element, and 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 generate a spatiotemporal spike train output signal in response to the received one or more synaptic output signals. Furthermore, the learning rule, weight storage mechanism, and / or communication protocol are non-uniformly applied across the neurons and / or synaptic elements in the spiking neural network.

[0030]

[0030] According to a fifth aspect of the present disclosure, a spiking neural network is disclosed, comprising a spiking neuron and a synaptic element connected to the spiking neuron. Here, the synaptic element comprises a first and a second receptor adapted to receive a synaptic input signal, and the first and the second receptor generate a first and a second receptor signal, respectively, based on the synaptic input signal. The synaptic element applies a weight to the first receptor signal to generate a synaptic output signal, and the synaptic element is configurable to adjust the weight applied by the synaptic element based at least on the second receptor signal. The spiking neuron is adapted to receive the synaptic output signal from the synaptic element and generate a spatiotemporal spike train output signal at least in response to the received synaptic output signal.

[0031] In one embodiment, the neuron emits a control signal which, together with a second receptor signal, modulates the weight applied by the synaptic element.

[0032] In one embodiment, the control signal is a backpropagation signal.

[0033] In one embodiment, the neuron comprises dendrites, an axon, and a cell body, and the control signal originates from the dendrites and / or axon and / or cell body of the neuron.

[0034] In one embodiment, the control signal comprises one or more spikes generated by an action potential of the neuron. In one embodiment, the decay time of the first receptor is faster than the decay time of the second receptor. In one embodiment, the first receptor generates a source or sink current for the spiking neuron.

[0035] In one embodiment, the first receptor comprises a low-pass filter. In one embodiment, the second receptor forms a voltage-gated receptor. In one embodiment, the second receptor comprises a low-pass filter, a band-pass filter, a high-pass filter, and / or an amplifier. In one embodiment, the first receptor is an AMPA receptor, or a GABA receptor, or an NMDA receptor.

[0036]

[0036] According to a sixth aspect of the present disclosure, a method for adjusting weights of synaptic elements in a spiking neural network is disclosed, the spiking neural network comprises a spiking neuron connected to a synaptic element. The synaptic element comprises a first and a second receptor adapted to receive a synaptic input signal, the first and the second receptor receive the synaptic input signal and generate a first and a second receptor signal based on the synaptic input signal, respectively. The synaptic element applies a weight to the first receptor signal to generate a synaptic output signal. The weight of the synaptic element is adjusted based on at least the second receptor signal, and the spiking neuron receives a synaptic output signal from the synaptic element and generates a spatiotemporal spike train output signal at least in response to the received synaptic output signal.

[0037] In one embodiment, the neuron emits a control signal which, together with a second receptor signal, modulates the weight applied by the synaptic element.

[0038] In one embodiment, the control signal is a backpropagation signal.

[0039] In one embodiment, the neuron comprises dendrites, an axon, and a cell body, and the control signal originates from the dendrites and / or axon and / or cell body of the neuron.

[0040] In one embodiment, the control signal comprises one or more spikes generated by an action potential of the neuron. In one embodiment, the decay time of the first receptor is faster than the decay time of the second receptor. In one embodiment, the first receptor generates a source or sink current for the spiking neuron.

[0041] In one embodiment, the first receptor comprises a low-pass filter. In one embodiment, the second receptor forms a voltage-dependent receptor. In one embodiment, the second receptor comprises a low-pass filter, a band-pass filter, a high-pass filter, and / or an amplifier. In one embodiment, the first receptor is an AMPA receptor, or a GABA receptor, or an NMDA receptor.

[0042]

[0042] According to a seventh aspect of the present disclosure, an integrated circuit is disclosed that includes the spiking neural network of the fifth aspect of the present disclosure.

[0043]

[0043] According to an eighth aspect of the present disclosure, a method for configuring a spiking neural network to reduce noise effects in the spiking neural network is disclosed. The spiking neural network comprises a plurality of spiking neurons implemented in hardware or a combination of hardware and software, and a plurality of synaptic elements interconnecting the spiking neurons to form a network. Each synaptic element is adapted to receive a synaptic input signal and apply a weight to the synaptic input signal to generate a synaptic output signal, the synaptic element being configurable to adjust the weight applied by each synaptic element. 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 generate a spatiotemporal spike train output signal in response to the received one or more synaptic output signals. A synaptic drive Γ of one spiking neuron i in the spiking neural network is i is a time-dependent function that describes the total transfer function of all synaptic elements connected within a neuron. i obtaining an expression for the variance of each synaptic drive of Γ, the variance expression depending on the control parameter. iThe variance of each synaptic drive in can be adjusted by adjusting the control parameter. The control parameter is the synaptic drive Γ i The variance of each synaptic drive of neuron i in the spiking neural network is adjusted to be below a predetermined value, so that the synaptic drive of each neuron Γ i However, the synaptic drive Γ with minimal noise effect i Equilibrium point

number

[0044] In one embodiment, the synaptic drive Γ of a spiking neuron i in a spiking neural network is i is expressed as Γ=(Γ1,...,Γ n ) T can be written in vector form for all neurons in a spiking neural network as, where n is the number of neurons in the spiking neural network, and Γ satisfies dΓ=N(Γ(t))dt+σ(Γ(t))dω(t)

[0045] where N(Γ(t)) is the nominal matrix part, σ(Γ) is the state-dependent noise matrix for a Gaussian white noise process dω(t), and ω is a Wiener process describing the noise in spiking neural networks, where the expression for dΓ can be rewritten in terms of the variance-covariance matrix K(t) of Γ(t) that satisfies the continuous-time algebraic Lyapunov equation N(Γ(t))K(t)+K(t)[N(Γ(t))] T +σ(Γ(t))[σ(Γ(t))] T =0

[0046] Synapse-driven Γ i The step of obtaining an expression for the variance of each synaptic drive of comprises determining the diagonal values ​​of a variance-covariance matrix K(t).

[0045]

[0047] In one embodiment, the synapse-driven Γ i Equilibrium point Γ i * is in vector form for all neurons in a spiking neural network,

number

[0046]

[0048] In one embodiment, the control parameters are tuned by computer simulation.

[0047]

[0049] According to a ninth aspect of the present disclosure, a spiking neural network is disclosed, comprising a plurality of spiking neurons implemented in hardware or a combination of hardware and software, and a plurality of synaptic elements interconnecting the spiking neurons to form a network. Each synaptic element is adapted to receive a synaptic input signal and apply a weight to the synaptic input signal to generate a synaptic output signal, the synaptic element being configurable to adjust the weight applied by each synaptic element. 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 generate a spatiotemporal spike train output signal in response to the received one or more synaptic output signals. A synaptic drive Γ of one spiking neuron i of the spiking neurons i in the spiking neural network is iis a time-dependent function that describes the total transfer function of all synaptic elements connected within a neuron. i The variance of each synaptic drive of neuron i in the spiking neural network is the synaptic drive of each neuron Γ i However, the synaptic drive Γ with minimal noise effect i Equilibrium point Γ i * below a given value, such that it is bounded around

[0048]

[0050] 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: [Brief description of the drawings]

[0049] [Figure 1]

[0051] 1 illustrates an exemplary neural network consisting of neurons and synaptic elements. [Diagram 2]

[0052] 1 illustrates a schematic of a spiking neural network within a microcontroller integrated circuit. [Diagram 3]

[0053] A high-level architecture for a learning system with an array of neuromorphic cores is presented. [Figure 4]

[0054] 1 shows a graphical representation of the boosting algorithm in a single core. [Diagram 5]

[0055] 1 shows a graphical representation of combining multiple ensemble systems within a single core. [Figure 6]

[0056] 1 shows a graphical representation of combining multiple ensemble systems across multiple cores in a multi-core implementation. [Figure 7A]

[0057] A conceptual diagram of synapse structure is shown. [Figure 7B]

[0058] A conceptual diagram of the neurosynapse structure is shown. [Figure 8]

[0059] Shows the net activity of an infinite network. [Figure 9]

[0060] Show the net activity of a finite network.

[0050]

[0061] The figures are intended for illustrative purposes only and do not serve as a limitation on the scope or protection afforded by the claims.

[0051]

[0062] Certain embodiments are described in more detail below, however, it should be understood that these embodiments should not be construed as limiting the scope of protection of the present disclosure.

[0052]

[0063] FIG. 1 is a simplified diagram of a neural network 10. Neurons 1 are connected to each other via synaptic elements 2. To avoid cluttering the drawing, only a few neurons and synaptic elements are shown (and only some are numbered). The connection topology shown in FIG. 1, i.e. the way in which the synaptic elements 2 connect the neurons 1 to each other, is merely an example, and many other topologies may be adopted. Each synaptic element 2 can send a signal to an input of a neuron 1, and each neuron 1 receiving the signal can process the signal and subsequently generate an output, which is transmitted to other neurons 1 via further synaptic elements 2. Each synaptic element 2 is assigned a certain weight, which is applied to each synaptic input signal it receives or receives so as to generate a weighted synaptic output signal. The weight of a synaptic element is thus a measure of the type of causal relationship between two neurons 1 connected by a synaptic element 2. This relationship can be causal (positive weight), anticausal (negative weight), or non-existent (zero weight).

[0053]

[0064] The neurons 1 and synaptic elements 2 may be implemented in hardware, for example using analog circuit elements or digital hardwired logic circuits. They may also be implemented partly in hardware and partly in software, or completely in software. A hardware or at least partial hardware implementation is preferred, i.e. hardware circuits or elements are used to perform the functions of individual neurons, rather than using large processors running software that mimics individual neurons. These (partial) hardware implementations achieve faster processing, allowing, for example, much faster pattern recognition and event-driven processing, where blocks of neurons and synaptic elements are activated only when necessary.

[0054]

[0065] The neural network 10 may be a spiking neural network. In this case, the neuron 1 is a spiking neuron that generates a neuron output signal in the form of one or more spikes or neuron-generated events. The 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 a received input signal, i.e., the synaptic output signal received by the neuron from a synaptic element is accumulated, integrated, or otherwise processed to modify the membrane potential. When the weight of a synaptic element 2 is positive, the synaptic output signal received from that synaptic element excites the spiking neuron 1 that receives the signal, raising their membrane potential. When the weight of a synaptic element 2 is negative, the synaptic output signal received from that synaptic element inhibits the spiking neuron 1 that receives the signal, lowering their membrane potential. When the weight of a synaptic element 2 is zero, the synaptic output signal received from that synaptic element has no effect on the spiking neuron 1 that receives the signal.

[0055]

[0066] When the membrane potential of a spiking neuron 1 reaches a threshold, the neuron fires, generating a spike upon firing, and the membrane potential decreases as a result of the firing. If the membrane potential subsequently 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 in response to input signals received from connected synaptic elements 2, the spikes forming a spatio-temporal spike train. Since a spiking neuron 1 fires only when its membrane potential reaches a predefined threshold, the encoding and processing of temporal information is built into the neural network 10. In this way, a spatio-temporal spike train is generated in the spiking neural network 110, which is a temporal sequence of spikes generated by the spiking neurons 1 of the network 10.

[0056]

[0067] The temporal properties of the spike train encode the amplitude and frequency characteristics 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 output of the neuron, the latency between successive spikes from the same neuron, and the number of spikes fired by the neuron during the duration that the input stimulus is applied.

[0057]

[0068] The synaptic elements 2 may be configurable, for example, such that the weights of each of the synaptic elements can be changed, for example, by training the neural network 10. The neurons 1 may be configurable to respond to signals from the synaptic elements. For example, in the case of a spiking neural network, the neurons 1 may be configured such that a particular signal increases or decreases the membrane potential, the time it takes for the membrane potential to naturally decay towards the resting membrane potential, the value of the resting membrane potential, the threshold that triggers a spike of the spiking neuron 1. The configuration of the neurons 1 may be kept constant, for example, during training, or may be variable and may be set in the training of the neural network 10 for a particular training set.

[0058]

[0069] The input signal 11 can be, for example, a number of different sampled input signals, or a spatio-temporal spike train. The input can be analog-to-digital converted values ​​of the signal samples, or for example the digital values ​​of the samples in case of an analog or digital integrator, or the analog values ​​of the samples in case of an analog integrator.

[0059]

[0070] The output signal 12 of the neural network 10 is, for example, a spatiotemporal spike train, which can be read out from the output neurons 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.

[0060]

[0071] 2 illustrates one embodiment of a high-level architecture of a microcontroller integrated circuit 100 with a spiking neural network 110. In this context, the microcontroller 100 is an economical means of data collection, sensing, pattern recognition, and actuation of physical signals.

[0061]

[0072] 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 to be 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 of 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 neuron 1 or synaptic element 2 of the spiking neural network 110, or the network topology.

[0062]

[0073] Each neuron 1 may have a set of configuration parameters that control the exact firing behavior of that neuron 1. For example, a neuron may be designed with a firing threshold, which represents a threshold of voltage, energy, or other variable that accumulates in the neuron as a result of receiving an input, and the neuron generates an output spike (such as a voltage, current, or energy spike) when the accumulated variable meets or exceeds the firing threshold. The neuron may implement an integral function that integrates the input to the neuron to determine an adjustment to the accumulated variable. Furthermore, the neuron may also be designed with (a) a leakage rate, which represents the rate at which the accumulated variable in 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 when there is no input signal to the neuron; (c) an integration time constant, which represents the time over which the input signal is integrated to determine an increase in the accumulated variable in the neuron; (d) a refractory level, which represents the value of the accumulated variable in the neuron immediately after the neuron fires; and (e) a refractory period, which represents the period of time required for the accumulated variable in the neuron to rise to the 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, leakage rate, integration time constant, and refractory period to match the energy content of a critical input signal feature, neuron 1 will generate one or more precisely timed spikes when stimulated with an input signal containing that feature.

[0063]

[0074] The configuration parameters of synaptic element 2 include the weight and gain of synaptic element 2. The weight of synaptic element 2 is typically used to tune synaptic element 2, while the gain of synaptic element 2 is used for signal amplification in hardware and typically relates to a low pass filter implementation. Typically, the gain is fixed at the initialization of network 110, while the weight can change based on the evolution / training of spiking neural network 110.

[0064]

[0075] The microcontroller integrated circuit 100 further comprises a microprocessor core 101 for performing computations and control of the integrated circuit 100. For example, the microprocessor core 101 may oversee communications between a memory map control and configuration interface 113 and the memory device 102.

[0065]

[0076] The memory device 102 may be any computer-readable storage medium. The memory device 102 may 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 may be permanently stored (e.g., a read-only memory device in a computer, such as a CD-ROM disk, a ROM chip, or any type of solid-state non-volatile semiconductor memory readable by a CD-ROM drive), 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.

[0066]

[0077] 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.

[0067]

[0078] 3 shows a high-level architecture for a learning system having multiple neuro-synaptic cores 210 arranged in an array of cores 200. Each core 210 comprises a network of neurons 1 implemented in hardware, the neurons being interconnected by synaptic elements 2. A single core 210 may implement a complete spiking neural network, or a portion of a spiking neural network forming a separate sub-network. In this way, a large spiking neural network may be partitioned into several smaller sub-networks, each sub-network being implemented in one of the cores 210 of the array 200. In one embodiment, the core 210 may implement the spiking neural network 110 shown in FIG. 2, with associated input data ports 111, output ports 112, and / or control and configuration interfaces 113.

[0068]

[0079] By partitioning the large spiking neural network 110 into smaller sub-networks and implementing each of the sub-networks on one or more cores 210, each with its own requisite circuitry, some of the non-idealities of the circuitry operating at smaller process geometries and lower operating currents are mitigated, especially for large arrays. Thus, the core-based implementation approach reduces the effects of physical non-idealities.

[0069]

[0080] Sub-networks or ensembles of neurons forming cooperating groups can form classifiers, ensembles of classifiers, groups of neurons, etc. that deal with, for example, data transformation, feature coding, or just classification.

[0070]

[0081] In such a regime, a large network of ensembles is partitioned and mapped onto an array of cores, each of which contains a programmable network of spiking neurons. As a result, each core implements a single ensemble, or multiple smaller ensembles (related to the number of neurons and synapses in the core), or in the case of a large ensemble, only a part of the single ensemble, with the other parts being implemented on other cores of the array. The manner in which the ensembles are partitioned and mapped onto the cores is determined by a mapping method. The mapping method may comprise a constraint-driven partitioning. The constraints may be performance metrics linked to the functionality of each respective sub-network. The performance metrics may depend on power domain limitations, memory structure, memory access, time constants, biases, technology limitations, resilience, the level of mismatch allowed, and network or physical artifacts.

[0071]

[0082]

[0083] The periphery of the array 200 includes rows of synaptic circuits that mimic the action of the cell body and axon hillock of a biological neuron. Additionally, each neurosynaptic core 210 in the array 200 has a local router 202 that communicates with routers 202 of other cores 210 in the dedicated real-time reconfigurable network-on-chip.

[0072]

[0084] The local routers 202 and their connections 201 form a programmable interconnection structure between the cores 210 of the core array 200. The cores 210 are connected through switchable matrices. Thus, the different cores 210 of the core array 200 are connected via the programmable interconnection structure. In particular, different parts of the spiking neural network implemented on the different cores 210 of the core array 200 are interconnected via the programmable interconnection structure. In this way, quantum effects and external noise only act on each core individually and not on the whole network. Thus, their effects are mitigated.

[0073]

[0085] The implemented spiking neural network 110 on the core array 200 may have high modularity in the sense that the spiking neural network 110 has dense connections between neurons within a core 210, but sparse connections between different cores 210. In this way, noise and quantum effects are further reduced between cores, yet still allowing sub-networks to increase classification accuracy, for example by allowing high complexity.

[0074]

[0086] The programmable interconnect structure can form a packet-switched network between the cores 210 in the core array 200. For example, the local router 202 can group data into packets, which are then transmitted over the connections 201. These connections 201 can form a digital network. The packets can consist of a header and a payload. The data in the header is used by the local router 202 to direct the packet to a destination core 210 where the payload in the packet is extracted and used. The data can be, for example, the output of one of the subnetworks of a spiking neural network partitioned and implemented on one or more cores 210 of the core array 200. The data can also be input from a sensor or another external device.

[0075]

[0087] The programmable interconnect structure may also use an address-event representation in which the addresses of the neurons to be connected are sent to the different cores.

[0076]

[0088] Furthermore, either synchronous or asynchronous communication may be used, an example of synchronous communication being any clock-based communication, an example of asynchronous communication being e.g. a handshake protocol.

[0077]

[0089] The cores in the core array may form a network-on-chip on the microcontroller integrated circuit 100. The network-on-chip improves the scalability and power efficiency of the microcontroller integrated circuit 100.

[0078]

[0090] The topology of the network-on-chip, e.g., mesh, torus, tree, ring, star, is selected based on the parameters of the system. The routing algorithm used depends on both the topology and the characteristics of the network 110. The complexity of the routing algorithm may be increased, for example, by considering multiple shortest paths to a target and then randomly selecting one of these paths for each flit (resulting in better network load). Further efficiency can be obtained by using information and heuristics about the network load at any given time. Other types of routing approaches exist, including static routing tables or source-based routing.

[0079]

[0091] The network-on-chip may be reconfigurable in real time or may be statically defined during the production phase. When the network-on-chip is reconfigurable in real time, the configuration of the cores 210 in the core array 200 and their interconnect structure configurations can be changed. This change can be based on, for example, changes in the inputs or outputs of the microcontroller integrated circuit, different requirements for classification accuracy or stability, the evolution of the network based on its learning rules, and changes in communication protocols.

[0080]

[0092] The present invention encompasses that the neurosynaptic core 210 can apply the same kind of learning rule, weight storage type, and / or communication protocol to the synaptic matrix. In other words, the same kind of learning rule, weight storage type, and / or communication protocol can be set within one core.

[0081]

[0093] The entire spiking neural network 110 may be implemented as one single entity implemented within one core 210 (e.g., as a crossbar structure of neurons 1 and synapses 2 with all feedback connections in place to implement any recurrent neural network structure (or a simple feedforward network with no feedback signals). Alternatively, the entire spiking neural network 110 may be spread across multiple cores 210, where each core 210 implements a particular layer(s), or part of a layer, e.g., in the context of a deep neural network. The decision on how to split the network is entirely based on the application (e.g., radar, lidar, imaging, ultrasound, biopotential sensing).

[0082]

[0094] In one embodiment, all neurosynaptic cores 210 may apply the same learning rules, weight storage types, and / or communication protocols to each of their synaptic matrices.

[0083]

[0095] In one embodiment, neurosynaptic cores 210 may apply different learning rules, weight storage types, and / or communication protocols to each of their synaptic matrices.

[0084]

[0096] In one embodiment, the neurosynaptic array of the entire spiking neural network 110 can be organized as a single core 210 that implements heterogeneous learning rules depending on the optimal performance-power-area tradeoff.

[0085]

[0097] In another embodiment, different regions of the synaptic matrix in the core 210 may be configured with the same learning rule that may be implemented by a common learning rule block driven by the same configuration register. The size of this region of the synaptic matrix in the core 210 may be configurable at design time to create specialized circuits, or may be dynamically configurable at run time, such that the synaptic matrix in the core 210 has a heterogeneous clustering of synaptic circuits, with each cluster implementing a different learning rule. This heterogeneous clustering may require a special mapping or composition algorithm to map each neuronal unit to a determined cluster of the synaptic matrix in the core 210 based on the applied learning rule.

[0086]

[0098] A learning rule block may be implemented for each synaptic element, for each group of synaptic elements, or for each core 210 of the core array 200. A learning rule block may consist of circuitry implemented in the configuration interface 113, for example.

[0087]

[0099] In another embodiment, different sets of one or more cores 210 in the core array 200 may be configured with the same learning rule, which may be implemented by a common learning rule block driven by the same configuration register. The size of this set of cores 210 in the core array 200 may be configurable at design time to create a specialized set of cores 210, or may be dynamically configurable at run time, such that the core array 200 has a heterogeneous clustering of cores 210, with each cluster implementing a different learning rule. This heterogeneous clustering may also require a special mapping or composition algorithm to map each neuronal unit to a determined cluster of cores 210 based on the applied learning rule.

[0088]

[0100] It is possible to implement different / heterogeneous learning rules not only across cores 210 but also within a single core 210 .

[0089]

[0101] The design is more flexible because different sets of cores 210 in the core array 200, and different regions of the synaptic matrices within the cores 210, can implement different learning rules. Furthermore, by applying different learning rules, a particular input will have a different effect on a particular region of the core array 200. In this way, optimization can be performed across different regions where different learning rules are implemented, resulting in better performance of the overall network 110. Finally, different learning rules can perform better for different input signals, resulting in the network being tailored to which learning rules are needed in which parts of the network.

[0090]

[0102] Furthermore, in one embodiment, each region of the synaptic matrix within one core 210, or a set of one or more cores 210, in the core array 200 may also implement a different weight storage mechanism, such as digital memory, capacitive storage, bistable mechanism, etc., and each region may have different power requirements.

[0091]

[0103] A set of cores 210 in a core array 200 can implement different communication protocols, which makes the design more flexible. Thus, the cores 210 of a core array 200 allow, among other things, the same or different learning rules, weight storage types, and / or communication protocols to be applied to each of their synaptic matrices, allowing for a practical and flexible implementation of the spiking neural network 110 in hardware.

[0092]

[0104] The brain's cognitive abilities emerge from ensemble neuronal activity or joint forms of computation, i.e., cooperating groups of neurons (sub-networks or ensembles) trigger learning, enhance global perceptual capabilities, and create functional neural states that compensate for the reduction of sensory modalities; the joint activity of groups of neurons overcomes the unreliable stochastic nature of single neuronal firing.

[0093]

[0105] Multiple architectures that attempt to mirror aspects of biology can then be used, i.e., multiple (parallel) classifiers, operating on the same stimuli or on different parts or features of the stimuli to improve various machine learning complexities, such as feature selection, confidence estimation, missing features, incremental learning, error correction, etc. The feature space of the input signal can be partitioned using multiple classifiers. Ensembles of similarly constructed neural networks can be utilized to improve classification performance. Using boosting procedures, powerful classifiers (with arbitrarily low error for binary classification problems) can be constructed from an ensemble of classifiers, where the error of any of these classifiers that make up the ensemble is simply better than that of random guessing.

[0094]

[0106] As a result, a number of ensemble systems can be devised, including, for example, bagging, random forests (ensembles of decision trees), composite classifier systems, mixture of experts (MoE), stacked generalization, consensus aggregation, combination of multiple classifiers, dynamic classifier selection, classifier fusion, committees of neural networks, and classifier ensembles, based on (i) the selection of training data for the individual classifiers, (ii) the particular procedure used to generate the ensemble members, and / or (iii) the combination rules for obtaining the ensemble decision.

[0095]

[0107] Figure 4 shows a graphical representation of the boosting algorithm within a single core 210 of the core array 200. In Figure 4, multiple classifiers 300 are shown implemented within a single core 210 of the core array 200. The classifiers 300 form an ensemble of classifiers 300.

[0096]

[0108] The ensemble of classifiers 300 is one example of a sub-network of spiking neural networks 110 that may be implemented on a single core 210 of the core array 200.

[0097]

[0109] The classifier 300 is assumed to have a set of output neurons (one for each class), each of which fires events (spikes) according to its firing probability distribution. A particular classifier 300 may comprise different layers 301, 302, 303 of neurons 1 (or neurons 1 and synaptic elements 2) connected by synaptic elements 2. Each layer may have different functionality, for example, layer 301 may perform data transformation, layer 302 may perform feature coding, layer 303 may perform classification. There may also be an output layer present in each classifier 300. The output layer may be connected to a local router 202 of a particular core 210 of the core array 200. Using the connections 201 between the local routers 202, the output of a particular classifier 300 or an ensemble of classifiers 300 may be directed to other cores 210 in the core array 200 and thus to other sub-networks of the spiking neural network 110.

[0098]

[0110] FIG. 5 shows a graphical representation of combining multiple ensemble classifier systems within a single core 210 of a core array 200.

[0099]

[0111] First, an ensemble 401 of classifiers 300 implemented based on a boosting algorithm receives an input 404 and then sends its output to a second ensemble 402 implemented based on a mixture of experts algorithm. The synaptic elements 410 between layers may be trained using triple-based spike timing dependent plasticity (T-STDP) as a learning rule. In STDP, if an input spike to a neuron 1 tends, on average, to occur just before an output spike of that neuron 1, that particular input is made somewhat stronger. If an input spike tends, on average, to occur just after an output spike, that particular input is made somewhat weaker by adjusting the weight. Thus, an input that may be responsible for the excitation of a post-synaptic neuron 1 is made more likely to contribute in the future by adjusting the weight of a particular synaptic element 2 up, while an input that is not responsible for a post-synaptic spike is made less likely to contribute in the future by lowering the weight of the particular synaptic element 2. This STDP method is pair-based since it uses pairs of spikes. T-STDP differs from conventional STDP in that triplets of spikes are used.

[0100]

[0112] A gating network 408 is used to set data partitions between experts during training and combine their outputs. The gating network 408 can be used during training as a so-called "golden reference" that defines a filter or transfer function for pooling different expert results and acts as a teacher signal. Experts compete to learn the training patterns and the gating network 408 mediates the competition. A pooling and combining system 409 pools and combines the outputs of all the different experts into one output 406 and outputs the output 406 of the mixed expert layer as an input to the third ensemble 403 of the classifier 300 implemented based on the boosting algorithm. The output 407 of the third ensemble 403 can be output to the local router 202 of the core 210 and can be sent via the interconnect 201 to other cores 210 of the core array 200 or to the output port 112 of the spiking neural network 110.

[0101]

[0113] FIG. 6 shows a graphical representation of combining multiple ensemble systems across multiple cores in a multi-core implementation.

[0102]

[0114] First, in the first core 510A, an ensemble 401 of classifiers 300 implemented based on a boosting algorithm receives an input 404 and then sends its output to a second ensemble 402 implemented based on a mixture of experts algorithm. The pooling and combining system 409 of the second ensemble 402 outputs its output to a first local router 502A of the first core 510A, which then sends this output via the interconnects 501A, 501B towards a second local router 502B of the second core 510B. This data is then used as input for a third ensemble 403 present on the second core 510B. This system therefore differs from the system shown in FIG. 5 in that the spiking neural network 110 is partitioned in a certain way and mapped onto the core array 200 in a different way. That is, in this embodiment, the ensemble is split into two different cores 510A, 510B, each with its own required circuitry.

[0103]

[0115] Devising a dynamic model that represents the subsequent merging of classifier outputs enforces a scheme for allocating each unique unlabeled pattern to the best-fitting classifier. The classifiers are defined as probabilistic winner-takes-all, where only a single neuron can spike for any data presented at the input. A trainable combiner can determine which classifiers are accurate in which parts of the feature space and can then be combined accordingly.

[0104]

[0116] Without loss of generality, in boosting algorithms, the equivalence of weak and strong probabilistically approximately correct (PAC) learning models is a prerequisite, i.e. they generate different distributions on which non-identical sub-hypotheses are trained. Probabilistically approximately correct (PAC) learning is a framework for mathematical analysis in computational learning theory in machine learning. Here, a learner receives samples and has to select a generalization function from a class of possible functions. The goal is that with high probability (the "probabilistic" part), the selected function has a low generalization error (the "approximately correct" part).

[0105]

[0117] In contrast, adaptive mixtures of experts obtain performance improvements by allocating different subtasks to diverse learners. Most ensemble learning algorithms, e.g., stacking, first train base predictors and then attempt to tune the combined model.

[0106]

[0118] Mixture of experts and boosting are intended for different categories of problems, which bring their own advantages and disadvantages. In boosting, the distributions are primarily specified to allow individual classifiers to proceed as experts on data patterns for which the former classifiers are incorrect or inconsistent. However, in mixture of experts, the data patterns are separated into basic but consistent subsets, and then the learning process required for the individual subsets is less complicated than for the original data pattern.

[0107]

[0119] The same gating functions are used both to set the data partition among the experts during training and to combine the outputs. Gating networks are typically trained using the expectation-maximization (EM) algorithm on the original training data.

[0108]

[0120] Each ensemble or combination of ensembles may be implemented in one or more cores 210.

[0109]

[0121] In one embodiment, shown in FIG. 7A, a hardware implementation of a synaptic matrix having synaptic elements 652A-D in core 210 is shown. Each of synaptic elements 652A-D is assigned a respective weight 653A-D. For first weight 653A, w 2,1 , for the second weight 653B w 1,1 , for the third weight 653C w 2,2 , 4th weight 653D w 1,2 can be written as:

[0110]

[0122] The input signal 650A to synaptic elements 652A, 652B can be written as X1, and the input signal 650B to synaptic elements 652C, 652D can be written as X2. This may be a voltage spike as part of a spatiotemporal spike train. At each synaptic element 652A, 652B, the input signal 650A is multiplied by a respective weight 653A, 653B. At each synaptic element 652C, 652D, the input signal 650B is multiplied by a respective weight 653C, 653D. The output signal 651A is expressed as Σ i w 1,i X i and the output signal 651B can be written as Σ i w 2,i X i These output signals 651A, 651B may be sent to one or more neurons.

[0111]

[0123] Weights 653A-D of synaptic elements 652A-D may be set by using one or more enable lines 654 and error lines 655. Enable line 654 enables weights 653A-D and sets them to correct values ​​known to the system, while error line 655 updates weights 653A-D based on perceived errors in the weight settings. These signals may be analog or digital.

[0112]

[0124] In another embodiment, each neurosynaptic ensemble may be implemented as a combination of multiple (synaptic) receptors, dendrites (inputs) and axonal somas (outputs) of a biological neuron, as shown in FIG.

[0113]

[0125] 7B shows one or more synaptic elements 602. The spike 606 portion of the spatiotemporal spike train enters the synaptic element 602. At the input of the synaptic element, three receptors are accessible: an NMDA receptor (rNMDA) 607A provides activity-dependent modification of the synaptic weight w, an AMPA receptor (rAMPA) 607B facilitates fast synaptic currents to drive the cell body with a source current, and finally, a GABA receptor (rGABA) 607C can facilitate fast or slow inhibitory synaptic currents with a sink current, depending on whether GABAA or GABAB receptors are used. The output of rAMPA 607A and / or rGABA 607C can be used as input to an amplifier 608.

[0114]

[0126] In this embodiment, one may also have rNMDA 607A and either rAMPA 607B or rGABA 607C. Importantly, receptors 607A-C with different temporal components may be used. rAMPA 607B and rGABA 607C may be implemented as low pass filters. On the other hand, rNMDA 607A may be implemented as a voltage-gated receptor with both amplification and filtering, connected to the presynaptic circuit of the implemented learning rule. Filtering may consist of band pass, low pass, and high pass filters.

[0115]

[0127] The output of the amplifier 608 is sent to a dendrite 601A that forms part of the neuron connected to the synaptic elements 602. The output from all synaptic elements 602 is then integrated over time by an integrator 609. The integrated signal is sent via a receptor 610 to the axon cell body 601B of the neuron, which consists of an axon 603 and a cell body 604. Signals from other clusters of synaptic elements 605, 606 may also enter the axon cell body 601B at this point. A different integrator 611 may be placed at the axon cell body 601B. When a spike is released by the axon 603 generating an action potential, it leaves the axon cell body 601B and may be sent towards other synaptic elements, for example as a spike in a spatiotemporal spike train.

[0116]

[0128] Receptors 612 and 613 receive back-propagated signals from dendrites 601A and axon soma 601B, respectively. These are summed, for example, to synaptic elements, time-derived 614, and the resulting signal is multiplied with the rNMDA 607A signal to modify the amplifier 608 and thus the weights of each synaptic element 602.

[0117]

[0129] This increase in dimensionality allows for more states and transitions (and time constants) and provides more flexibility in the implementation of the interplay of homeostatic and metaplasticity, i.e., providing a mechanism for adaptation to environmental changes and thus providing a means to achieve and maintain robust neural computation.

[0118]

[0130] In engineering terms, homeostatic plasticity is a form of backpropagation signal control that balances the effects of drift in neuronal activity or internal connectivity as a result of changing external conditions or temperature fluctuations.

[0119]

[0131] This concept has an important role in the hardware design of spiking neural networks 110 because it provides robustness against variations in operating conditions, but the limited number of previous implementations include constant plasticity, primarily as a result of the technological constraints involved in realizing long time constants in silicon.

[0120]

[0132] The implemented synaptic element 602 has multiple computational capabilities, e.g., a wide range of receptor time constants and conductance values ​​g m , axonal and dendritic delays, allowing for the realization of optimal synaptic transfer function.

[0121]

[0133] Activity-dependent plasticity of synaptic transmission, which underlies learning and memory, is primarily due to postsynaptic changes in the biophysical properties of receptors.

[0122]

[0134] To regulate synaptic current flow, each receptor is specified as a multi-trans-conductance channel, modeling nonlinear properties such as multi-binding cooperativity of neurotransmitters to the receptor. Synaptic currents mediated by NMDA receptors 607A are modulated by synaptic activity. Fast modification of receptor numbers at synaptic elements 602 serves as a control mechanism for activity-dependent changes in synaptic efficacy through the formation, stability, morphology, and density of synaptic contacts. If a group (burst) of dendritic spikes is sufficient to cross a threshold, the axon 603 generates an action potential, and the subsequent spike is back-propagated into the dendrite 601A, multiplied with the cell body 604 signal, and added to the NMDA receptor 607A signal, which then generates a weight control. This quasi-local homeostatic framework provides a normalization function without interfering with Hebbian plasticity, effectively keeping net synaptic activation constant through stepwise adjustments of postsynaptic strength.

[0123]

[0135] Although the hardware implementations of both Figures 7A and 7B need not necessarily be used in a multi-core implementation of the present invention, any spiking neural network 110 may be implemented in any hardware in this manner.

[0124]

[0136] Each of the combined ensemble systems operates according to a biologically plausible mechanism, e.g., a mixed-expert nonlinear gating mechanism based on spike-timing-dependent plasticity (STDP), the combination of STDP and activity-dependent changes in neuronal excitability eliciting a Bayesian information process (called a spiking expectation-maximization (SEM) network).

[0125]

[0137] The proposed ensemble architecture can utilize such SEM networks as individual ensemble units (classifiers). The learning rule used in SEM is the weight-dependent Hebbian triplet TSTDP rule. A classifier is assumed to have a set of output neurons (one for each class), each of which fires events (spikes) according to its firing probability distribution. The classifier follows a probabilistic winner-takes-all (sWTA) mechanism, where only a single neuron 1 can fire for any presented input data.

[0126]

[0138] All synapses ij is the input y i We collect activation statistics of the (presynaptic neurons) at each synapse. These statistics can be collected at run-time from samples of the augmented input distribution. From this data, each weight can be interpreted as the log product of two local virtual counters at each synapse, i.e., one representing the number of events and one representing the local learning rate η ij Represents.

[0127]

[0139] As a result, spike-event-triggered probabilistic online learning rules can be derived.

number

[0128]

[0140] This is Synapse ij In the neuron z i Approximate the logarithm of the moving average of the output over the spiking time of

[0129]

[0141] The (random) variables used to set the probabilistic weight updates in the spiking neural network 110 are independent and identically distributed. Due to the stochasticity introduced in the transitions, the classifiers learn different representations of the target classes. The classifier responses are then combined into an aggregated improved classifier. Furthermore, by adding uncertainty to the bound values ​​of the synaptic weights, the level at which the classifier decisions are activated can be controlled.

[0130]

[0142] Neurons 1 communicate in the spiking neural network 110 primarily through fast all-or-none events, i.e., spikes in their membrane potential. The relative spike firing times in neuronal populations are assumed to be information codes, and synchronization between neuronal populations is assumed to be signals for encoding and decoding information. As a result, the neuroelectrical properties of excitatory and inhibitory neuronal networks can be expressed as:

number

[0131]

[0143] Here, Γ i (t) is the i-th synaptic drive at time t, which essentially indicates the synaptic strength, defined in terms of synaptic weights and gain functions, λ i is the gain that regulates the exponential decay of the synaptic voltage and mimics the spike-time-dependent scaling of the input conductance, and is the function f i (...) represents the firing rate of the i-th neuron, and w ji is a constant that specifies the connection between the jth neuron on the ith neuron, and v r,i (t) denotes the input voltage to the i-th neuron, e.g., a nerve impulse from a sensory receptor. The weight w ji For a connected synaptic element with i, the i-th neuron is the post-synaptic neuron and the j-th neuron is the pre-synaptic neuron. ican be considered as the total transfer function of all synaptic elements driving neuron i, and may be, for example, a function of the exponential decay values, weights, and gains of the synaptic elements driving neuron i, etc.

[0132]

[0144] However, neurons are noisy both in the generation of spikes and in the transmission of synaptic signals. Noise arises from the quantal release of neurotransmitters, the random opening of ion channels, the coupling of background neural activity, etc.

[0133]

[0145] Noise then induces neuronal variability, increases neuronal sensitivity to environmental stimuli, and affects interneuron synchronization, facilitating probabilistic inference.

[0134]

[0146] We derive the uncertainty model as a Markov process, where the stochastic integrals are interpreted as a system of Ito-type stochastic differential equations, thereby extending equation (1) above with time-varying delays and stochastic input uncertainties.

number

[0135]

[0147] The first two terms on the right hand side of equation (2) are the deterministic drift and stochastic diffusion parts of the stochastic differential equation, respectively, where

number

number

[0136]

[0148] where ω(t)=[ω1(t),ω2(t),...,ω n (t)] Tdescribes the noise in the input voltage and is represented by Brownian motion, i.e., the n-dimensional standard Wiener process, σ(Γ)=diag([σ1(Γ1),σ2(Γ2),...,σ n (Γ n )] T ) represents the state-dependent noise matrix for a Gaussian white noise process dω(t). The time-dependent function δ ji (t) denotes the continuous time-varying time delay of the signal from the jth neuron to the ith neuron at time t, and δ ji (t)≧0, t≧0.

[0137]

[0149] If all the entries of the mean dynamics matrices Ψ and w are, for example, (v r,i The assumption is made that the signals are synchronously perturbed (at t)=0.

[0138]

[0150] The assumption is that synaptic inputs resemble delta-like current pulses and can then be modeled as Gaussian white noise which is valid for fast AMPA currents (decays on the order of 2 ms) and GABAA currents (decays on the order of 5-10 ms), but for slow currents such as NMDA or GABAB currents (decays on the order of 50-100 ms), this assumption may not hold. As a result, they are modeled as colored noise, including the fact that non-zero synaptic time constants within the steady-state firing rate require complex 3-D Fokker-Planck expressions.

[0139]

[0151] While such a methodology is feasible, a more effective approach is to replace the variance by a rescaled variance, i.e., the synaptic time constant τ s It consists in noting that mainly affects the amplitude of the fluctuations. Therefore, from the ratio of the voltage variances of the colored noise and the white noise, we obtain the synaptic time constant dependent reduction factor

number

number

number

[0140]

[0152] The goal is to find the variance-covariance matrix of Γ(t). Equation 2 is a set of (Ito) stochastic differential equations. From there, we derive the variance-covariance matrix of Γ(t), which is essentially a system of linear ordinary differential equations with time-varying coefficients. To find the solution, we apply a discretization using the backward Euler method, which results in a continuous-time algebraic Lyapunov equation, shown below in Equation 5. These are solved by an iterative approach, and the solution is given in Equation 6.

[0141]

[0153] The goal is to find the synaptic drive Γ for every neuron i. i The aim is to find an equation for the variance of (t), which is i This is because it measures how far the random values ​​of (t) are spread from their mean value. The variance of the synaptic drive is the expected squared deviation of the synaptic drive from its mean. By finding such an expression, it is possible to bound spiking neural networks. That is, the spread of the values ​​of synaptic drive resulting from noise in a spiking neural network can be limited by tweaking the parameters that make up the expression for the variance so that it is smaller, and in particular so that the variance is minimized in a particular hardware implementation.

[0142]

[0154] Diagonal element K of the variance-covariance matrix K(t) of Γ(t) i,i (t) are the synaptic driving i (t). Therefore, by solving the variance-covariance matrix K(t) of Γ(t), we can obtain the synaptic drive Γ for all neurons i. i The equation for the variance of (t) is obtained.

[0143]

[0155] To solve (2), it is necessary to first find σ(Γ) and then take the square root of that matrix. The stationary statistics of the open channels in the Markov channel model are used to define the noise process in the conductance model. The general method for constructing σ(Γ) from M ion channels (or the number of synaptic elements in a spiking neural network) by the deterministic drift and Goldwin method can be expressed as follows:

number

[0144]

[0156] where N is the number of states and I N×1 is an N×1 column vector with all entries equal to 1, and I N×N is an N×N unit matrix. σ(Γ * )=0 * Using equation (4) above, the noise part of equation (2) is zero at * The equilibrium point Γ can be obtained. * Then, the spiking neural network is stable, i.e., noise does not affect the behavior of the spiking neural network. Then, the boundary of the network is such that the synaptic drive Γ for every neuron i i It is determined by finding the equation for the variance of (t).

[0145]

[0157] Applying Ito’s theorem to stochastic differentiation and then the backward Euler method, we obtain the variance-covariance matrix K(t) of Γ(t) (initial value K(0) = E[ΓΓ T]) can be expressed as a continuous-time algebraic Lyapunov equation.

number

[0146]

[0158] At a specific time t r K(t) in is computed by solving the system in (5). Small dense Lyapunov equations are computed efficiently by the Bartels-Stewart or Hammarling algorithms. Alternatively, large dense Lyapunov equations can be computed by sign function based techniques. In general, the matrices for deterministic drift in neural networks are not full rank matrices, and as a result, (5) is rewritten as a sparse linear matrix-vector system in the standard form and solved with an adjusted alternating direction algorithm.

number

[0147]

[0159] Here, K j has rank j×n for iterations j=1, 2, ..., where n is the sum of [σ(Γ)] r is the number of vectors in

[0148]

[0160] (2) The network is moved to the equilibrium point Γ where the stochastic perturbation decreases. * For example, the point σ(Γ * ) = 0, the transformation Θ(t) = Γ(t) - Γ * Using the equilibrium point Γ * Shift to the origin,

number

[0161] Here, g(Θ(t))=f(Γ(t)+Γ * )-f(Γ(t)), and the controller u(t) is

number

[0162] Where:

number

[0149]

[0163] Without loss of generality, Figures 8 and 9 illustrate the network activity and thus the average network activity histograms for an unbounded network and a bounded network, respectively.

[0150]

[0164] In both figures, the input signal to the spiking neural network 110 is the same. However, the synaptic drive Γ of every neuron i in the spiking neural network 110 i (t) is the equilibrium point Γ i * By bounding the network to lie within a certain range around , a finite neural network is less susceptible to noise than an infinite network. i * A certain range around the synaptic drive Γ i It is set by selecting the parameters that occur in the equation for the variance of (t). i is below a certain value that depends on the maximum tolerance for the amount of noise in the network 110. This value can be predefined and depends on the particular application, the input signal, the accuracy requirements of the classification performed by the spiking neural network 110, etc.

[0151]

[0165] In practice, the network 110 can be bounded, for example, by simulating the network in either hardware or software and then bounding the network as described above and evaluating the effect. Bounding of the network 110 can be performed during the learning process when the weights are set. Noise can be artificially added or simulated to test the stability of the network when affected by noise.

[0152]

[0166] The above can be implemented in any spiking neural network 110, either in a multi-core solution in one particular core 210 in the core array 200, or in a general-purpose hardware implementation of the spiking neural network 110.

[0153]

[0167] The present invention can be realized as an implementation of a reconfigurable multi-layer computing network as an ensemble of multi-receptor, multi-dendrite neurosynaptic structures, where each ensemble realizes a unique input-output transfer function, e.g., filtering, amplification, multiplication, summation.

[0154]

[0168] In one embodiment, a method is devised for setting confidence and uncertainty bounds on a neurosynaptic structure or an ensemble of such structures for control and adaptability of classifier responses and activation functions.

[0155]

[0169] In one embodiment, an increased dimensionality of the neurosynaptic structure is realized, allowing for more states and transitions, providing more flexibility in the implementation of plasticity and metaplasticity interactions, as well as various neuronal properties, such as delays or synaptic transfer functions.

[0156]

[0170] In one embodiment, multiple learning rules may be implemented where each neurosynaptic structure in an ensemble or set of ensembles implements a unique learning rule.

[0157]

[0171] In one embodiment, unique input / output transfer functions can be generated by adaptive control of charge or amounts of representative chemicals, such as calcium, potassium, and sodium.

[0158]

[0172] In one embodiment, fine-grained temporal precision can be achieved to drive precise learning behavior and improve the learning capacity of neurosynaptic structures.

[0159]

[0173] In one embodiment, multiple temporal control mechanisms can be implemented to allow for the implementation of homeostatic regulation in the resulting network.

[0160]

[0174] In one embodiment, multiple signals can be implemented to model local and global post-synaptic influences.

[0161]

[0175] In one embodiment, the set of ensembles can be organized as a reconfigurable neurosynaptic array.

[0162]

[0176] In one embodiment, the reconfigurable neurosynaptic structure can be organized as computational / signal processing cores, where each core can be organized as a single heterogeneous or homogeneous type implementing a particular learning rule, weight storage, etc. depending on the optimal performance-power-area tradeoff.

[0163]

[0177] The large discrepancy in energy efficiency and cognitive performance of biological neural systems and traditional computing is largely illustrated with tasks related to real-time interaction with physical surroundings, especially when uncontrolled or noisy sensory inputs are present. However, neuromorphic event-based neuronal networks are a unique choice for small, low-power cognitive systems that learn and adapt to changes in the statistics of complex sensory signals due to their ability to learn by example, parallelism of operations, associative memory, multifactorial optimization, and scalability. This new hardware-resilient methodology for neuromorphic networks enables designs that provide energy-efficient solutions for applications ranging from detecting patterns in biomedical signals (e.g., spike classification, seizure detection, etc.), speech commands, and classification of images (e.g., handwritten digits), and can be applied in a wide range of devices, including smart sensors or wearable devices in cyber-physical systems and the Internet of Things.

[0164]

[0178] 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 the 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-writeable 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, a ROM chip, or any type of solid-state non-volatile semiconductor memory readable by a CD-ROM drive), and (ii) writeable 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.

[0165]

[0179] Two or more of the above embodiments may be combined in any suitable manner.

Claims

1. A spiking neural network for classifying an input signal, comprising a plurality of spiking neurons and a plurality of synaptic elements interconnecting said spiking neurons to form said network; each synaptic element is adapted to receive a synaptic input signal and 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 synaptic element; 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; The network is divided into a plurality of sub-networks; each sub-network comprising a subset of the spiking neurons connected to receive synaptic output signals from a subset of the synaptic elements; the sub-network is adapted to generate a sub-network output pattern signal in response to a sub-network input pattern signal applied to the sub-network; each sub-network forms part of one or more cores in an array of cores, each core consisting of a programmable network of spiking neurons implemented in hardware or a combination of hardware and software; A spiking neural network, wherein communication between cores in the core array is arranged via a programmable interconnect structure.

2. The spiking neural network of claim 1 , wherein the programmable interconnect structure comprises a switchable matrix.

3. 3. The spiking neural network of claim 2, wherein the switchable matrix comprises a plurality of routers adapted to route a plurality of output signals from a first core in the core array to a plurality of inputs of a second core in the core array.

4. 4. A spiking neural network as claimed in claim 1, wherein the programmable interconnect structure forms a packet-switched network between the cores in the core array.

5. 5. A spiking neural network as claimed in claim 1, wherein the programmable interconnect structure uses address-event representation.

6. 6. A spiking neural network as claimed in any preceding claim, wherein the programmable interconnect structure uses synchronous or asynchronous communication.

7. 7. A spiking neural network as claimed in any preceding claim, wherein the core array is implemented in an integrated circuit and the programmable interconnect structure comprises a network-on-chip.

8. 8. The spiking neural network of claim 7, wherein the network-on-chip is configurable in real-time.

9. 9. A spiking neural network as claimed in any one of claims 1 to 8, wherein any of the learning rules, weight storage mechanisms, or communication protocols for the synaptic elements are applied non-uniformly within a single core of the core array.

10. 10. The spiking neural network of claim 1, wherein any of the learning rules, weight storage mechanisms, or communication protocols for the synaptic elements are applied non-uniformly across multiple cores in the core array.

11. 11. A spiking neural network as claimed in any preceding claim, wherein a learning rule configuration of one or more of the cores in the core array is implemented using a learning rule block.

12. 12. The spiking neural network of claim 11, wherein the cores that use the same learning rule use a common learning rule block that implements the learning rule.

13. 13. A spiking neural network as claimed in any preceding claim, wherein a distribution of learning rules within and / or across each core in the array of cores is dynamically configurable at run-time.

14. 14. A spiking neural network according to any one of claims 1 to 13, wherein one of the sub-networks is part of a classifier for classifying an input signal of the sub-network.

15. 15. A spiking neural network according to any one of claims 1 to 14, wherein one of the sub-networks is part of an ensemble of classifiers for classifying input signals to that sub-network.

16. 16. A spiking neural network as claimed in any preceding claim, in which the weights of the synaptic elements are constructed using probabilistic weight updates.

17. 17. A spiking neural network according to claim 1, wherein the weights of the synaptic elements are bounded by boundary values, the boundary values ​​being probability values.

18. An integrated circuit comprising the spiking neural network implemented in the core array of any one of claims 1 to 17.

19. 1. A method of partitioning and mapping a spiking neural network onto a core array, the spiking neural network comprising a plurality of spiking neurons and a plurality of synaptic elements interconnecting the spiking neurons to form the network; each synaptic element is adapted to receive a synaptic input signal and 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 synaptic element; 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; the core array comprises a plurality of cores; Each core consists of a programmable network of spiking neurons implemented in hardware or a combination of hardware and software; Communications between cores in the core array are arranged via a programmable interconnect structure, the method comprising: partitioning the network into a plurality of sub-networks, each sub-network comprising a subset of spiking neurons connected to receive synaptic output signals from a subset of the synaptic elements, the sub-network adapted to generate sub-network output pattern signals from the subset of spiking neurons in response to sub-network input pattern signals applied to the subset of synaptic elements; mapping each sub-network onto said one or more cores; A method comprising:

20. the partitioning of the spiking neural network into the sub-networks implemented on the one or more cores is determined by a mapping method; The mapping method comprises constraint-driven partitioning; The method of claim 19 , wherein the constraint is a performance metric linked to a function of each respective sub-network.

21. A spiking neural network, a plurality of spiking neurons implemented in hardware or a combination of hardware and software; and a plurality of synaptic elements interconnecting said spiking neurons to form said network. Equipped with each synaptic element is adapted to receive a synaptic input signal and 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 synaptic element; 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 spiking neural network, wherein learning rules, weight storage mechanisms, and / or communication protocols are applied non-uniformly across the neurons and / or synaptic elements within the spiking neural network.

22. A spiking neural network comprising a spiking neuron and a synapse element connected to the spiking neuron, the synaptic element comprises first and second receptors adapted to receive a synaptic input signal; the first and second receptors respectively generating first and second receptor signals based on the synaptic input signal; the synaptic element applies a weight to the first receptor signal to generate a synaptic output signal; the synaptic element is configurable to adjust the weight applied by the synaptic element based at least on the second receptor signal; the spiking neurons are adapted to receive the synaptic output signals from the synaptic elements and to generate spatiotemporal spike train output signals in response at least to the received synaptic output signals.

23. 23. The spiking neural network of claim 22, wherein the neuron emits a control signal which, together with the second receptor signal, modulates the weight applied by the synaptic element.

24. 24. The spiking neural network of claim 23, wherein the control signal is a backpropagation signal.

25. 25. A spiking neural network as claimed in claim 23 or 24, wherein the neurons comprise dendrites, axons and cell bodies, and the control signals arise from the dendrites and / or axons and / or cell bodies of the neurons.

26. 26. A spiking neural network as claimed in any one of claims 23 to 25, wherein the control signal comprises one or more spikes generated by action potentials of the neurons.

27. 27. A spiking neural network as described in any one of claims 22 to 26, wherein a decay time of the first receptor is faster than a decay time of the second receptor.

28. 28. A spiking neural network as claimed in any one of claims 22 to 27, wherein the first receptor generates a source or sink current for the spiking neuron.

29. 29. A spiking neural network as claimed in any one of claims 22 to 28, wherein the first receptor comprises a low pass filter.

30. 30. A spiking neural network as claimed in any one of claims 22 to 29, wherein the second receptor forms a voltage-dependent receptor.

31. 31. A spiking neural network as claimed in any one of claims 22 to 30, wherein the second receptor comprises a low-pass filter, a band-pass filter, a high-pass filter, and / or an amplifier.

32. 32. A spiking neural network as claimed in any one of claims 22 to 31, wherein the first receptor is an AMPA receptor.

33. 32. A spiking neural network as claimed in any one of claims 22 to 31, wherein the first receptor is a GABA receptor.

34. 34. A spiking neural network as claimed in any one of claims 22 to 33, wherein the second receptor is an NMDA receptor.

35. 1. A method for adjusting weights of synaptic elements in a spiking neural network, the spiking neural network comprising spiking neurons connected to the synaptic elements, the method comprising: the synaptic element comprises first and second receptors adapted to receive a synaptic input signal; the first and second receptors receive the synaptic input signal and generate first and second receptor signals, respectively, based on the synaptic input signal; the synaptic element applies a weight to the first receptor signal to generate a synaptic output signal; The weights of the synaptic elements are adjusted based on at least the second receptor signal; The spiking neuron receives the synaptic output signal from the synaptic element and generates a spatiotemporal spike train output signal in response at least to the received synaptic output signal.

36. 36. The method of claim 35, wherein the neuron emits a control signal that, together with the second receptor signal, modulates the weight applied by the synaptic element.

37. 37. The method of claim 36, wherein the control signal is an error backpropagation signal.

38. 38. The method of claim 36 or 37, wherein the neuron comprises dendrites, an axon and a cell body, and the control signal arises from the dendrites and / or axon and / or cell body of the neuron.

39. 39. The method of any one of claims 36 to 38, wherein the control signal comprises one or more spikes generated by action potentials of the neuron.

40. 40. The method of any one of claims 35 to 39, wherein the decay time of the first receptor is faster than the decay time of the second receptor.

41. 41. The method of any one of claims 35 to 40, wherein the first receptor generates a source or sink current for the spiking neuron.

42. A method according to any one of claims 35 to 41, wherein the first receptor comprises a low pass filter.

43. The method of any one of claims 35 to 42, wherein the second receptor forms a voltage-dependent receptor.

44. 44. A spiking neural network as claimed in any one of claims 35 to 43, wherein the second receptor comprises a low-pass filter, a band-pass filter, a high-pass filter, and / or an amplifier.

45. 45. The method of any one of claims 35 to 44, wherein the first receptor is an AMPA receptor.

46. 45. The method of any one of claims 35 to 44, wherein the first receptor is a GABA receptor.

47. 47. The method of any one of claims 35 to 46, wherein the second receptor is an NMDA receptor.

48. An integrated circuit comprising a spiking neural network according to any one of claims 22 to 34.

49. 1. A method for configuring a spiking neural network to reduce the effects of noise in the network, comprising: the spiking neural network comprises a plurality of spiking neurons implemented in hardware or a combination of hardware and software, and a plurality of synaptic elements interconnecting the spiking neurons to form the network; each synaptic element is adapted to receive a synaptic input signal and 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 synaptic element; 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; The synaptic drive Γ of one of the spiking neurons i in the spiking neural network i is a time-dependent function describing the total transfer function of all synaptic elements connected within the neuron, and the method further comprises: The synapse-driven Γ i obtaining an equation for the variance of each synaptic drive of the synapse, the equation for the variance being dependent on a control parameter; The synapse-driven Γ i the variance of each synaptic drive of is adjustable by adjusting the control parameter; The control parameter is the synaptic drive Γ i is adjusted to be below a predetermined value, so that the synaptic drive Γ i is the synaptic drive Γ where noise effects are minimal. i Equilibrium point Γ i * A method of being bounded around.

50. The synaptic drive Γ of the spiking neuron i in the spiking neural network i Γ = (Γ 1 , . . . , Γ n ) T where n is the number of neurons in the spiking neural network and Γ satisfies: dΓ=N(Γ(t))dt+σ(Γ(t))dω(t), where N(Γ(t)) is the nominal matrix part, σ(Γ) is the state-dependent noise matrix for a Gaussian white noise process dω(t), and ω is a Wiener process describing the noise in the spiking neural network. The equation for dΓ can be rewritten in terms of the variance-covariance matrix K(t) of Γ(t) that satisfies the following continuous-time algebraic Lyapunov equation: N(Γt)K(t)+K(t)[N(t)] T +σ(Γ(τ))[σ(Γ(τ))] T =0、 The synapse-driven Γ i 50. The method of claim 49, wherein obtaining an expression for the variance of each synaptic drive of comprises determining diagonal values ​​of the variance-covariance matrix K(t).

51. The synapse-driven Γ i The equilibrium point Γ i * is expressed in vector form for all neurons in the spiking neural network, and Γ * =(Γ 1 * , . . . , Γ n * ) T where the noise matrix σ is zero, i.e., σ(Γ * 51. The method of claim 50, wherein: .times. ...

52. A method according to any one of claims 49 to 51, wherein the control parameters are adjusted by computer simulation.

53. A spiking neural network, a plurality of spiking neurons implemented in hardware or a combination of hardware and software; and a plurality of synaptic elements interconnecting said spiking neurons to form said network. Equipped with each synaptic element is adapted to receive a synaptic input signal and 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 synaptic element; 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; The synaptic drive Γ of one of the spiking neurons i in the spiking neural network i is a time-dependent function describing the total transfer function of all synaptic elements connected within the neuron, The synapse-driven Γ i The variance of each synaptic drive of the neuron i in the spiking neural network is the synaptic drive Γ i However, the synaptic drive Γ i Equilibrium point Γ i * A spiking neural network below a given value, bounded around

Citation Information

Patent Citations

  • cold neuron spike timing backpropagation

    JP2017515205A

  • Method and system for reward-modulated spike timing-dependent plasticity

    JP5512895B2

  • Neural processor based accelerator system and method

    US20170024644A1

  • Synapse and neuromorphic device including the same

    US20170061278A1