Systems and methods for reconfigurable modular neurosynaptic computing structures

Abstract

The invention discloses a nerve synapse structure for a spiking neural network, and the nerve synapse structure comprises one or more input ports; and one or more synaptic elements, each synaptic element connected to at least one input port and configured to receive an input signal and output a weighted post-synaptic signal. In addition, the neurosynaptic structure includes neurons connected to the one or more synaptic elements. The neurosynaptic structures are provided with different feedback and control structures, such as AMPA, GABA and NMDA receptors, axon, dendritic and neuronal back propagation channels, and / or astrocyte structures.

Description

Technical Field

[0001] The present disclosure relates generally to automatic signal recognition techniques, and more particularly to systems and methods for reconfigurable modular neurosynaptic computing structures that allow for increased flexibility and efficiency of signal processing units and for achieving complex spatiotemporal and plastic behaviors in spiking neuronal networks. Background Art

[0002] To minimize the energy per inference or specific task (e.g., systems known as energy-proportional systems or always-optimal circuits, i.e., systems with some form of dynamic feedback control or adaptivity), co-design across the computing stack is required (so that the algorithms and applications can influence the underlying hardware design, and likewise, the underlying hardware implementation can be tailored to the needs or constraints of the specific application), greater hardware-software co-optimization (to minimize network size for a given accuracy / maximize the number of operations within the allowed power envelope), and fully modular and flexible hardware systems to unify the design process.

[0003] Smart sensor systems are not only good at modifying or adjusting their internal processing or capabilities, but are also able to modify or alert or refocus / implement expectations within the system (based on new information). Therefore, the reasoning engine associated with such smart systems continuously updates its knowledge of internal and external parameter sets. To save space, time and energy, new information is stored at the synaptic site, where it is processed and from where it can be recalled efficiently.

[0004] For short-term memory, short-term changes in input are relayed through low-level synapses. High-level synapses encode information (quantity) after multiple processing stages, so by strengthening synaptic capacity and allocating additional capacity, their memory is longer and the encoding is more stable.

[0005] Reconfigurable neuromorphic networks are typically composed of circuits that only partially contain receptors, dendrites, and subsequent synaptic properties. A typical synaptic array consists of a neural network matrix that connects n×n (or some portion) of programmable synapses to n neurons. However, the detailed description of the synaptic computational structure that stipulates and promotes complex spatiotemporal pulse behavior and consolidates plasticity includes both functional modifications and structural mechanisms (circuit reconstruction through synapse generation, i.e., formation, elimination, remodeling, and degradation) of synaptic elements. These include a variety of receptors and dendritic channels (for example, see I. Segev, M. London, "Untangling dendrites with quantitative models", Science, vol. 290, pp. 744-750, 2000), which in addition to changing various forms of short-term and long-term plasticity, intrinsic excitatory plasticity, multi-receptor plasticity, and non-Hebbian plasticity (including steady-state synaptic scaling and metaplasticity and rapid structural (synaptic and dendritic) plasticity), also change synaptic responses by amplification, regulation, dendritic structural scaling, etc. Therefore, general synaptic structures (as simple point processing units) cannot capture the diverse temporal dynamics of different types of receptors in biological synapses, which is crucial for achieving biophysically accurate neural behavior in spiking neural networks (SNNs). See, for example, A. Morrison, M. Diesmann, and W. Gerstner, "Phenomenological models of synaptic plasticity based on spike timing," Biolog. Cybern., vol. 98, no. 6, pp. 459-478, 2008. Summary of the invention

[0006] The present invention relates to the implementation of a hierarchical modular hardware platform that promotes heterogeneity of computing elements and systems, but also enables targeted reconfigurability, full programmability, and parameter adaptability. The implementation of this modularity and heterogeneity means that systems or system definitions can be reused, and the various parts of the system can be adjusted / adjusted accordingly. Through partitioning, complexity, and reconfigurability, neural synaptic resources are allocated to the signals they need to process; reliable signals require multiple synapses to maintain the signal-to-noise ratio, however, when the signal-to-noise ratio is low, fewer synapses can be used, thereby avoiding excess capacity waste. In addition, in order to maximize efficiency, signals are concentrated / sparse in space and time (i.e., information is concentrated at each hierarchical scale); therefore, highly active computing elements (e.g., neurons, synapses) are spatially concentrated in specific areas, while their activities are concentrated in time. Signal processing is performed in the continuous time (analog) domain to maximize information rate. In addition, the hierarchical modular approach enables plasticity mechanisms to be implemented at different time scales (e.g., short-term, long-term, steady-state, structural plasticity), and subsequently across different integrated verticals. Thus, compensation can be performed at different levels of time and space and across different framework modularities, along different computing stacks, at the circuit level, the system level, and the architectural level.

[0007] The modular framework promotes the advantages of (biological / impulse) signal processing at different time-granularities or hierarchies and enables full software / hardware co-alignment in terms of distributed processing, neural network definition and mapping.

[0008] According to a first aspect, a neural synaptic structure for a pulse neural network is disclosed, wherein the neural synaptic structure includes a plurality of synaptic elements and neurons. Each synaptic element may include: one or more input ports, each input port is configured to receive an input signal; an output port, configured to output a postsynaptic signal; and a signal path connecting one or more input ports to the output port, the signal path including a weight element, which is configured to apply a weight for generating a postsynaptic signal. The signal path of at least a portion of the synaptic elements may include an AMPA receptor, which is configured to integrate the input signal to make the postsynaptic signal more excited. The signal path of at least a portion of the synaptic elements may include a GABBA receptor, which is configured to integrate the input signal to make the postsynaptic signal more inhibitory. The neuron may include a soma and an axon, wherein the soma may be configured to receive and combine one or more postsynaptic signals, and the axon may be configured to generate a pulse output when the membrane potential of the neuron reaches a predetermined threshold in response to the output of the soma.

[0009] According to an embodiment of the first aspect, at least one synaptic element may include a signaling path including both an AMPA receptor and a GABBA receptor.At least one synaptic element may be configured to combine outputs from the AMPA receptor and the GABBA receptor for generating a postsynaptic signal.

[0010] According to an embodiment of the first aspect, the AMPA receptor and / or GABBA receptor of the neural synapse element may be configurable to be active or passive at a particular moment, and / or the AMPA receptor and / or GABBA receptor of the neural synapse element may be configurable to select a frequency range of input signals in which the AMPA receptor and / or GABBA receptor can operate.

[0011] According to an embodiment of the first aspect, the weighting element may be located after the AMPA receptor and / or the GABBA receptor along the signaling pathway.

[0012] According to an embodiment of the first aspect, the weight elements may be dynamically adjustable, preferably by modulating the weight elements.

[0013] According to an embodiment of the first aspect, the AMPA receptor comprises a voltage-driven integrator configured to generate an excitatory postsynaptic current for inducing charge inflow from the neuronal membrane potential, and / or the GABBA receptor is a voltage-driven integrator configured to generate an inhibitory postsynaptic current for inducing charge outflow from the neuronal membrane potential. The AMPA and / or GABBA receptors can be formed using one or more gm-C filters. The AMPA receptor can be configured to mediate glutamatergic synaptic currents to drive the cell body of the neuron, and / or the GABBA receptor can drive a rapid Hebbian weakening of the synaptic element. The GABBA receptor preferably can have a temporal dynamic behavior comparable to that of the AMPA receptor.

[0014] According to a second aspect, a neural synaptic structure for a spiking neural network is disclosed. The neural synaptic structure includes a plurality of synaptic elements and neurons. Each synaptic element includes: one or more input ports, each input port is configured to receive an input signal; an output port, configured to output a postsynaptic signal; and a signal path connecting the one or more input ports to the output port, the signal path including a weight element, the weight element being configured to: apply a weight for generating a postsynaptic signal. The neuron may include a soma and an axon, wherein the soma is configured to: receive and combine one or more postsynaptic signals, and the axon is configured to: generate a pulse output when the membrane potential of the neuron reaches a predetermined threshold in response to the output of the soma. The neural synaptic structure may also include an axon backpropagation signal path, which is connected to receive an axon signal from an axon of the neuron. The axon backpropagation signal path may include an axon backpropagation receptor, which is configured to integrate the axon signal. A first subset of synaptic elements may be configured to: receive the integrated axon signal, and based on the integrated axon signal, modulate the weight element of the first subset of synaptic elements.

[0015] According to an embodiment of the second aspect, each synaptic element in the first subset of synaptic elements may include an NMDA receptor configured to integrate the input signal to form an integrated NMDA signal. The synaptic elements in the first subset of synaptic elements may be configured to combine the integrated NMDA signal and the integrated axon signal, and modulate the weight element of the synaptic element based on the combined signal.

[0016] According to an embodiment of the second aspect, the integrated axon signal may be time differentiated by a derivation unit (preferably a differentiator), and then the differentiated integrated axon signal may be used to modulate the weight element.

[0017] According to an embodiment of the second aspect, the axon backpropagating receptor may be a voltage driven pure integrator, preferably, the axon backpropagating receptor is formed using one or more gm-C filters.

[0018] According to an embodiment of the second aspect, at least two axonal backpropagation receptors may be used to backpropagate to at least two groups of synaptic elements.

[0019] According to an embodiment of the second aspect, the neural synapse structure may also be the neural synapse structure according to the first aspect.

[0020] According to a third aspect, a neural synaptic structure for a spiking neural network is disclosed. The neural synaptic structure may include a plurality of synaptic elements and neurons. Each synaptic element includes: one or more input ports, each input port is configured to receive an input signal; an output port, configured to output a postsynaptic signal; and a signal path connecting the one or more input ports to the output port, the signal path including a weight element, the weight element being configured to: apply a weight for generating a postsynaptic signal. The neuron may include a soma and an axon, wherein the soma may be configured to: receive and combine one or more postsynaptic signals, and the axon may be configured to: generate a pulse output when the membrane potential of the neuron reaches a predetermined threshold in response to the output of the soma. The neural synaptic structure may also include a dendritic backpropagation signal path, which is configured to: receive and combine postsynaptic signals from one or more synaptic elements. The dendritic backpropagation signal path may include a dendritic receptor, which is configured to: integrate the combined postsynaptic signal. A second subset of synaptic elements may be configured to: receive the integrated combined postsynaptic signal, and modulate the weight elements of the second subset of synaptic elements based on the integrated combined postsynaptic signal.

[0021] According to an embodiment of the third aspect, each synaptic element in the second subset of synaptic elements may include an NMDA receptor configured to: integrate the input signal to form an integrated NMDA signal. The synaptic elements in the second subset of synaptic elements may be configured to: combine the integrated NMDA signal and the integrated combined post-synaptic signal, and modulate the weight element of the synaptic element based on the combined signal.

[0022] According to an embodiment of the third aspect, the integrated combined post-synaptic signal can be time differentiated by a derivation unit (preferably a differentiator), and then the differentiated integrated summed post-synaptic signal is used to modulate the weight element.

[0023] According to an embodiment of the third aspect, the dendritic receptors may be voltage driven pure integrators, preferably, the dendritic receptors may be formed using one or more gm-C filters.

[0024] According to an embodiment of the third aspect, the neuron may be configured to receive one or more postsynaptic signals from one or more synaptic elements at the cell body via a second dendritic receptor, the second dendritic receptor integrating the combined postsynaptic signals of each of the one or more synaptic elements. Preferably, the dendritic receptor is a voltage-driven pure integrator, more preferably, the dendritic receptor is formed using one or more gm-C filters.

[0025] According to an embodiment of the third aspect, one or more postsynaptic signals can be combined at the cell body by combining the combined (preferably also integrated) postsynaptic signals of each synaptic element in the second subset with one or more postsynaptic signals from one or more synaptic elements not in the second subset. Preferably, the one or more synaptic elements not in the second subset are also connected to the neuron via one or more different dendritic structures.

[0026] According to an embodiment of the third aspect, the neural synapse structure is also the neural synapse structure of the first and / or second aspect.

[0027] According to a fourth aspect, a neural synaptic structure for a spiking neural network is disclosed. The neural synaptic structure may include a plurality of synaptic elements and neurons. Each synaptic element may include: one or more input ports, each input port being configured to receive an input signal; an output port being configured to output a postsynaptic signal; and a signal path connecting one or more input ports to the output port, the signal path including a weight element, the weight element being configured to: apply a weight for generating a postsynaptic signal. The neuron may include a soma and an axon, wherein the soma is configured to: receive and combine one or more postsynaptic signals, and the axon is configured to: generate a pulse output when the membrane potential of the neuron reaches a predetermined threshold in response to the output of the soma. The neuron may include a neuron backpropagation signal path, which may be connected from the axon of the neuron to the soma of the neuron, and may include a local receptor configured to integrate the axon signal and a calcium element configured to set the influence of the integrated axon signal, thereby generating a neuron feedback signal. The neuron backpropagation signal path may be configured to: feed the neuron feedback signal to the soma of the neuron, wherein the neuron feedback signal may be combined with one or more postsynaptic signals.

[0028] According to an embodiment of the fourth aspect, the calcium element may be an amplifier, preferably a faster amplifier than the weight element of each synaptic element. Additionally or alternatively, the calcium element may provide adaptive feedback.

[0029] According to an embodiment of the fourth aspect, the local receptor may be a voltage driven pure integrator.Preferably, the local receptor is formed using one or more gm-C filters.

[0030] According to an embodiment of the fourth aspect, the neural synapse structure may also be a neural synapse structure according to the first, second and / or third aspect.

[0031] According to a fifth aspect, a neural synaptic structure for a spiking neural network is disclosed. The neural synaptic structure may include a plurality of synaptic elements and neurons. Each synaptic element may include: one or more input ports, each input port being configured to receive an input signal; an output port being configured to output a postsynaptic signal; and a signal path connecting one or more input ports to the output port, the signal path including a weight element, the weight element being configured to: apply a weight for generating a postsynaptic signal. The neuron may include a soma and an axon, wherein the soma is configured to: receive and combine one or more postsynaptic signals, and the axon is configured to: generate a pulse output when the membrane potential of the neuron reaches a predetermined threshold in response to the output of the soma. The neural synaptic structure may include an astrocyte structure connected to a third subset of the synaptic elements. The astrocyte structure may be configured to: receive and sum the postsynaptic signals of each synaptic element, and may include a modulation circuit, the modulation circuit being configured to: modulate the summed postsynaptic signal into a modulation signal. The modulation signal may be used to change the behavior at one or more synaptic elements of the third subset or at the neuron.

[0032] According to an embodiment of the fifth aspect, the astrocyte structure may include an mGluR receptor configured to integrate the modulation signal. The astrocyte structure may be configured to output the integrated modulation signal to each synaptic element in the third subset. Each of the plurality of synaptic elements may be configured to combine the integrated modulation signal with the integrated input signal. The integrated input signal may be obtained by integrating the input signal using AMPA and / or GABBA receptors positioned along a signal path configured to integrate the input signal.

[0033] According to an embodiment of the fifth aspect, the mGluR receptor may be a voltage driven pure integrator.Preferably, the mGluR receptor is formed using one or more gm-C filters.

[0034] According to an embodiment of the fifth aspect, the astrocyte structure may include an NMDA receptor configured to integrate the modulation signal. The astrocyte structure may be configured to output the integrated modulation signal to the neuron.

[0035] According to an embodiment of the fifth aspect, a neuron may include a neuron back propagation signal path connecting an axon of the neuron to a cell body, which may include a calcium element, and the calcium element may be configured to: set the influence of the integrated axon signal to generate a neuron feedback signal. The neuron back propagation signal path may be configured to: feed the neuron feedback signal to the cell body of the neuron. The integral modulation signal may modulate the calcium element of the neuron to change the influence of the integrated axon signal of the neuron.

[0036] According to an embodiment of the fifth aspect, the modulation circuit may include an IP3 receptor configured to integrate the summed postsynaptic signal and an astrocyte calcium element configured to set the influence of the modulation signal. The IP3 receptor and the astrocyte calcium element may generate the modulation signal.

[0037] According to an embodiment of the fifth aspect, the IP3 receptor can be a voltage driven pure integrator. Preferably, the IP3 receptor is formed using one or more gm-C filters. Additionally or alternatively, the astrocyte calcium element can be an amplifier, preferably an amplifier faster than the weight element of each synaptic element. Additionally or alternatively, the astrocyte calcium element can provide adaptive feedback.

[0038] According to an embodiment of the fifth aspect, the neural synaptic structure may include a synaptic element that is not directly connected to a neuron. Astrocyte calcium elements may be modulated using presynaptic neuronal signals or postsynaptic signals from synaptic elements that are not directly connected to a neuron, and these postsynaptic signals may be integrated using dendritic receptors.

[0039] According to an embodiment of the fifth aspect, the axon of the neuron can be connected to the astrocyte structure, so that the axon back-propagating receptor of the astrocyte can be configured to integrate the axon signal of the neuron. The integrated axon signal can be used to modulate the astrocyte calcium element.

[0040] Preferably, the integrated axon signal can be combined with other axonal back propagation signals of other neurons contained in the neural synaptic structure to form a combined axon signal. The combined axon signal can be used to regulate the astrocyte calcium element. More preferably, the combined axon input signal is time-derived by a derivation unit, and even more preferably, wherein the derivation unit is a differentiator.

[0041] According to an embodiment of the fifth aspect, the neural synapse structure may also be a neural synapse structure according to the first, second, third and / or fourth aspect.

[0042] According to a sixth aspect, a method for controlling a neural synaptic structure of a pulse neural network is disclosed, wherein the neural synaptic structure includes a plurality of synaptic elements and neurons. Each synaptic element may include: one or more input ports, each input port being configured to receive an input signal; an output port, configured to output a postsynaptic signal; and a signal path connecting one or more input ports to the output port, the signal path including a weight element, the weight element being configured to: apply a weight for generating a postsynaptic signal. The signal path of at least a portion of the synaptic element may include a GABBA receptor, which is configured to: integrate the input signal to make the postsynaptic signal more inhibitory. The neuron may include a cell body and an axon, wherein the cell body may be configured to: receive and combine one or more postsynaptic signals, and the axon may be configured to: generate a pulse output when the membrane potential of the neuron reaches a predetermined threshold in response to the output of the cell body. The method may start by: receiving an input signal at a synaptic element. Thereafter, the input signal may be integrated to make the postsynaptic signal more excitable and / or inhibitory.

[0043] According to a seventh aspect, a method for controlling a neural synaptic structure of a pulse neural network is disclosed. The neural synaptic structure includes a plurality of synaptic elements and neurons. Each synaptic element includes: one or more input ports, each input port is configured to receive an input signal; an output port, configured to output a postsynaptic signal; and a signal path connecting one or more input ports to the output port, the signal path including a weight element, the weight element being configured to: apply a weight for generating a postsynaptic signal. The neuron may include a soma and an axon, wherein the soma is configured to: receive and combine one or more postsynaptic signals, and the axon is configured to: generate a pulse output when the membrane potential of the neuron reaches a predetermined threshold in response to the output of the soma. The method may include: receiving an axon signal from an axon of the neuron. Subsequently, the received axon signal may be integrated. The integrated axon signal may be received at a first subset of the synaptic elements. Next, the weight elements of the synaptic elements of the first subset may be modulated based on the integrated axon signal.

[0044] According to an eighth aspect, a method for controlling a neural synaptic structure of a spiking neural network is disclosed. The neural synaptic structure may include a plurality of synaptic elements and neurons. Each synaptic element includes: one or more input ports, each input port is configured to receive an input signal; an output port, configured to output a postsynaptic signal; and a signal path connecting one or more input ports to the output port, the signal path including a weight element, the weight element being configured to: apply a weight for generating a postsynaptic signal. The neuron may include a soma and an axon, wherein the soma may be configured to: receive and combine one or more postsynaptic signals, and the axon may be configured to: generate a pulse output when the membrane potential of the neuron reaches a predetermined threshold in response to the output of the soma. The method may include: receiving and combining postsynaptic signals from one or more synaptic elements. Next, the combined postsynaptic signal may be integrated. The integrated combined postsynaptic signal may be received at a second subset of the synaptic elements, and the synaptic elements of the second subset may modulate the weight elements of the synaptic elements of the second subset based on the integrated combined postsynaptic signal.

[0045] According to the ninth aspect, a method for controlling a neural synaptic structure of a pulse neural network is disclosed. The neural synaptic structure may include a plurality of synaptic elements and neurons. Each synaptic element may include: one or more input ports, each input port being configured to receive an input signal; an output port being configured to output a postsynaptic signal; and a signal path connecting one or more input ports to the output port, the signal path including a weight element, the weight element being configured to: apply a weight for generating a postsynaptic signal. The neuron may include a soma and an axon, wherein the soma is configured to: receive and combine one or more postsynaptic signals, and the axon is configured to: generate a pulse output when the membrane potential of the neuron reaches a predetermined threshold in response to the output of the soma. The method may include: integrating the signal of the axon. In addition, the influence of the integrated axon signal may be set to generate a neuron feedback signal. Next, the neuron feedback signal may be fed to the soma of the neuron, where the neuron feedback signal may be combined with one or more postsynaptic signals.

[0046] According to the tenth aspect, a method for controlling a neural synaptic structure of a pulse neural network is disclosed. The neural synaptic structure may include a plurality of synaptic elements and neurons. Each synaptic element may include: one or more input ports, each input port being configured to receive an input signal; an output port being configured to output a postsynaptic signal; and a signal path connecting one or more input ports to the output port, the signal path including a weight element, the weight element being configured to: apply a weight for generating a postsynaptic signal. The neuron may include a soma and an axon, wherein the soma is configured to: receive and combine one or more postsynaptic signals, and the axon is configured to: generate a pulse output when the membrane potential of the neuron reaches a predetermined threshold in response to the output of the soma. The method may include: receiving and summing the postsynaptic signals of each synaptic element. Next, the summed postsynaptic signal may be modulated into a modulation signal. Next, the modulation signal is used to change the behavior at one or more synaptic elements or neurons of the third subset. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Embodiments will now be described, by way of example only, with reference to the accompanying drawings, in which corresponding reference characters indicate corresponding parts, and in which:

[0048] Figure 1A shows a typical implementation of a neural synapse array, Figure 1B The corresponding representation of a synaptic column with nominal (AMPA-like) synaptic integration and weight multiplication and neuronal summation is shown;

[0049] Figure 2 A representation of a synaptic column with AMPA synaptic integration and weight multiplication and neuronal summation is shown, but also including GABA (inhibitory) receptors;

[0050] Figure 3 A representation of a synaptic column with AMPA synaptic integration and weight multiplication and neuronal summation is shown, but also including GABA (inhibitory) receptors, and further extending the representation with NMDA receptors and the corresponding axonal (back-propagating) feedback system;

[0051] Figure 4 shows a schematic representation of a reconfigurable modular network consisting of neuromorphic synapses with multiple receptors, dendrites, and axonal feedback signals, respectively;

[0052] Figure 5 shows a schematic representation of a reconfigurable modular network consisting of neuromorphic synapses with multiple receptors, adaptive exponential neurons, and an astrocyte system;

[0053] Fig. 6A and 6BIt is shown that coordinated events in the source population convey feedforward a) excitation and b) inhibition; transitions to higher excitation / inhibition put neurons in fast / slow firing behavior with sparse bursts; for each synchronization partition, a population of postsynaptic neurons is assigned; synchronization partitions are mapped to postsynaptic activity patterns. DETAILED DESCRIPTION

[0054] Some embodiments will be described in more detail below. However, it should be understood that these embodiments should not be interpreted as limiting the scope of protection of the present disclosure.

[0055] Figure 1A A typical implementation of a neural synapse array 100 of a spiking neural network (SNN) is shown.

[0056] In this embodiment, the neural synapse array includes m×n synapses 102, each having a corresponding synaptic weight w i,j , where i∈{1,…,m}, j∈{1,…,n}, which connect m input signals 101 (e.g., digital or analog input pulses, preferably voltage spikes) to n neurons 103, with N k Represented by, where k∈{1,…,n}. Each neuron integrates the incoming signal and, when a specific membrane potential threshold is reached, may emit a pulse signal 105, which forms a spatiotemporal pulse train. In this example, an input signal 101 may be fed to a row of synapses 102, and each column of synapses may be fed to a specific neuron 103. Typically, the input signal 101 may be fed through one or more synapses 102 to one or more neurons 103. The synaptic array and its connections may be dynamically adjustable to optimize the performance and power consumption of the spiking neural network.

[0057] Input signals 101 can come from a variety of sources, such as sensors, images, audio, or any other form of data. Each synapse 102 between neurons in an SNN is associated with a synaptic weight, which determines the strength of the connection. During learning, the synaptic weights can be adjusted to optimize the performance of the network for a specific task. Neuron 103 integrates incoming signals from input connections or other connected neurons via synapses 102. The membrane potential of the neuron is updated based on incoming pulses from connected neurons and is determined by the synaptic strength w i,j When the membrane potential of neurons 103 exceeds a certain threshold, they will emit a pulse.

[0058] Neurons 103 can generate complex spatiotemporal spike patterns that contribute to the processing power of the network. SNNs can adjust to and learn from data through synaptic plasticity mechanisms. This means that the synaptic weights w i,jThe weights can vary over time based on the spike pattern and its timing. A learning rule is used to adjust the weights to improve the network's performance for a given task.

[0059] Typically, neurosynaptic structures, such as those implemented in spiking neural network cores, employ a hybrid analog-digital computing platform, i.e., spike trains can incorporate analog information in the timing of events, which can subsequently be converted back to analog representation at the input of the synaptic matrix. Each hybrid analog / digital core can consist of an input decoder and an I / O network communication layer that connects m×n programmable synapses to n neurons. Neurosynaptic computing elements can generate complex spatiotemporal dynamics that can be extended to specific features that can assist in the target signal processing function. Neuronal spike characteristics can be controlled by specific sets of parameters. The topology of neurosynaptic elements controls the regularity of spontaneous neuronal spikes, such as the firing rate scaling factor and the strength of the intrinsic noise, resulting in coherence / resonance phenomena.

[0060] The core router of the spiking neural network core (connected to the rest of the network via I / O ports) can provide input pulses to the row decoder. The row decoder receives the input pulses and can determine which row of the synaptic array to transmit a specific part of the signal to. There can be a learning and adaptation module. In addition, there can be a column decoder that receives the output of a column of synaptic elements and sends the decoded output of the column to one or more neurons. The neuron control module can change the neuron configuration, and the neuron control module can also obtain feedback from the learning and adaptation module. Pulses from one or more neurons can be sent to other cores of the spiking neural network via the core router.

[0061] The structures shown can all be implemented using electronic circuits, either completely or at least partially in hardware.

[0062] Figure 1B Shows about Figure 1A Corresponding representation of a typical synaptic column with nominal (AMPA-like) synaptic integration and weight multiplication and neuronal summation. Figure 1B express Figure 1A A specific implementation of a single synaptic column 102 (which is connected to a specific neuron 103) in a synaptic array.

[0063] Each synapse 102 in this exemplary embodiment includes an AMPA receptor that acts as an integrator. AMPA 111, which is connected to synaptic weights 112 (which may be adjustable). An integrator may be an element whose output signal is the time integral of its input signal. An integrator may accumulate input quantities over a defined time to produce a representative output.

[0064] AMPA receptor 111 may be a component that simulates the behavior of biological AMPA receptors at synapses, and is used to capture the manner in which excitatory synaptic transmission occurs in the brain, and integrate input signal 101 over time, thereby simulating the accumulation of excitatory currents in biological synapses.

[0065] Each synapse 102 may include an associated weight 112 that represents the strength of the connection. When a spike arrives at a synapse as an input signal (e.g., from a presynaptic neuron or from an encoder), the spike signal may be integrated by the AMPA receptor, after which the weight w of the synapse 102 may be calculated. i,j Vice versa, when a pulse arrives at the synapse as an input signal (e.g., from a presynaptic neuron or from an encoder), the weight w of the synapse 102 is applied to i,j applied to the pulse, and the resulting contribution can be integrated over the "AMPA receptor" unit, i.e., the order of integration and weight application can be reversed.

[0066] The cell body 113 of the neuron 103 can be used as a location for integrating incoming signals. These incoming signals are usually pulses (discrete events) that come from the input synapses 102 of the neuron. This causes the membrane potential of the neuron to change, and when a certain threshold is reached, the neuron 103 can be discharged, specifically, the neuron can be discharged from the axon 115 of the neuron 103. Once the artificial neuron 103 generates an output pulse, it will be further transmitted along the axon 115 of the neuron to, for example, other neurons to which the current neuron is connected. The transmission of this pulse imitates the propagation of action potentials along biological axons. In addition, a signal recovery unit 114 can be present in the neuron, which can restore the signal after signal attenuation has occurred. The signal recovery unit 114 may include, for example, a digital buffer.

[0067] Without loss of generality, in order to perform computations most efficiently in space and energy, i.e., to minimize the energy per bit of information, the present invention enables computations to be performed in a modular multi-layered manner across different hierarchical granularities, wherein each hierarchical level is optimized or selected based on the minimum associated cost. Thus, in this multi-layered structure, each set of basic signal processing components is modularly combined, e.g., synaptic integration is regulated by a gain-bandwidth product characteristic, a subsequent weight application / retention mechanism is controlled by a digital code-current transfer (i.e., digital (signal) to analog (current) conversion), and the final pulse generation is obtained by a current-frequency transfer characteristic, which is controlled by an adaptive mechanism, etc., to match the current to the environment. In an integrated system, this efficiency path is followed across different hierarchical levels while preserving information embodied in signal-to-noise ratio (SNR) and bandwidth. The structural implications are valid at different hierarchical levels (at finer scales and lower levels, as well as at larger scales and higher levels), i.e., the diversified circuits allow the inference engine to send only relevant information at a lower information rate, thereby providing a competitive computational performance / associated cost curve.

[0068] The multi-layered synaptic architecture enables different (e.g., fast, slow, inhibitory, potentiating) responses to input stimuli that are controlled by multiple mechanisms, such as non-competitive homosynaptic plasticity driven by residual activity, competitive heterosynaptic driven neighbor sensing processes, or mediated potentiation and homeostatic regulation in response to active inputs during normal sensory use. The model allows the assignment of various neuronal properties (e.g., axonal and dendritic delays, synaptic transfer functions), provides calculation of optimal input-output transfer functions and unique signal processing functions (e.g., filtering, amplification, multiplication, addition, etc.), allowing spike train kernel convolution in the time domain.

[0069] In integrated neuromorphic systems, the sensed signal, once encoded, is propagated passively (along the axon) to the synaptic terminal. If the physical distance on the chip is short enough for passive signaling (i.e., does not sufficiently degrade the signal (e.g., by exponential decay), then the signal can be transmitted passively. ) control), so no buffering / signal recovery is required), then the inference engine can read out the encoded information directly (analog, and therefore very efficiently). If the inference engine's analog signal contains more information than the action potential can encode, then in situ processing is required. Optimizing silicon circuits for best performance vs. associated cost (e.g., space, time, energy) may require a large diversity of circuit architectures at all spatial scales. Note that passive signals propagate in space and time as the square root of the diameter of the dendritic / synaptic column.

[0070] The present invention attempts to solve this problem by proposing a more adaptive synaptic structure. The proposed structure is based on Figure 1A and Figure 1BSimilar features will not be repeated in the embodiments shown, and can be easily discussed in conjunction with these embodiments.

[0071] Figure 2 Shown is a representation of a synaptic column with AMPA synaptic integration and weight multiplication and neuronal summation, but also including GABA (inhibitory) receptors.

[0072] Synaptic structure 200 includes synapse 202 (also referred to as synaptic element) including AMPA receptor 203 and / or GABA receptor 204. Each receptor may exhibit different temporal dynamics in response to an input signal. For example, each receptor may act as a filter for a different frequency.

[0073] The receptors trigger excitatory and inhibitory postsynaptic currents (EPSC and IPSC), respectively, thereby inducing charge inflow and outflow from the membrane voltage. AMPA receptors mediate fast (glutamatergic) synaptic currents to drive the cell body of the postsynaptic neuron. AMPA receptors have a high reaction rate to transmitter binding in both the rise and decay phases because the transmitter is quickly removed and the channel is quickly closed. See S. Hestrin, "Activation and desensitization of glutamate-activated channels mediating fast excitatory synaptic currents in the visual cortex," Neuron, vol. 9, no. 5, pp. 991–999, 1992.

[0074] GABA receptors are the main contributors to fast IPSCs, see A. Destexhe, ZF Mainen, TJ Sejnowski, "Kinetic models of synaptic transmission," Meth. Neuron. Model., vol. 2, pp. 1-25, 1998. In contrast, deprivation reduces the amplitude of inhibitory postsynaptic currents entering principal neurons and preferentially reduces the activation of fast-spiking interneurons. GABA receptor-mediated inhibition is a prerequisite for balancing excitation and inhibition, thereby stabilizing neural networks and achieving fine-grained temporal changes to drive precise plastic behavior. See SH Wu, CL Ma, JB Kelly, "Contribution of ampa, nmda, and gabaa receptors to temporal pattern of postsynaptic responses in the inferior colliculus of the rat," J. Neurosc., vol. 24, no. 19, pp. 4625-4634, 2004.

[0075] Similarly, in the homeostatic model, deprivation of a subset of inputs through GABAergic circuits drives a rapid Hebbian weakening on the deprived pathways and more slowly drives a homeostatic increase in global synaptic strength and / or intrinsic excitability, thereby increasing responses to the spared inputs. GABA receptors preferably have temporal dynamic behavior comparable to that of AMPA.

[0076] The synaptic element 202 is connected to at least one input port 201 and is configured to receive an input signal and output a postsynaptic signal. The synaptic element 202 includes an AMPA receptor 203, which is configured to integrate the input signal so as to make the postsynaptic signal more excitatory. Additionally or alternatively, the synaptic element 202 may include a GABA receptor 204, which is configured to integrate the input signal so as to make the postsynaptic signal more inhibitory. Both AMPA and GABA receptors may be present, but it may be dynamically determined which one is active and / or which frequency range they filter, i.e., over which frequency range these receptors are active. The outputs of the AMPA and / or GABA receptors are combined (if necessary, for example by summing).

[0077] There is also a weight element 205 in the synapse 202. The weight element 205 is configured to set the influence of the post-synaptic signal by having a specific weight that determines the strength of the connection represented by the synapse 202. For example, if the weight is high, the weight element 205 can be used to amplify signals from AMPA receptors and / or GABA receptors. If the weight is low, the weight element 205 can be used to attenuate signals from AMPA receptors and / or GABA receptors. The weight of the weight element, i.e., its influence, can be dynamically adjustable (e.g., by modulating the weight element 205), or can be static.

[0078] Neuron 206 can be connected to Figure 1B 206. That is, it may include a cell body 207 and an axon 209 that can fire once the membrane potential of neuron 206 reaches a certain threshold. Thus, the neuron may be configured to receive a postsynaptic signal from at least one synaptic element 202. Note again that the synaptic element 202 connected to neuron 206 represents Figure 1A The columns of the synaptic array are connected to synaptic elements of specific neurons, although other configurations are possible.

[0079] Figure 3 A representation of a synaptic column with AMPA synaptic integration and weight multiplication and neuronal summation is shown, but also includes GABA (inhibitory) receptors, and further extends the representation with NMDA receptors and the corresponding axonal (back-propagating) feedback system.

[0080] Without loss of generality, all major glutamate receptors, AMPA and NMDA, and GABAa receptors are included, which exhibit different temporal dynamics in their respective responses to neurotransmitters. These receptors elicit excitatory and inhibitory postsynaptic currents (EPSCs and IPSCs), respectively, inducing charge influx and efflux from the membrane voltage. Without loss of generality, scaling of excitatory synapses to principal neurons is expressed primarily through modulation of AMPA receptor insertion via NMDA potentiation and inhibition.

[0081] NMDA receptors provide activity-dependent modifications to synaptic weights, i.e., more complex mechanisms including NMDA receptor-dependent inhibition, which inhibits processes mediated by synaptic AMPA receptor currents, and NMDA potentiation, which enables AMPA receptor currents to occur at weak synapses, thereby functionalizing these synapses. They also homeostatically regulate the intrinsic excitation and excitation-inhibition balance by setting the permissive excitation-inhibition balance, or editing firing patterns to promote excitatory synaptic plasticity.

[0082] The neural synapse structure 300 includes an axon backpropagation signal path 311 connected to a neuron 307 and at least one synaptic element 302, wherein the axon backpropagation signal path 311 is configured to receive an axon signal from an axon 308 of the neuron 307 and includes an axon backpropagation receptor 312 configured to integrate the axon signal. The axon backpropagation signal path 311 is configured to output the integrated axon signal to the at least one synaptic element 302. The axon backpropagation receptor 312 may be a pure integrator element, which may be voltage driven.

[0083] At least one synaptic element includes an NMDA receptor 305, which is configured to integrate the input signal received at the input port 301 into an integrated NMDA signal. The integrated NMDA signal and the integrated axon signal are combined and modulate the weight element 306 of at least one synaptic element 302 to change the influence of the postsynaptic signal of at least one synaptic element 302. Therefore, by modulating the weight element 306, the weight of the weight element can be dynamically changed. Note that the integrated axon signal can be used to combine with the integrated NMDA signal, but the time derivative of the integrated axon signal is preferably used. Therefore, a derivation unit 313 can be placed along the axonal back propagation signal path 311, for example, in the synaptic element 302. The derivation unit 313 can be, for example, a differentiator, which is a circuit designed to produce an output that is roughly proportional to the rate of change (time derivative) of the input. The differentiator can essentially be a high-pass filter. An active differentiator can include some form of amplifier, while a passive differentiator can be made of only resistors, capacitors and / or inductors.

[0084] The remaining synaptic elements 302 and neurons 307 can be connected to Figure 2 Similarly. That is, there may also be AMPA 303 and / or GABA 304 receptors that integrate the input signal, and their individual or combined outputs are sent to the weight element 306. The neuron receives a weighted integrated postsynaptic signal from each synapse 302 of the neuron 307 connected to the cell body 309, wherein the different postsynaptic signals are combined, for example, by summing. There may be a signal recovery unit 310 in the neuron 307. When the membrane potential of the neuron 307 reaches a certain threshold, the neuron may generate a pulse via its axon 308.

[0085] Note that while each synaptic element 302 is shown as having AMPA, GABA, and NMDA receptors and an axonal backpropagation signal path 311, this is not required. Some synaptic elements connected to neurons 307 may have only AMPA receptors and weight elements, some may have only GABA receptors and weight elements, and some may have AMPA and GABA receptors and weight elements. Synaptic elements may also have AMPA and / or GABA receptors, and weight elements modulated by NMDA receptors and an axonal backpropagation signal path 311 (preferably with a derivative unit 313).

[0086] Note that the weight element 306 may also be modulated based on the NMDA receptor output alone, or the output of the axonal back-propagating signal path 311 (again, preferably with a derivative unit 313).

[0087] Figure 4 A schematic representation of a reconfigurable modular neurosynaptic structure 400 is shown, which consists of a neuromorphic synapse with multiple receptors and dendritic and axonal feedback signals, respectively.

[0088] The neural synapse structure 400 may further include a dendritic backpropagation signal path 408 connected to one or more synaptic elements 402, wherein the dendritic backpropagation signal path 408 is configured to receive and sum the postsynaptic signals of each synaptic element 402. Preferably, the dendritic backpropagation signal path 408 includes a signal recovery unit 407, which may have a similar working mode as the signal recovery unit 114. The dendritic backpropagation signal path 408 may include a dendritic receptor 409, which is configured to integrate the summed postsynaptic signal. The dendritic backpropagation signal path 408 is configured to output the obtained integrated summed postsynaptic signal (also referred to as a dendritic signal) to each of the one or more synaptic elements 402 connected thereto.

[0089] Each of these one or more synaptic elements 402 may include an NMDA receptor 405, which is configured to integrate the input signal of the synaptic element received at the input port 401 into an integrated NMDA signal. The integrated NMDA signal and the integrated dendritic signal can be combined, and the weight element 406 of the synaptic element can be modulated to change the effect of the postsynaptic signal of at least one synaptic element 402 on, for example, a neuron 414. Note that the integrated dendritic signal can be used to combine with the integrated NMDA signal, but the time derivative of the integrated dendritic signal is preferably used. Therefore, a derivation unit 410 (similar to the unit described above) can be placed along the dendritic backpropagation signal path 408, for example, in the synaptic element 402. Preferably, if there is also an axon backpropagation signal path 416, as in the present exemplary embodiment, the dendritic signal and the axon signal are combined (for example, by summing the signals), and then the derivative of the combined signal is obtained using the derivation unit. Of course, the derivatives of the two signals can also be determined separately and then combined.

[0090] In one embodiment, all synapses connected to a particular neuron are also connected to the same dendritic back-propagating signal path 408. In this case, the summed postsynaptic signal of each synaptic element can be forwarded to the cell body 413 of the neuron. Alternatively, a separate post-dendritic receptor 411 integrates the summed postsynaptic signal and then sends it to the cell body 413 of the neuron 414.

[0091] In another embodiment, the dendrites are connected only to a specific subset of the synaptic elements 402 that are connected to neurons 414. In this case, other synaptic outputs can be received at the cell body 413 via other synaptic segment outputs 412. These segments can also have dendritic structures with dendritic back-propagating signal paths 408. Each dendritic structure can have a separate post-dendritic receptor 411. This is the present exemplary embodiment shown in the figure.

[0092] Other structures have the same functions as Figure 3 404. That is, there may also be AMPA 403 and / or GABA 404 receptors that integrate the input signal, and their individual or combined outputs are sent to the weight element 406. The neuron receives the weighted integrated postsynaptic signal at the cell body 413, wherein different postsynaptic signals are combined, for example, by summing. There may be a signal recovery unit 407 in the neuron 414. When the membrane potential of the neuron 414 reaches a certain threshold, the neuron may generate a pulse via its axon 415. There may also be an axon back-propagation signal path 416.

[0093] Similarly, as before, different feedback paths can be implemented individually or in combination. Note that it is also possible to modulate the weight element 406 based only on the NMDA receptor output, or based only on the output of the dendritic backpropagation signal path 408 (again, preferably with a derivation unit 409). This time, the axonal backpropagation signal path 416 is also optionally combined.

[0094] To achieve an efficient energy-delay product over short distances, the signal is converted into a current that charges the membrane capacitance; for longer distances, the current is regenerated by appropriately clustered voltage-gated NMDA channels. Thus, in the postsynaptic part, the temporal summation of the postsynaptic backpropagating impulses (i.e., the dendritic impulses and the somatic impulses that initiated the dendritic impulses, respectively) is accomplished. This implementation allows a wide range of temporal behaviors, since each receptor can operate on a different time scale, and a higher noise tolerance is obtained, since the ability of the network to perform pattern recognition even in noisy signals is increased.

[0095] If the dendritic spike group (burst) is strong enough to drive the neuron, the neuron can generate an action potential; for this, the resulting spike can be back-propagated into the dendrites. See RC Froemke, Y. Dan, "Spike-timing-dependent synaptic modification induced by natural spike trains", Nature, vol. 416, pp. 434-438, 2002. The back-propagating dendritic and neuronal signals can be multiplied and added with the NMDA receptor signal to form a weight control signal.

[0096] Individual receptors are susceptible to random fluctuations (e.g., thermal noise); therefore, neurons may need to improve the signal-to-noise ratio of the signal transmitted by one receptor by averaging over a population of the same type. Increasing the number of receptors n increases the SNR by However, since it comes with an associated higher cost, it is reserved for special purposes, e.g., synapses that transmit with high temporal precision. To achieve an optimal allocation of resources, the noise resource equation can be used, i.e., weighting power and / or area to reduce thermal / flicker noise. If noise is to be minimized, more resources should be distributed to parts of the system that affect all other parts of it (initial stages) and to complex parts of the system (i.e., parts with a larger number of receptors, with high n i express).

[0097] Typically, in biochemical systems, noise sources can be traced to different hierarchical scales: i) biophysical sources (microscopic fluctuations of small structural components of neurons (e.g., ion channels or synaptic machinery involved in vesicle release)), ii) neural circuit dynamics (excitatory / inhibitory behavior, sparsity), iii) global network properties (average instantaneous population activity, synchrony). Similarly, variability effects associated with modern IC manufacturability can be broken down into different levels of spatial granularity in the physical dimensions of the chip, ranging from global, i.e., IC level (including technology (process corner), electrical (supply voltage), thermal (temperature), lithography (line width)), to local effects, i.e., at the sub-circuit (substrate noise, temperature gradients, layer density), transistor (well proximity, local heating, jitter, crosstalk, aging effects), or atomic (thermal noise, mobility variations, random dopants, mechanical stress, line edge roughness) level.

[0098] Figure 2-5 The structure shown in allows this hierarchical granularity to be achieved.

[0099] Therefore, variability effects when broken down across different time granularities can be linked to methods to mitigate performance or yield loss, i.e., can be broken down into static, slow, medium, and fast varying events based on the temporal dependency of the variability artifacts. Static effects allow for one-time corrections (e.g., fine tuning) of associated devices in the circuit, while slow time-dependent effects (i.e., static over the observation period) can be corrected during operation using on-chip calibration mechanisms. Variations over time spans comparable to the maximum speed of the process (e.g., jitter, substrate noise, kT noise, dynamic IR drop) cannot be calibrated directly and must be accommodated by expanding the design margin. Comparability and plasticity mechanisms occur at different timescales and subsequently across different integration verticals, such as short-term plasticity (adaptive neuronal behavior, short-term synaptic adaptation), long-term plasticity (spike-based plasticity rules), homeostatic plasticity (enlargement of excitatory synaptic strength, specific changes in synaptic elements in inhibitory circuits and changes in intrinsic excitability, depending on the precise deprivation / depression definition / paradigm, all in order to stabilize the range of neuronal firing frequencies and, therefore, network activity), structural plasticity (modification of network connectivity). Compensation can therefore be performed at different levels of time and space and across different framework modules, different computational stacks, circuit level, system and architecture level, across hardware and software lines.

[0100] The synaptic weight induced by Hebbian STDP co-activation will deteriorate unless it is consolidated. The initial synaptic plasticity value can set a label at the synapse, thereby defining a mark for prospective consolidation (i.e., transfer of information through the writing process) when the synaptic efficacy changes. The intrinsic plasticity rule based on the principle of information maximization is as follows as a non-limiting example:

[0101]

[0102] w ij is the weight, label η ij The coupling with the nearest neighbors is done via a time-dependent gating variable, as given by the formula above. is the external input, χ is the steady-state scaling factor, is an independent Gaussian white noise process, α ηw is the coupling parameter term, which determines the strength of the interaction between variables, τ w is the time constant of the association and g acts as a gating variable. Furthermore, f is a (possibly discontinuous) function that converts the net current arriving at the population into population activity and specifies the firing rate (in Hz). While the regulation time constant determines the direction in the conductance space, the steady-state scaling factor controls the trajectory until equilibrium is reached. It is assumed here that a node can only excite its own inhibitory population. The neuronal membrane potential creates the permissive time window for inducing perceptual context-dependent bidirectional plasticity mechanisms. This plasticity rule for different classes of excitatory and inhibitory synapses is consistent with an overall steady-state shift in the balance between excitation and inhibition.

[0103] Synaptic scaling is mediated by multiple mechanisms that vary at various hierarchical levels depending on synapse type, time course, array region, and architectural stage. The receptor types in synaptic implementations are more diverse than the neurons they connect.

[0104] For the inference machine, the synapse count is more restricted, i.e., in a system where each uniformly defined receptor forming a synapse does not have a dedicated neuron associated with a synaptic column, neurons can be shared with synaptic clusters containing a set of receptors that lead to the same final action. Therefore, the input signals provided to the neuromorphic array are not assembled by high-level circuits to perform actions, but by dedicated neurons capable of providing complex functions, to save on the number of neurons.

[0105] Morphological features are considered as specific control parameters that contribute significantly to the dynamics of neural synaptic elements and the patterns of network activity. Without loss of generality, the formation of synaptic clusters in terms of size and functionality can be defined based on predefined contributions of subcellular morphological features to intracellular dynamics, such as specific signal processing functions and coding schemes, specific data rates, frequency bands of interest, feature sets or complexity, target SNR, time constants, etc.

[0106] This provides a paradigm shift in how to provide sufficient resources for complex inference tasks within a feasible power-performance-area design space, whether by increasing the synapse count of simplified (or specialized) computational elements, or by increasing the computational power of (alternatively complex, multifunctional) elements within a reduced synapse count. In these complex, multifunctional elements, separate parts can be defined for each task, i.e., synapses include different receptors, where each receptor can be independently tuned for speed, sensitivity can be independently adjusted, and more opportunities for further refinement and redundancy for fault protection can be provided.

[0107] Without loss of generality, complexity can include the realization of synaptic element(s) with nominally similar functions but supporting multiple pathways and having slightly different properties, such as axons and astrocytes both expressing activation / responsibility capabilities facilitated by different mechanisms. Nominally, communication and information transfer within neural networks has been the sole domain of pre- and postsynaptic coupling between neurons. However, the coupling of astrocytes and neurons provides additional pathways for intercellular communication; communication between neurons and astrocytes is facilitated by additional extrasynaptic compounds released upon arrival of presynaptic action potentials.

[0108] Figure 5 Shown is a schematic representation of a reconfigurable modular network consisting of neuromorphic synapses with multiple receptors, adaptive exponential neurons, and an astrocyte system.

[0109] This embodiment shows AMPA receptors 503 and / or GABA receptors 504, NMDA receptors 505, axonal back propagation signal paths 516, and dendritic back propagation signal paths 508. They work in the same manner as described above, and can act alone or in different combinations for each synapse 502 connected to the input port 501.

[0110] Two new structures are introduced in this embodiment, namely the neuron back-propagation signal path 520 and the astrocyte unit / structure 530.

[0111] Thus, the neuron 512 may include a neuron backpropagation signal path 520 connected from the axon 515 of the neuron 512 to the cell body of the neuron. For neural frequency / firing regulation; when the pulse frequency is high enough, the astrocytes use internal memory to reach a steady state, so the neural pulse frequency will persist. The neuron backpropagation signal path 520 may include a local receptor 518 configured to integrate the axon signal. The local receptor 518 may be a pure integrator element and is voltage driven. It can act as a filter for a specific frequency range of the axon signal. In addition, the neuron backpropagation signal path 520 may include a calcium element 519, which is configured to: set the influence of the integrated axon signal, thereby generating a neuron feedback signal. The calcium element 519 can be a fast amplifier, preferably providing adjustable feedback. The neuron backpropagation signal path 520 is configured to: feed the neuron feedback signal to the cell body 514 of the neuron 512.

[0112] At the cell body 514, the neuron feedback signal can be combined with signals from the dendrites of a set of synapses 502 (which can be all synapses connected to the neuron 512, or a subset) and / or signals from other segments 513 (which can have a dendritic structure or not between the synapse 502 and the neuron 512), similar to Figure 4 Embodiment of the invention.

[0113] The neural synapse structure 500 may additionally or alternatively include an astrocyte structure 530 connected to one or more synapse elements 502. The astrocyte structure 530 may be configured to receive and sum postsynaptic signals from each of the plurality of synapse elements, and may include a modulation circuit configured to modulate the summed postsynaptic signal into a modulation signal.

[0114] The astrocyte structure 530 may include an mGluR receptor 523 configured to integrate the modulation signal. Thereafter, the astrocyte structure may be configured to output the integrated modulation signal to each synaptic element 502 (or a subset thereof). For example, each synaptic element 502 (or a subset thereof) for which the postsynaptic signal is used as an input to be modulated by the astrocyte structure may be configured to sum the integrated modulation signal received from the astrocyte structure with the integrated input signal of the AMPA receptor 503 and / or the GABA receptor 504.

[0115] In addition to or as an alternative to feedback from the astrocyte structure to the synaptic element, the astrocyte structure can include an NMDA receptor 524 configured to integrate the modulation signal from the modulation circuit. The NMDA receptor 524 can be similar to the NMDA receptor 505, but optimized differently, such as operating in a different frequency domain. If there is a neuronal back-propagation signal path 520, the astrocyte structure 530 can be configured to output the integrated modulation signal from the NMDA receptor 524 to the neuron 512. The integrated modulation signal from the NMDA receptor 524 can modulate the calcium element of the neuron to change the effect of the integrated axon signal of the neuron from the local receptor 518 on the cell body 514 of the neuron.

[0116] The modulation circuit of the astrocyte structure 530 may include an IP3 receptor 521, which is configured to integrate the summed postsynaptic signal. The IP3 receptor 521 can be a pure integrator, which is voltage driven. In addition, the modulation circuit may include an astrocyte calcium element 522, which can be a fast amplifier, preferably providing adjustable feedback. The astrocyte calcium element 522 can be configured to set the influence of the modulation signal. The IP3 receptor 521 and the astrocyte calcium element 522 can generate a modulation signal. Optionally, there are one or more signal recovery units 507 in the modulation circuit, which have a similar working mode to the above-mentioned signal recovery unit.

[0117] The astrocyte calcium element 522 can be adaptive in that it can be modulated. By modulating the astrocyte calcium element 522, the effect of the modulating signal on the structure of the astrocyte structure 530 that provides feedback (e.g., the synaptic element 502 or the neuron 512) can be changed.

[0118] Presynaptic neuron signals or postsynaptic signals from synaptic elements not directly connected to neuron 512 may be received at input 525 and integrated using dendritic receptors 526. Dendritic receptors 526 may be similar to dendritic receptors 509 or 511, but may also operate over a different frequency range or have other differences in their operating parameters.

[0119] The axon backpropagation signal may be integrated at an axon backpropagation receptor 527, which may operate in a manner similar to the axon backpropagation receptor 517, but may also operate over a different frequency range or have other differences in its operating parameters. The integrated axon backpropagation signal may be combined with other axon backpropagation signals from the postsynaptic neuron at input 528, or only axon backpropagation signals from other postsynaptic neurons may be used as input. Preferably, the (combined) axon input signal is time-derivatived by a derivation unit 529, which operates in a manner similar to the integration unit 510.

[0120] Input from the dendritic receptors 526 and axonal feedback can be combined (if both are present, but these can also be used separately), such as by a multiplication unit, and the resulting signal can be fed to the astrocyte calcium element 522 so as to modulate the astrocyte calcium element. As described above, the effect of the modulating signal on the structure (e.g., synaptic element 502 or neuron 512) to which the astrocyte structure 530 provides feedback can thus be changed.

[0121] Astrocyte units employed within spiking neurons in pulsating astrocyte-neuron networks can perform distributed and fine-grained compensation and / or self-repair, e.g., if synaptic operation is disrupted, retrograde control mechanisms effectively reinitiate control and, by extension, plasticity processes.

[0122] Although astrocytes do not directly initiate propagating action potentials like neurons, they can communicate bidirectionally with neurons and other astrocytes. At synapses, which include connections between astrocytes and several neurons, i.e., synapses exchange signals at three terminals, when an action potential arrives, the accumulated charge is released at the synapse, causing depolarization of the postsynaptic neuron.

[0123] Activation of astrocyte receptors results in a transient increase in calcium levels within the astrocyte cell, which represents the fundamental excitatory mode of astrocytes, and astrocyte-neuron signaling, i.e., a slow inward current operating over much longer rise and decay times than typical excitatory postsynaptic currents.

[0124] When the postsynaptic neuron is sufficiently depolarized, voltage-gated calcium channels on the dendrites allow calcium to flow into the dendrites, which also enables control mechanisms to regulate back-propagating signals from the postsynaptic to the presynaptic neurons. Synaptic-driven calcium transients in astrocytes modulate synaptic transmission via the release of astrocytic glutamate (mGluR) bound to presynaptic receptors. This then initiates the production and release of IP3 within astrocytes and triggers a transient intracellular release of calcium, which subsequently leads to the activation of NMDA receptors on the postsynaptic neuron.

[0125] Astrocytes can also support the synchronous or partially synchronous construction of neighboring neurons, which can be enhanced by increasing inhibitory dynamics. In heterogeneous inhibitory networks with sparse random connections, inhibitory dynamics have a dual nature in suppressing population activity and generating neural reactivation. When the input from excitatory neurons to neighboring inhibitory neurons is strong enough, the pulse transmission within the circuit will be synchronized. The coupling strength is controlled by inhibitory innervation: when inhibitory input is present, that is, intermediate coupling, neurons will generate pulses with slight rhythmicity / synchrony. Fig. 6A and Figure 6BThe effect of this coupling is shown.

[0126] Fig. 6A and Figure 6B It is shown that coordinated events in the source population convey feedforward a) excitation and b) inhibition; transitions to higher excitation / inhibition put neurons in fast / slow firing behavior with sparse bursts; for each synchronization partition, a population of postsynaptic neurons is assigned; synchronization partitions are mapped to postsynaptic activity patterns.

[0127] The duration selectivity curve is asymmetric to account for the heterogeneity of spike delays and resulting delay profiles. The initial transient dynamics set by synaptic coupling guide the network towards the stationary dynamics region. As coupling increases, there is a transition from a group of individually firing neurons to a coherent synchronized network. Neurons generate time-locked patterns; due to the interaction between conductance delays and plasticity rules, the network forms a group of neurons with reproducible and precise firing sequences that depend on the activation pattern. As conductance increases (thus, resulting in net excitation of the network), the firing rate increases and becomes more uniform with lower coefficients of variability. The sequencing of synaptic currents (i.e., outward, inward) reduces the temporal jitter in the generation of action potentials in individual neurons, thereby creating a network with enhanced controllability of synchronized activity and homeostatic regulation.

[0128] The two figures show the positions on the two axes. Through the present invention, the excitement (A) and inhibition (B) events are located in space, as shown by the black spots. If these figures are superimposed, the areas in space where excitement and inhibition occur can be used to control dynamics using excitement and inhibition.

[0129] Note that the features of any embodiment disclosed herein can be combined in an appropriate manner. For example, the different feedback and control structures disclosed above, such as AMPA receptors, GABA receptors and NMDA receptors, axonal backpropagation channels, dendritic backpropagation channels and neuronal backpropagation channels and astrocyte structures, can be used alone or in an appropriate manner. The same is true for dendritic receptors, IP3 receptors, mGluR receptors, local neuronal receptors and axonal backpropagation receptors, as well as weight elements and calcium elements, derivative units (differentiators) and signal recovery units.

[0130] As an example, the receptors can be formed using gm-C filters, the neural pulse generators can be (positive feedback reinforced) inverters, and the weight elements can be implemented by different (static or dynamic) memory elements (latches, leakage compensation capacitors, SRAM cells, etc.).

Claims

1. A neural synapse structure (200) for a spiking neural network, wherein: The neural synapse structure (200) includes a plurality of synapse elements (202) and neurons (206), wherein each synapse element (202) includes: one or more input ports (201), each input port configured to receive an input signal; an output port configured to output a post-synaptic signal; and a signal path connecting the one or more input ports to the output port, the signal path comprising a weight element (205) configured to: apply a weight for generating the post-synaptic signal; Wherein, the signal path of at least a portion of the synaptic elements (202) includes an AMPA receptor (203), and the AMPA receptor is configured to: integrate the input signal to make the post-synaptic signal more excited; wherein the signaling pathway of at least a portion of the synaptic element (202) includes a GABBA receptor (204), the GABBA receptor being configured to: integrate the input signal to make the postsynaptic signal more inhibitory; and The neuron (206) includes a cell body (207) and an axon (209), wherein the cell body (207) is configured to receive and combine one or more post-synaptic signals, and the axon (209) is configured to generate a pulse output when the membrane potential of the neuron (206) reaches a predetermined threshold in response to the output of the cell body.

2. The neural synapse structure (200) according to claim 1, wherein: At least one synaptic element (202) includes a signaling pathway comprising an AMPA receptor (203) and a GABBA receptor (204), and Wherein, the at least one synaptic element is configured to combine outputs from the AMPA receptor and the GABBA receptor to generate the postsynaptic signal.

3. The neural synapse structure (200) according to claim 2, wherein: The AMPA receptors and / or GABBA receptors of the neuronal synaptic element may be configured to be active or passive at a particular moment, and / or Wherein, the AMPA receptor and / or the GABBA receptor of the neural synapse element may be configured to: select a frequency range of the input signal within which the AMPA receptor and / or the GABBA receptor may operate.

4. The neural synapse structure (200) according to any one of the preceding claims, wherein: The weighting element is located along the signaling pathway after the AMPA receptor and / or the GABBA receptor.

5. The neural synapse structure (200) according to any one of the preceding claims, wherein: The weight element (205) is dynamically adjustable, preferably by modulating the weight element (205).

6. The neural synapse structure (200) according to any one of the preceding claims, wherein: The AMPA receptor comprises a voltage driven integrator configured to generate an excitatory postsynaptic current for inducing charge inflow from a membrane potential of the neuron, and / or The GABBA receptor is a voltage-driven integrator configured to generate an inhibitory postsynaptic current for inducing charge outflow from the membrane potential of the neuron.

7. A neural synapse structure (300) for a spiking neural network, wherein: The neural synapse structure (300) includes a plurality of synapse elements (302) and neurons (307), wherein each synapse element (302) includes: one or more input ports (301), each input port configured to receive an input signal; an output port configured to output a post-synaptic signal; and a signal path connecting the one or more input ports to the output port, the signal path comprising a weight element (306) configured to: apply a weight for generating the post-synaptic signal; The neuron (307) includes a cell body (309) and an axon (308), wherein the cell body (309) is configured to: receive and combine one or more post-synaptic signals, and the axon (308) is configured to: generate a pulse output when the membrane potential of the neuron (307) reaches a predetermined threshold in response to the output of the cell body; Wherein, the neural synapse structure (300) further comprises an axon back propagation signal path (311), wherein the axon back propagation signal path is connected to receive an axon signal from an axon (308) of the neuron (307); Wherein, the axon backpropagation signal path (311) comprises an axon backpropagation receptor (312), and the backpropagation receptor is configured to: integrate the axon signal; and The first subset of synaptic elements (302) is configured to receive the integrated axon signal and modulate the weight elements (306) of the synaptic elements (302) of the first subset based on the integrated axon signal.

8. The neural synapse structure (300) according to claim 7, wherein: Each synaptic element (302) of the first subset of synaptic elements comprises an NMDA receptor configured to: integrate the input signal to form an integrated NMDA signal, and The synaptic element (302) in the first subset of synaptic elements is configured to: combine the integrated NMDA signal and the integrated axon signal, and modulate a weight element (306) of the synaptic element (302) based on the combined signal.

9. The neural synapse structure (300) according to claim 7 or 8, wherein: The integrated axon signal is time-differentiated by a derivation unit (313), preferably a differentiator, and then the differentiated integrated axon signal is used to modulate the weight element (306).

10. The neural synapse structure (300) according to any one of claims 7 to 9, wherein: The axonal backpropagation receptor (312) is a voltage driven pure integrator, Preferably, wherein the axonal backpropagation receptors (312) are formed using one or more gm-C filters.

11. The neural synapse structure (300) according to any one of claims 7 to 10, wherein: At least two axonal backpropagation receptors are used to backpropagate to at least two sets of synaptic elements.

12. The neural synapse structure (300) according to any one of claims 7 to 11, wherein: The neural synapse structure (300) is also the neural synapse structure (200) according to any one of claims 1 to 6.

13. A neural synapse structure (400) for a spiking neural network, wherein: The neural synapse structure (400) includes a plurality of synapse elements (402) and neurons, wherein each synapse element (402) includes: one or more input ports (401), each input port configured to receive an input signal; an output port configured to output a post-synaptic signal; and a signal path connecting the one or more input ports to the output port, the signal path comprising a weight element (406) configured to: apply a weight for generating the post-synaptic signal; The neuron comprises a cell body (413) and an axon (415), wherein the cell body (413) is configured to: receive and combine one or more post-synaptic signals, and the axon (415) is configured to: generate a pulse output when the membrane potential of the neuron reaches a predetermined threshold in response to the output of the cell body; Wherein, the neural synapse structure further comprises a dendritic back propagation signal path (408), wherein the dendritic back propagation signal path is configured to: receive and combine post-synaptic signals from one or more synaptic elements (402); wherein the dendritic back propagation signal path (408) includes a dendritic receptor (409) configured to: integrate a combined postsynaptic signal; and The second subset of synaptic elements (402) is configured to: receive the integrated combined post-synaptic signal, and modulate the weight elements (406) of the synaptic elements (402) of the second subset based on the integrated combined post-synaptic signal.

14. The neural synapse structure (400) according to claim 13, wherein: Each synaptic element (402) of the second subset of synaptic elements comprises an NMDA receptor (405) configured to: integrate the input signal to form an integrated NMDA signal, and The synaptic element (402) in the second subset of synaptic elements is configured to: combine the integrated NMDA signal and the integrated combined post-synaptic signal, and modulate a weight element (406) of the synaptic element based on the combined signal.

15. The neural synapse structure (400) according to claim 13 or 14, wherein: The integrated combined post-synaptic signal is time-differentiated by a derivation unit (313), preferably a differentiator, and then the differentiated integrated combined post-synaptic signal is used to modulate the weight element (306).

16. The neural synapse structure (400) according to any one of claims 13 to 15, wherein: The dendritic receptor (409) is a voltage-driven pure integrator, Preferably, wherein the dendritic receptors (409) are formed using one or more gm-C filters.

17. The neural synapse structure (400) according to any one of claims 13 to 16, wherein: The neuron is configured to receive one or more postsynaptic signals from the one or more synaptic elements at the cell body (413) via a second dendritic receptor (409), the second dendritic receptor (409) integrating the combined postsynaptic signals of each of the one or more synaptic elements (402), Preferably, wherein the dendritic receptor (409) is a voltage-driven pure integrator, More preferably, wherein said dendritic receptors (409) are formed using one or more gm-C filters.

18. The neural synapse structure (400) according to any one of claims 13 to 17, wherein: combining one or more postsynaptic signals at the cell body by combining the combined, preferably also integrated, postsynaptic signals of each synaptic element (402) in the second subset with one or more postsynaptic signals from one or more synaptic elements not in the second subset, Preferably, one or more synaptic elements not in the second subset are also connected to the neuron via one or more different dendritic structures.

19. The neural synapse structure (400) according to any one of claims 13 to 18, wherein: The neural synapse structure (400) is also the neural synapse structure (200, 300) according to any one of claims 1 to 12.

20. A neural synapse structure (500) for a spiking neural network, wherein: The neural synapse structure (500) includes a plurality of synapse elements (502) and neurons (512), wherein each synapse element (502) includes: one or more input ports (501), each input port configured to receive an input signal; an output port configured to output a post-synaptic signal; and a signal path connecting the one or more input ports to the output port, the signal path comprising a weight element (506) configured to: apply a weight for generating the post-synaptic signal; The neuron (512) includes a cell body (514) and an axon (515), wherein the cell body (514) is configured to: receive and combine one or more post-synaptic signals, and the axon (515) is configured to: generate a pulse output when the membrane potential of the neuron (512) reaches a predetermined threshold in response to the output of the cell body; wherein the neuron (512) comprises a neuron back propagation signal path (520), the neuron back propagation signal path being connected from an axon (515) of the neuron (512) to a cell body (514) of the neuron (512), and comprising a local receptor (518) configured to integrate the axon signal and a calcium element (519) configured to set an effect of the integrated axon signal to generate a neuron feedback signal, and The neuron back propagation signal path (520) is configured to feed the neuron feedback signal to the cell body (514) of the neuron (512), where the neuron feedback signal is combined with the one or more post-synaptic signals.

21. The neural synapse structure (500) according to claim 20, wherein: The calcium element (519) is an amplifier, preferably a faster amplifier than the weight element (506) of each synaptic element, and / or Therein, the calcium element (519) provides adaptive feedback.

22. The neural synapse structure (500) according to claim 20 or 21, wherein: The local receptor (518) is a voltage driven pure integrator, Preferably, wherein the local receptor (518) is formed using one or more gm-C filters.

23. The neural synapse structure (500) according to any one of claims 20 to 22, wherein: The neural synapse structure (500) is also the neural synapse structure (200, 300, 400) according to any one of claims 1 to 19.

24. A neural synapse structure (500) for a spiking neural network, wherein: The neural synapse structure (500) includes a plurality of synapse elements (502) and neurons (512), wherein each synapse element (502) includes: one or more input ports (501), each input port configured to receive an input signal; an output port configured to output a post-synaptic signal; and a signal path connecting the one or more input ports to the output port, the signal path comprising a weight element (506) configured to: apply a weight for generating the post-synaptic signal; The neuron (512) includes a cell body (514) and an axon (515), wherein the cell body (514) is configured to receive and combine one or more post-synaptic signals, and the axon (515) is configured to generate a pulse output when the membrane potential of the neuron (512) reaches a predetermined threshold in response to the output of the cell body, wherein the neural synaptic structure comprises an astrocyte structure (530) connected to a third subset of the synaptic elements; Wherein, the astrocyte structure is configured to: receive and sum the postsynaptic signals of each synaptic element, and comprises a modulation circuit, wherein the modulation circuit is configured to: modulate the summed postsynaptic signals into a modulation signal; The modulation signal is used to change the behavior of one or more synaptic elements (502) or neurons (512) of the third subset.

25. The neural synapse structure (500) according to claim 24, wherein: The astrocyte structure comprises an mGluR receptor (523) configured to integrate the modulating signal, and wherein the astrocyte structure is configured to: output an integrated modulation signal to each synaptic element in the third subset, and wherein each of the plurality of synaptic elements is configured to: combine the integral modulation signal with an integral input signal, Wherein, the integrated input signal is obtained by integrating the input signal using AMPA and / or GABBA receptors located along a signal path configured to integrate the input signal.

26. The neural synapse structure (500) according to claim 25, wherein: The mGluR receptor (523) is a voltage-driven pure integrator. Preferably, wherein the mGluR receptor (523) is formed using one or more gm-C filters.

27. The neural synapse structure (500) according to any one of claims 24 to 26, wherein: The astrocyte structure comprises an NMDA receptor (524) configured to integrate the modulating signal, and The astrocyte structure is configured to output an integrated modulation signal to the neuron (512).

28. The neural synapse structure (500) according to claim 27, wherein: The neuron (512) includes a neuron back propagation signal path (520) connecting an axon of the neuron (512) to a cell body (514) and includes a calcium element (519) configured to: configure an effect of an integrated axon signal to generate a neuron feedback signal, and wherein the neuron back propagation signal path (520) is configured to: feed the neuron feedback signal to the cell body (514) of the neuron (512), and The integrated modulation signal modulates the calcium element (519) of the neuron (512) to change the influence of the integrated axon signal of the neuron (512).

29. The neural synapse structure (500) according to any one of claims 24 to 28, wherein: The modulation circuit includes an IP3 receptor (521) configured to integrate the summed postsynaptic signal and an astrocyte calcium element (522) configured to set the effect of the modulation signal, Wherein, the IP3 receptor (521) and the astrocyte calcium element (522) generate the modulation signal.

30. The neural synapse structure (500) according to claim 29, wherein: The IP3 receptor (521) is a voltage-driven pure integrator. Preferably, wherein the IP3 receptor (521) is formed using one or more gm-C filters, and / or wherein the astrocyte calcium element (522) is an amplifier, preferably a faster amplifier than the weight element (506) of each synaptic element, and / or Therein, the astrocyte calcium element (522) provides adaptive feedback.

31. The neural synapse structure according to claim 29 or 30, wherein: The neural synapse structure includes a synaptic element that is not directly connected to the neuron (512), wherein the astrocyte calcium element is modulated using a presynaptic neuronal signal or a postsynaptic signal integrated using a dendritic receptor (526) from a synaptic element not directly connected to the neuron (512).

32. The neural synapse structure according to any one of claims 29 to 31, wherein: The axon (515) of the neuron (512) is connected to the astrocyte structure (530) such that the axonal backpropagation receptors of the astrocyte are configured to integrate the axonal signal of the neuron (512), The integrated axonal signal is used to modulate the astrocyte calcium element (522).

33. The neural synapse structure according to claim 32, wherein: The integrated axon signal is combined with other axonal back-propagating signals from other neurons included in the neural synapse structure (500) to form a combined axon signal, and wherein the combined axon signal is used to modulate the astrocyte calcium element (522), Preferably, the derivation unit (529) performs a temporal derivation on the combined axon input signal. More preferably, the derivation unit is a differentiator.

34. The neural synapse structure (500) according to any one of claims 24 to 33, wherein: The neural synapse structure (500) is also the neural synapse structure (200, 300, 400, 500) according to any one of claims 1 to 23.

Images