Systems and methods for reconfigurable modular neurosynaptic computational structures

Abstract

The present invention discloses a neurosynaptic structure for a spiking neural network, the neurosynaptic structure comprising one or more input ports, one or more synaptic elements, each synaptic element connected to at least one of the input ports and configured to receive an input signal and output a weighted post-synaptic signal. The neurosynaptic structure further comprises neurons connected to the one or more synaptic elements. The neurosynaptic structure is provided with different feedback and control structures, such as AMPA, GABA, and NMDA receptors, axons, dendrites, and neuronal backpropagation channels, and / or astroglial structures.

Description

[Technical Field]

[0001]

[0001] The present disclosure relates generally to automatic signal recognition techniques, and more particularly to systems and methods for reconfigurable modular neurosynaptic computational structures that enable increased flexibility and effectiveness of signal processing units and enable complex spatiotemporal and plastic behavior in networks of spiking neurons. [Background technology]

[0002]

[0002] The quest to minimize energy per inference or specific task (such as what are called energy proportional systems or constantly optimal circuits, i.e., systems with some form of dynamic feedback control or adaptability) requires co-design across the computational stack (so that algorithms and applications can influence the underlying hardware design, and similarly, the underlying hardware implementation can adapt to the needs or constraints of a particular application), significant hardware-software co-optimization (to minimize network size for a given degree of accuracy / maximize the number of operations within an allowable power envelope), and fully modular and flexible hardware systems to unify the design process.

[0003]

[0003] Smart sensor systems are not only adept at modifying or adjusting their internal processes or capabilities, but are also capable of modifying, warning, or refocusing / enabling predictions within the system (based on new information). Therefore, the inference machine associated with such a smart system continuously updates its knowledge of a set of internal and external parameters. To save space, time, and energy, new information is stored at synaptic sites, where it is processed and from which it can be efficiently recalled.

[0004]

[0004] In the case of short-term memory, short-term changes in input are relayed through lower-level synapses. Higher-level synapses encode information after multiple processing stages, so by strengthening synaptic capacity and allocating additional synaptic capacity, their memories become longer-lasting and more stably encoded.

[0005]

[0005] Reconfigurable neuromorphic networks typically comprise circuits that only partially contain receptors, dendrites, and subsequent synaptic features. A typical neurosynaptic array comprises a neural network matrix connecting n × n (or some division thereof) programmable synapses to n neurons. However, specifying the neurosynaptic computational architecture that defines and facilitates complex spatiotemporal spiking behavior and integrates plasticity involves both functional modification of neurosynaptic elements and structural mechanisms (circuit reconfiguration through synaptogenesis, i.e., formation, elimination, remodeling, and degeneration). These include a variety of receptors and dendritic channels (see, e.g., I. Segev, M. London, “Untangling dendrites with quantitative models,” Science, vol. 290, pp. 744-750, 2000), which alter synaptic responses through multiple forms of short- and long-term plasticity, intrinsic excitatory plasticity, multireceptor plasticity, and non-Hebbian plasticity, including homeostatic synaptic scaling and metaplasticity, as well as rapid structural (synaptic and dendritic) plasticity, as well as amplification, modulation, and dendritic structural scaling. Therefore, a generic synaptic architecture (as a simple point-processing unit) does not capture the diverse temporal dynamics of different types of receptors in biological synapses, which is essential 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 enabling a hierarchical, modular hardware platform that facilitates heterogeneity of computational elements and systems, but also allows for targeted reconfigurability, full programmability, and parameter adaptability. This modularity and heterogeneity means that systems or system definitions can be reused and parts of the system can be tuned / adjusted accordingly. Using compartmentalization, compounding, and reconfigurability, neurosynaptic resources are allocated to the signals they need to process; reliable signals require multiple synapses to conserve signal-to-noise ratios, while fewer synapses can be used when the signal-to-noise ratio is low, thereby avoiding wasted excess capacity. Furthermore, to maximize efficiency, signals are concentrated / sparsed in space and time (i.e., information is concentrated at all hierarchical scales); hence, highly active computational elements (e.g., neurons, synapses) are concentrated spatially in specific regions, while their activation is concentrated temporally. Signal processing is performed in the continuous-time (analog) domain to maximize information rate. Additionally, the hierarchical modular approach allows for the implementation of plasticity mechanisms at different time scales (e.g., short-term, long-term, homeostatic, structural plasticity) and then across different integral verticals. Consequently, compensation can be performed at different levels in both time and space, across different framework modularities, along different computational stacks, at the circuit, system, and architecture levels.

[0007]

[0007] This modular framework facilitates the benefits of (biological / spiking) signal processing at different levels of temporal granularity or hierarchy, allowing full software / hardware co-alignment with respect to distributed processing, neural network definition and mapping.

[0008] According to a first aspect, a neurosynaptic structure for a spiking neural network is disclosed, the neurosynaptic structure comprising a plurality of synaptic elements and neurons. Each synaptic element may comprise one or more input ports, where each input port is 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, where the signal path includes a weight element configured to apply a weight to generate the post-synaptic signal. At least a portion of the signal path of the synaptic element may include an AMPA receptor configured to integrate the input signal to make the post-synaptic signal more excitatory. At least a portion of the signal path of the synaptic element may include a GABBA receptor configured to integrate the input signal to make the post-synaptic signal more inhibitory. A neuron may comprise a soma and an axon, the soma may be configured to receive and combine one or more postsynaptic signals, and the axon may be configured to generate a spike output when the neuron's membrane potential reaches a predetermined threshold in response to the output of the soma.

[0009] According to an embodiment of the first aspect, at least one of the synaptic elements may comprise a signaling pathway that includes both an AMPA receptor and a GABBA receptor, and the at least one synaptic element may be configured to combine outputs from the AMPA receptor and the GABBA receptor to generate a postsynaptic signal.

[0010]

[0010] According to an embodiment of the first aspect, the AMPA receptors and / or GABBA receptors of the neurosynaptic element may be configurable to be activated or inactivated at a particular moment, and / or the AMPA receptors and / or GABBA receptors of the neurosynaptic element may be configurable to select a frequency range of the input signal within which the AMPA receptors and / or GABBA receptors are operable.

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

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

[0013]

[0013] According to an embodiment of the first aspect, the AMPA receptor comprises a voltage-driven integrator configured to generate an excitatory postsynaptic current to induce charge inflow from the neuron's membrane potential, and / or the GABBA receptor is a voltage-driven integrator configured to generate an inhibitory postsynaptic current to induce charge outflow from the neuron's 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 neuron's cell body, and / or the GABBA receptor can drive rapid Hebbian weakening of synaptic elements. The GABBA receptor preferably has a temporal dynamic behavior comparable to that of the AMPA receptor.

[0014] According to a second aspect, a neurosynaptic structure for a spiking neural network is disclosed. The neurosynaptic structure includes a plurality of synaptic elements and a neuron. Each synaptic element includes one or more input ports, where each input port is 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, where the signal path includes a weight element configured to apply a weight to generate the post-synaptic signal. The neuron may include a cell body and an axon, where the cell body is configured to receive and combine one or more of the post-synaptic signals, and the axon is configured to generate a spike output when the neuron's membrane potential reaches a predetermined threshold in response to the cell body output. The neurosynaptic structure may further include an axonal error backpropagation signal path connected to receive an axonal signal from the axon of the neuron. The axonal error backpropagation signal path may include an axonal error backpropagation receptor configured to integrate the axonal signal. The first subset of synaptic elements may be configured to receive the integrated axonal signal and to modulate weight elements of the first subset of synaptic elements based on the integrated axonal signal.

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

[0016]

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

[0017]

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

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

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

[0020] According to a third aspect, a neurosynaptic structure for a spiking neural network is disclosed. The neurosynaptic structure may include a plurality of synaptic elements and a neuron. Each synaptic element includes one or more input ports, where each input port is 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, where the signal path includes a weight element configured to apply a weight to generate the post-synaptic signal. The neuron may include a cell body and an axon, where the cell body may be configured to receive and combine one or more of the post-synaptic signals, and the axon may be configured to generate a spike output when the neuron's membrane potential reaches a predetermined threshold in response to the cell body output. The neurosynaptic structure may further include a dendrite backpropagation signal path configured to receive and combine the post-synaptic signals from one or more of the synaptic elements. The dendritic backpropagation signal path may include a dendritic receptor configured to integrate the combined post-synaptic signal. A second subset of synaptic elements may be configured to receive the integrated combined post-synaptic signal and to modulate weight elements of the second subset of synaptic elements based on the integrated combined post-synaptic signal.

[0021] According to an embodiment of the third aspect, each of the synaptic elements of the second subset of synaptic elements may comprise an NMDA receptor configured to integrate an input signal to form an integrated NMDA signal. A synaptic element of the second subset of synaptic elements may be configured to combine the integrated NMDA signal with the integrated combined postsynaptic signal and modulate a weight element of the synaptic element based on the combined signal.

[0022]

[0022] According to an embodiment of the third aspect, the integrated combined post-synaptic signal may be time-differentiated by a derivation unit, preferably a differentiator, before the differentiated integrated summed post-synaptic signal is used to modulate the weight elements.

[0023]

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

[0024] According to an embodiment of the third aspect, a 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 that integrates the combined postsynaptic signals of each of the one or more synaptic elements. Preferably, the dendritic receptor is a voltage-driven pure integrator, and more preferably, the dendritic receptor is formed using one or more gm-C filters.

[0025] According to an embodiment of the third aspect, the one or more postsynaptic signals may be combined at the cell body by combining the combined, preferably also integrated, postsynaptic signals of each of the synaptic elements 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 neurosynaptic structure is also the neurosynaptic structure of the first and / or second aspect.

[0027] According to a fourth aspect, a neurosynaptic structure for a spiking neural network is disclosed. The neurosynaptic structure may include a plurality of synaptic elements and a neuron. Each synaptic element may include one or more input ports, where each input port is configured to receive an input signal and 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, where the signal path includes a weight element configured to apply a weight to generate the post-synaptic signal. The neuron may include a cell body and an axon, where the cell body is configured to receive and combine one or more of the post-synaptic signals, and the axon is configured to generate a spike output when the neuron's membrane potential reaches a predetermined threshold in response to the cell body output. The neuron may include a neuron backpropagation signal path that may connect from the axon to the neuron's cell body and may include a local receptor configured to integrate the axon signal and a calcium element configured to establish an influence of the integrated axon signal and create a neuron feedback signal therefrom. The neuron backpropagation signal path may be configured to provide a neuron feedback signal to the soma of the neuron, where the neuron feedback signal may be combined with one or more postsynaptic signals.

[0028] According to an embodiment of the fourth aspect, the calcium elements may be amplifiers, preferably amplifiers that are faster than the weight elements of each synaptic element. Additionally or alternatively, the calcium elements 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 neurosynaptic structure may also be a neurosynaptic structure according to the first, second and / or third aspects.

[0031]

[0031] According to a fifth aspect, a neurosynaptic structure for a spiking neural network is disclosed. The neurosynaptic structure may include a plurality of synaptic elements and a neuron. Each synaptic element may include one or more input ports, where each input port is 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, where the signal path includes a weight element configured to apply a weight to generate the post-synaptic signal. The neuron may include a cell body and an axon, where the cell body is configured to receive and combine one or more of the post-synaptic signals, and the axon is configured to generate a spike output when the neuron's membrane potential reaches a predetermined threshold in response to the cell body output. The neurosynaptic 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 post-synaptic signals of each of the synaptic elements and may include a modulation circuit configured to modulate the summed post-synaptic signal into a modulated signal. The modulated signal may be used to change the behavior at one or more of the synaptic elements of the third subset or at the neuron.

[0032]

[0032] According to an embodiment of the fifth aspect, the astroglial structure may comprise an mGluR receptor configured to integrate a modulated signal. The astroglial structure may be configured to output an integrated modulated signal to each of the synaptic elements in the third subset. Each of the plurality of synaptic elements may be configured to combine the integrated modulated signal with an integrated input signal. The integrated input signal may be obtained from integrating the input signal using AMPA and / or GABBA receptors positioned along the signal pathway 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 astroglial structure may comprise an NMDA receptor configured to integrate the modulated signal, and the astroglial structure may be configured to output the integrated modulated signal to the neuron.

[0035] According to an embodiment of the fifth aspect, a neuron may include a neuron backpropagation signal path connecting an axon to a cell body of the neuron and including a calcium element configured to set an influence of an integrated axon signal to create a neuron feedback signal. The neuron backpropagation signal path may be configured to provide the neuron feedback signal to the cell body of the neuron. The integrated modulated signal may modulate the calcium element of the neuron to change the influence of the neuron's integrated axon signal.

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

[0037]

[0037] According to an embodiment of the fifth aspect, the IP3 receptor may 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 may be an amplifier, preferably an amplifier faster than the weight element of each synaptic element. Additionally or alternatively, the astrocyte calcium element may provide adaptive feedback.

[0038] According to an embodiment of the fifth aspect, the neurosynaptic structure may comprise synaptic elements that are not directly connected to neurons, and astrocyte calcium elements may be modulated using presynaptic neuronal signals or postsynaptic signals coming from synaptic elements that are not directly connected to neurons, which may be integrated using dendritic receptors.

[0039] According to an embodiment of the fifth aspect, the axons of neurons can be connected to astroglial structures, so that the axonal backpropagation receptors of the astrocytes can be configured to integrate the axonal signals of the neurons, and the integrated axonal signals can be used to modulate the astrocyte calcium elements.

[0040]

[0040] Preferably, the integrated axonal signal can be combined with other axonal error backpropagation signals from other neurons included in the neurosynaptic structure to form a combined axonal signal. The combined axonal signal can be used to modulate astrocyte calcium elements. More preferably, a time derivative of the combined axonal input signal is obtained by a derivation unit, and even more preferably, the derivation unit is a differentiator.

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

[0042]

[0042] According to a sixth aspect, a method for controlling a neurosynaptic structure for a spiking neural network is disclosed, the neurosynaptic structure comprising a plurality of synaptic elements and a neuron. Each synaptic element may comprise one or more input ports, where each input port is 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, where the signal path includes a weight element configured to apply a weight to generate the post-synaptic signal. The signal path of at least a portion of the synaptic elements may include a GABBA receptor configured to integrate the input signal to make the post-synaptic signal more inhibitory. The neuron may comprise a cell body and an axon, where the cell body may be configured to receive and combine one or more of the post-synaptic signals, and the axon may be configured to generate a spike output when the neuron's membrane potential reaches a predetermined threshold in response to the cell body output. The method may begin by receiving an input signal at a synaptic element. The input signal may then be integrated to make the post-synaptic signal more excitatory and / or inhibitory.

[0043] According to a seventh aspect, a method for controlling a neurosynaptic structure for a spiking neural network is disclosed. The neurosynaptic structure includes a plurality of synaptic elements and a neuron. Each synaptic element includes one or more input ports, where each input port is 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, where the signal path includes a weight element configured to apply a weight to generate the post-synaptic signal. The neuron may include a cell body and an axon, where the cell body is configured to receive and combine one or more of the post-synaptic signals, and the axon is configured to generate a spike output when the membrane potential of the neuron reaches a predetermined threshold in response to the cell body output. The method may include receiving an axon signal from an axon of the neuron. The received axon signal may then be integrated. The integrated axon signal may be received at a first subset of synaptic elements. The weight elements of the first subset of synaptic elements may then be modulated based on the integrated axon signal.

[0044] According to an eighth aspect, a method for controlling a neurosynaptic structure for a spiking neural network is disclosed. The neurosynaptic structure may include a plurality of synaptic elements and a neuron. Each synaptic element includes one or more input ports, where each input port is 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, where the signal path includes a weight element configured to apply a weight to generate the post-synaptic signal. The neuron may include a cell body and an axon, where the cell body may be configured to receive and combine one or more of the post-synaptic signals, and the axon may be configured to generate a spike output when the neuron's membrane potential reaches a predetermined threshold in response to the cell body output. The method may include receiving and combining the post-synaptic signals from one or more of the synaptic elements. The combined post-synaptic signal may then be integrated. The integrated combined postsynaptic signal may be received at a second subset of synaptic elements, and the second subset of synaptic elements may modulate weight elements of the second subset of synaptic elements based on the integrated combined postsynaptic signal.

[0045] According to a ninth aspect, a method for controlling a neurosynaptic structure for a spiking neural network is disclosed. The neurosynaptic structure may include a plurality of synaptic elements and a neuron. Each synaptic element may include one or more input ports, where each input port is 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, where the signal path includes a weight element configured to apply a weight to generate the post-synaptic signal. The neuron may include a cell body and an axon, where the cell body is configured to receive and combine one or more of the post-synaptic signals, and the axon is configured to generate a spike output when the neuron's membrane potential reaches a predetermined threshold in response to the cell body output. The method may include integrating the axon signal. Further, an influence of the integrated axon signal may be established, from which a neuron feedback signal may be created. The neuron feedback signal may then be provided to the neuron's cell body, where the neuron feedback signal may be combined with one or more post-synaptic signals.

[0046] According to a tenth aspect, a method for controlling a neurosynaptic structure for a spiking neural network is disclosed. The neurosynaptic structure may include a plurality of synaptic elements and a neuron. Each synaptic element may include one or more input ports, where each input port is 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, where the signal path includes a weight element configured to apply a weight to generate the post-synaptic signal. The neuron may include a cell body and an axon, where the cell body is configured to receive and combine one or more of the post-synaptic signals, and the axon is configured to generate a spike output when the membrane potential of the neuron reaches a predetermined threshold in response to the cell body output. The method may include receiving and summing the post-synaptic signals of each of the synaptic elements. The summed post-synaptic signal may then be modulated into a modulated signal. Behavior at one or more of a third subset of synaptic elements or at the neuron is then changed using the modulated signal.

[0047]

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

[0048] [Figure 1A]

[0048] A typical implementation of a neurosynaptic array is shown. [Figure 1B] The corresponding diagram of a synaptic train with nominal (AMPA-like) synaptic integration, weight multiplication, and neuronal summation is shown. [Figure 2]

[0049] A diagram of a synaptic train with AMPA synaptic integration, weight multiplication, and neuronal summation, but also including GABA (inhibitory) receptors, is shown. [Figure 3]

[0050] Shown is a diagram of a synaptic train with AMPA synaptic integration, weight multiplication, and neuronal summation, but also including GABA (inhibitory) receptors, and further extending this diagram with NMDA receptors and the corresponding axonal (backpropagated) feedback system. [Figure 4]

[0051] FIG. 1 shows a schematic diagram of a reconfigurable modular network comprising neuromorphic synapses with multiple receptors and dendritic and axonal feedback signals, respectively. [Figure 5]

[0052] FIG. 1 shows a schematic diagram of a reconfigurable modular network comprising neuromorphic synapses with multiple receptors, fitness index neurons, and an astrocyte system. [Figure 6A]

[0053] Cooperative events in source populations are shown to transmit feedforward a) excitation, and a shift to higher excitation sets neurons to fast-firing behavior with sparse bursts. For each synchrony category, a population of postsynaptic neurons is assigned, and the synchrony category is mapped to a pattern of postsynaptic activity. [Figure 6B] Cooperative events in the source population are shown to transmit feedforward b) inhibition, and a shift to higher inhibition sets neurons to slow-firing behavior with sparse bursts. For each synchrony category, a population of postsynaptic neurons is assigned, and the synchrony category is mapped to a pattern of postsynaptic activity. DETAILED DESCRIPTION OF THE INVENTION

[0049]

[0054] Certain specific embodiments are described in further detail below, however, it should be recognized that these embodiments should not be construed as limiting the scope of protection for the present disclosure.

[0050]

[0055] FIG. 1A shows a typical implementation of a neurosynaptic array 100 of a spiking neural network (SNN).

[0051]

[0056] In this embodiment, the neurosynapse array includes m×n synapses 102, each with a synaptic weight w i,j is a function of dividing m input signals 101 (e.g., digital or analog input pulses, preferably voltage spikes) into N k The spiking neural network connects to n neurons 103, denoted by . Each neuron may integrate an incoming signal and fire a spike signal 105, forming a spatiotemporal spike train, when a certain membrane potential threshold is reached. In this example, an input signal 101 may be fed to a row of synapses 102, and each row of synapses may be fed to a particular neuron 103. In general, an input signal 101 may be fed to one or more neurons 103 through one or more synapses 102. The synaptic arrangement and its connections may be dynamically adaptable to optimize the performance and power consumption of the spiking neural network.

[0052]

[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 training, the synaptic weights can be adjusted to optimize the network's performance for a particular task. Neurons 103 integrate incoming signals, for example, from input connections or other connected neurons via synapses 102. The membrane potential of a neuron is determined by the synaptic weight w i,j The neural network is updated based on incoming spikes from connected neurons, weighted by . Neurons 103 fire spikes when their membrane potential exceeds a certain threshold.

[0053]

[0058] Neurons 103 can generate complex spatiotemporal spike patterns that contribute to the network's processing power. SNNs can adapt to and learn from data through synaptic plasticity mechanisms. This is achieved by adjusting the synaptic weights w i,jThis means that σ can change over time based on the pattern of spikes and their timing. Learning rules are used to adjust the weights to improve the performance of the network for a given task.

[0054]

[0059] Typically, neurosynaptic structures implemented in, for example, spiking neural network cores use a mixed analog-digital computing platform; that is, spike trains can incorporate analog information in the timing of events, which can then be converted back to analog representation at the input of the synaptic matrix. Each mixed analog / digital core can include an input decoder connecting m × n programmable synapses to n neurons and an I / O network communication layer. Neurosynaptic computing elements can generate complex spatiotemporal dynamics that can be scaled toward specific characteristics that can support target signal processing functions. Neuronal spiking characteristics can be controlled through specific parameter sets. The topology of the neurosynaptic elements controls the regularity of spontaneous neuronal spiking, e.g., firing rate scaling factor and the strength of intrinsic noise, resulting in coherence / resonance generation.

[0055]

[0060] A core router (connected to the rest of the network via I / O ports) of a spiking neural network core may provide input spikes to a row decoder. The row decoder may take the input spikes and determine which row of the synaptic array a particular portion of the signal is to be propagated to. There may be a learning and adaptation module. Additionally, there may be a column decoder that takes the output of a column of synaptic elements and sends the decoded column output to one or more neurons. The neuron configuration may be changed by a neuron control module, which may also receive feedback from the learning and adaptation module. Spikes from one or more neurons may be sent to other cores of the spiking neural network via the core router.

[0056]

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

[0057]

[0062] Figure 1B shows a diagram corresponding to Figure 1A of a typical synaptic train with nominal (AMPA-like) synaptic integration, weight multiplication, and neuronal summation. Figure 1B represents a specific implementation of a single train of synapses 102 connecting to a particular neuron 103 of the synaptic arrangement of Figure 1A.

[0058]

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

[0059]

[0064] The AMPA receptor 111 can be a component that mimics the behavior of biological AMPA receptors at synapses. It can be used to capture the way excitatory synaptic transmission occurs in the brain, integrating the input signal 101 over time and simulating the buildup of excitatory current as in biological synapses.

[0060]

[0065] Each synapse 102 may have 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 an AMPA receptor, which then assigns the weight w of the synapse 102 i,j can be applied to the spike. Conversely, when a spike arrives at a synapse as an input signal (e.g., from a presynaptic neuron or from an encoder), the weight w of the synapse 102 i,j can be applied to the spike and the resulting contribution integrated by the "AMPA receptor" unit, i.e. the order of integration and weight application can be reversed.

[0061]

[0066] The cell body 113 of the neuron 103 may serve as a location where incoming signals are integrated. These incoming signals are typically spikes (discrete events) that arrive from the neuron's input synapses 102. This changes the neuron's membrane potential, and when a certain threshold is reached, the neuron 103 may fire, specifically, from its axon 115. When the artificial neuron 103 generates an output spike, it is further transmitted along the neuron's axon 115, for example, to other neurons to which the current neuron is connected. This spike transmission mimics the propagation of an action potential along a biological axon. Furthermore, a signal restoration unit 114 may be present in the neuron, which may restore the signal after signal attenuation has occurred. The signal restoration unit 114 may comprise, for example, a digital buffer.

[0062]

[0067] Without loss of generality, to compute most efficiently in space and energy, i.e., to minimize the energy per bit of information, the present invention allows computation in a modular, multi-layered manner across different hierarchical granularities, where each hierarchical level is optimized or selected based on the lowest associated cost. Thus, in this multi-layered structure, each set of basic signal processing components is modularly combined; for example, synaptic integration is tuned through gain-bandwidth product characteristics, the subsequent weight application / retention mechanism is governed by digital code-current transfer (i.e., digital (signal)-analog (current) conversion), and the final spike generation is derived through a current-frequency transfer characteristic controlled, among other things, through an adaptation mechanism to match the current to the conditions. The route to this efficiency is followed in the integration system across different hierarchical levels while preserving information, embodied as signal-to-noise ratio (SNR) and bandwidth. The structural implications are valid across different levels of the hierarchy, at finer scales and lower levels, as well as at larger scales and higher levels: diverse circuits allow the inference machine to send only relevant information at lower information rates, providing a competitive computational performance / associated cost curve.

[0063]

[0068] The multilayer neurosynaptic structure allows for differential (e.g., fast, slow, depressive, potentiating) responses to input stimuli governed by several mechanisms, such as noncompetitive homosynaptic plasticity driven by residual activity, competitive heterosynaptically driven neighbor-recognition processes, or mediating potentiation and homeostatic adjustment in response to activating inputs during normal sensory use. The model allows for the assignment of various neuronal properties, e.g., axonal and dendritic delays, synaptic transfer functions, provides for the calculation of optimal input-output transfer functions, and allows for specific signal processing functions, such as filtering, amplification, multiplication, summation, etc., to convolve spike train kernels in the time domain.

[0064]

[0069] In an integrated neuromorphic system, the sensed signal, once encoded, propagates passively (down the axon) to the synaptic terminal. The physical distance on the chip is short enough for passive signal transmission (i.e., does not degrade the signal sufficiently (e.g., exponential decay)).

[0065]

number

[0066] If the inference machine's analog signal contains more information than an action potential can encode (dominated by σ and hence no buffering / signal restoration is required), then the inference machine can read out the encoded information directly (analog, and therefore very efficient). If the inference machine's analog signal contains more information than an 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 structures at all spatial scales. Note that passive signals spread in space and time only as the square root of the dendrite / synaptic column wire diameter.

[0067]

[0070] The present invention seeks to solve this problem by proposing a more adaptable synapse structure. The proposed structure is built on the basis of the embodiment shown in Figures 1A and 1B. Similar features are not repeated and can be easily combined and discussed in relation to these embodiments.

[0068]

[0071] FIG. 2 shows a diagram of a synaptic train with AMPA synaptic integration, weight multiplication, and neuronal summation, but also including GABA (inhibitory) receptors.

[0069]

[0072] The synaptic structure 200 includes synapses 202 (also called synaptic elements) that include AMPA receptors 203 and / or GABA receptors 204. Each receptor can exhibit different temporal dynamics in response to an input signal. For example, each receptor can function as a filter for different frequencies.

[0070]

[0073] Receptors induce excitatory and inhibitory postsynaptic currents (EPSCs and IPSCs), respectively, eliciting charge inflow and outflow from the membrane voltage. AMPA receptors mediate fast (glutamatergic) synaptic currents to drive the soma of the postsynaptic neuron. AMPA receptors have a high response rate to transmitter binding, both in the upstroke and decay phases, due to both fast transmitter clearance and rapid channel closure. 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.

[0071]

[0074] GABA receptors are a major contributor to fast IPSCs (see A. Destexhe, ZF Mainen, TJ Sejnowski, “Kinetic models of synaptic transmission,” Meth. Neuron. Model., vol. 2, pp. 1–25, 1998). Conversely, deprivation reduces the magnitude of inhibitory postsynaptic currents to principal neurons, preferentially reducing the activation of fast-spiking interneurons. GABA receptor-mediated inhibition is a prerequisite for balancing excitation and inhibition, thereby enabling fine-grained temporal fluctuations to stabilize neural networks and drive precise plastic behavior (see SH Wu, CL Ma, JB Kelly, “Contribution of AMPA, NMDA, and GABA receptors to the temporal pattern of postsynaptic responses in the inferior colliculus of the rat,” J. Neurosc., vol. 24, no. 19, pp. 4625–4634, 2004).

[0072]

[0075] Similarly, in the homeostatic model, deprivation of a subset of inputs through a GABAergic circuit drives rapid Hebbian weakening along the deprived pathway and a slower homeostatic increase in overall synaptic strength and / or intrinsic excitability, which increases the response to spared inputs. GABA receptors preferably have temporal dynamics comparable to AMPA receptors.

[0073]

[0076] The synaptic element 202 is connected to at least one of the input ports 201 and is configured to receive an input signal and output a post-synaptic signal. The synaptic element 202 comprises an AMPA receptor 203 configured to integrate the input signal to make the post-synaptic signal more excitatory. Additionally or alternatively, the synaptic element 202 can comprise a GABA receptor 204 configured to integrate the input signal to make the post-synaptic signal more inhibitory. Both AMPA and GABA receptors can be present, but which are active and / or which frequency ranges they filter, i.e., in which frequency ranges these receptors are active, can be dynamically determined. The outputs of the AMPA and / or GABA receptors are combined (e.g., by summation, if necessary).

[0074]

[0077] Weight elements 205 are also present in synapses 202. Weight elements 205 are configured to set the influence of postsynaptic signals by having a particular weight that determines the strength of the connection represented by synapse 202. For example, if the weight is higher, weight elements 205 may function to amplify signals coming from AMPA receptors and / or GABA receptors. If the weight is lower, weight elements 205 may function to attenuate signals coming from AMPA receptors and / or GABA receptors. The weight, and therefore the influence, of a weight element may be dynamically adjustable (e.g., by modulating weight elements 205) or may be static.

[0075]

[0078] Neuron 206 may be the same as the neuron of Figure 1B. That is, neuron 206 may comprise a cell body 207 and an axon 209, which may fire when the membrane potential of neuron 206 reaches a certain threshold. The neuron may thus be configured to receive a postsynaptic signal from at least one synaptic element 202. Again, it should be noted that the synaptic element 202 connected to neuron 206 represents a synaptic element in the column of the synaptic array of Figure 1A connected to a particular neuron, although other configurations are of course possible.

[0076]

[0079] Figure 3 shows a diagram of a synaptic train with AMPA synaptic integration, weight multiplication, and neuronal summation, but also including GABA (inhibitory) receptors, and further extending this diagram with NMDA receptors and corresponding axonal (backpropagated) feedback systems.

[0077]

[0080] Without loss of generality, all major glutamate receptors, AMPA and NMDA, as well as GABAa receptors, are included, each exhibiting different temporal dynamics in their response to neurotransmitters. Receptors elicit excitatory and inhibitory postsynaptic currents (EPSCs and IPSCs), respectively, eliciting charge inflow and outflow from membrane voltage. Without loss of generality, scaling of excitatory synapses onto principal neurons is primarily expressed by modulating AMPA receptor insertion through NMDA potentiation and inhibition.

[0078]

[0081] NMDA receptors provide a more complex mechanism, including activity-dependent modification of synaptic weights (i.e., NMDA receptor-dependent inhibition, which suppresses the process mediating synaptic AMPA receptor currents), and NMDA potentiation, which triggers the appearance of AMPA receptor currents at weak synapses, thus functionalizing these synapses. Intrinsic excitability and excitatory-inhibitory balance are also homeostatically regulated by establishing a permissive excitatory-inhibitory balance or by editing firing patterns to promote excitatory synaptic plasticity.

[0079]

[0082] The neurosynaptic structure 300 comprises an axonal error backpropagation signal path 311 connected to a neuron 307 and at least one synaptic element 302, the axonal error backpropagation signal path 311 configured to receive an axonal signal from an axon 308 of the neuron 307 and comprising an axonal error backpropagation receptor 312 configured to integrate the axonal signal. The axonal error backpropagation signal path 311 is configured to output the integrated axonal signal to the at least one synaptic element 302. The axonal error backpropagation receptor 312 may be a pure integrator element that may be voltage-driven.

[0080]

[0083] At least one synaptic element includes an NMDA receptor 305 configured to integrate an input signal received at input port 301 into an integrated NMDA signal. The integrated NMDA signal and the integrated axonal signal are combined and modulate a weight element 306 of at least one synaptic element 302 to change the influence of the postsynaptic signal of the 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 while the integrated axonal signal can be used to combine with the integrated NMDA signal, it is preferable to use the time derivative of the integrated axonal signal. A derivation unit 313 can therefore be located along the axonal error backpropagation signal path 311, for example, within the synaptic element 302. The derivation unit 313 can be, for example, a differentiator, which is a circuit designed to generate an output that is approximately proportional to the rate of change (time derivative) of the input. The differentiator can essentially be a high-pass filter. An active differentiator may include some form of amplifier, while an inactive differentiator may be made solely from resistors, capacitors, and / or inductors.

[0081]

[0084] The synaptic elements 302 and the rest of the neuron 307 may be similar to those in Figure 2. That is, there may also be AMPA 303 and / or GABA 304 receptors that integrate the input signal, and their single or combined outputs are sent to the weight element 306. The neuron receives, at its cell body 309, weighted and integrated postsynaptic signals from each synapse 302 connected to the neuron 307, where different postsynaptic signals are combined, for example, by summation. A signal restoration unit 310 may be present in the neuron 307. When the membrane potential of the neuron 307 reaches a certain threshold, the neuron may spike via its axon 308.

[0082]

[0085] Note that while all synaptic elements 302 are shown with AMPA, GABA, and NMDA receptors and axonal backpropagation signal pathway 311, this is not required. Some of the synaptic elements connecting to neuron 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 NMDA receptors and weight elements modulated by axonal backpropagation signal pathway 311 (preferably with derivation unit 313).

[0083]

[0086] Note that the weight elements 306 may also be modulated based solely on the NMDA receptor output, or solely on the output of the axonal error backpropagation signal path 311 (again, preferably with the derivation unit 313).

[0084]

[0087] FIG. 4 shows a schematic diagram of a reconfigurable modular neurosynaptic structure 400 comprising neuromorphic synapses with multiple receptors and dendritic and axonal feedback signals, respectively.

[0085]

[0088] The neurosynaptic structure 400 may further include a dendritic error backpropagation signal path 408 connected to one or more synaptic elements 402, the dendritic error backpropagation signal path 408 configured to receive and sum the post-synaptic signals of each of the synaptic elements 402. Preferably, the dendritic error backpropagation signal path 408 includes a signal reconstruction unit 407, which may have a function similar to that of the signal reconstruction unit 114. The dendritic error backpropagation signal path 408 may include a dendritic receptor 409 configured to integrate the summed post-synaptic signal. The dendritic error backpropagation signal path 408 is configured to output the resulting integrated summed post-synaptic signal, also referred to as a dendritic signal, to each of the one or more synaptic elements 402 to which it is connected.

[0086]

[0089] Each of these one or more synaptic elements 402 may include an NMDA receptor 405 configured to integrate an input signal received at an input port 401 to the synaptic element into an integrated NMDA signal. The integrated NMDA signal and the integrated dendritic signal may be combined, and the weight element 406 of the synaptic element may be modulated to change the influence of the postsynaptic signal of at least one synaptic element 402, e.g., on a neuron 414. Note that while the integrated dendritic signal may be used to combine with the integrated NMDA signal, it is preferable to use the time derivative of the integrated dendritic signal. A derivation unit 410 (similar to that described above) may therefore be located along the dendritic error backpropagation signal path 408, e.g., within the synaptic element 402. Preferably, if an axonal error backpropagation signal path 416 is also present, as in the illustrative embodiment, the dendritic and axonal signals are combined (e.g., by signal summation), and then the derivation unit is used to obtain the derivative of the combined signal. Of course, the derivatives of both signals can also be determined separately and then combined.

[0087]

[0090] In one embodiment, all synapses connected to a particular neuron are also connected to the same dendritic backpropagation signal pathway 408. In that case, the summed postsynaptic signal of each of the synaptic elements may be forwarded to the neuron's soma 413. Alternatively, separate postsynaptic receptors 411 integrate the summed postsynaptic signal before it is sent to the neuron's soma 413.

[0088]

[0091] In another embodiment, dendrites connect only to a specific subset of synaptic elements 402 that connect to neurons 414. In that case, other synaptic outputs may be received at the cell body 413 via other synaptic segment outputs 412. These segments may also have dendritic structures that exist with dendritic backpropagation signal paths 408. Each dendritic structure may have a distinct post-dendritic receptor 411. This is the illustrative embodiment shown in the figure.

[0089]

[0092] The functionality of the other structures is similar to the embodiment of Fig. 3. That is, AMPA 403 and / or GABA 404 receptors may also be present that integrate the input signal, and their single or combined outputs are sent to the weight element 406. The neuron receives a weighted integrated post-synaptic signal at the cell body 413, where different post-synaptic signals are combined, for example, by summation. A signal restoration unit 407 may be present in the neuron 414. When the membrane potential of the neuron 414 reaches a certain threshold, the neuron may spike via its axon 415. An axonal error backpropagation signal path 416 may also be present.

[0090]

[0093] Similarly, as previously mentioned, the different feedback paths can be implemented separately or together. Note that the weight element 406 can also be modulated based solely on the NMDA receptor output, or solely on the output of the dendritic backpropagation signal path 408 (again, preferably with the derivation unit 409). This time is also optionally combined with the axonal backpropagation signal path 416.

[0091]

[0094] Due to the efficient energy-delay product over short distances, the signal is converted into a current that charges the membrane capacitance. Over longer distances, the current is reproduced by appropriately clustered voltage-gated NMDA channels. At the postsynaptic site, the postsynaptic backpropagation spike that initiates the dendritic spike—the temporal summation of the dendritic spike and the soma spike—is then completed. This implementation allows for a wide range of temporal behavior, since each receptor can operate on a different time scale. It also increases the network's ability to perform pattern recognition even in noisy signals, resulting in higher noise tolerance.

[0092]

[0095] If a group (burst) of dendritic spikes is strong enough to drive the cell body, the neuron can generate an action potential, and the resulting spike can then be backpropagated to the dendrites. See R.C. Froemke, Y. Dan, "Spike-timing-dependent synaptic modification induced by natural spike trains," Nature, vol. 416, pp. 434-438, 2002. The backpropagated dendritic and cell body signals can be multiplied and added to the NMDA receptor signal to form a weighted control signal.

[0093]

[0096] A single receptor is subject to stochastic fluctuations such as thermal noise, and 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 improves the SNR by √n. However, it comes with an associated higher cost and is therefore reserved for special purposes, e.g., synapses transmitting with high temporal precision. For optimal allocation of resources, the noise-resource equation can be applied, i.e., trading power and / or area to reduce thermal / flicker noise. If noise is to be minimized, it is important to consider the parts of the system that affect all other parts of the system (initial stages), and the parts of complex systems (i.e., high n i More resources should be allocated to the region with a larger number of receptors (denoted by ).

[0094]

[0097] In general, in biochemical systems, the origins of noise can be traced to different hierarchical scales: i) biophysical origins (microscopic fluctuations in small structural components of neurons, such as ion channels or synaptic mechanisms involved in vesicle release); ii) neural circuit dynamics (excitatory / inhibitory behavior, sparsity); and iii) global network properties (average instantaneous population activity, synchrony). Similarly, variability effects related to modern IC manufacturability can be subdivided into different levels of spatial granularity, ranging across the physical dimensions of the chip, from global effects at the IC level, including technology (process corners), electrical (supply voltage), thermal (temperature), and lithography (linewidth), to local effects at the subcircuit (substrate noise, temperature gradients, layer density), transistor (well proximity, local heating, jitter, crosstalk, aging effects), or atomic (thermal noise, mobility fluctuations, random dopants, mechanical stress, line-edge roughness) level.

[0095]

[0098] The structure shown in Figures 2-5 allows for the implementation of this hierarchical granularity.

[0096]

[0099] Variability effects, when subdivided across different temporal granularities, can therefore be relevant to methodologies for mitigating performance or yield loss, i.e., can be subdivided into static, slow-, medium-, and fast-varying events based on the time dependency of the variability artifacts. Static effects allow for one-time correction (e.g., trimming) of the relevant devices in the circuit, while slow time-dependent effects, i.e., static effects within the observation period, can be corrected during operation using on-chip calibration mechanisms. Variations with time spans comparable to the maximum speed of the process (e.g., jitter, substrate noise, kT noise, dynamic IR drop) cannot be directly calibrated and must be accommodated by extending the design margin. Equivalently, plasticity mechanisms occur at different time scales and across different integration verticals, e.g., short-term plasticity (adaptive neuronal behavior, short-term synaptic adaptation), long-term plasticity (spike-based plasticity rules), homeostatic plasticity (scaling up of excitatory synaptic strength, neurosynaptic element-specific changes in inhibitory circuits, and intrinsic excitability changes depending on the precise deprivation / depression definition / paradigm, all to stabilize the neuronal firing frequency range and hence network activity), and structural plasticity (modifying network connectivity). Thus, compensation can be performed at different levels in both time and space, across different framework modularities, along different computational stacks, at the circuit, system, and architectural levels, and across hardware and software lines.

[0097]

[0100] Synaptic weights induced by Hebbian STDP coactivation will deteriorate unless consolidated. Initial synaptic plasticity values may set tags at synapses and define markers for future consolidation (i.e., transmission of information through the write process) when changes in synaptic efficacy occur. An intrinsic plasticity rule based on the information maximization principle is given as a non-limiting example below:

[0098]

number

[0099]

[0101] Weight w ij , and tags η connected to their nearest neighbors via a time-dependent gating variable as given by the above equation. ij .

[0100]

number

[0101] is the external input, χ is the homeostatic scaling factor,

[0102]

number

[0103] are independent Gaussian white noise processes, and α ηw is a coupling parameter term that determines the strength of the interaction between variables, and τ w is the relevant time constant, and g serves as the gating variable. Furthermore, f is a (possibly discontinuous) function that converts the net current reaching the population into population activity and specifies the firing frequency (in Hz). The regulation time constant determines the direction in conductance space, while the homeostatic scale factor controls the trajectory before reaching equilibrium. Here, the assumption is that a node can only excite its own inhibitory population. Neuronal membrane potentials create a permissive time window for the induction of sensory context-dependent bidirectional plasticity mechanisms. This plasticity regulation, targeting different classes of excitatory and inhibitory synapses, coincides with an overall homeostatic shift in the balance between excitation and inhibition.

[0104]

[0102] Synaptic scaling is mediated by multiple mechanisms that vary across different hierarchical levels depending on neurosynaptic type, time course, sequence domain, and structural stage. Receptor types in synaptic implementations are more diverse than the neurons to which they connect.

[0105]

[0103] For inference machines that are more limited in the number of neurosynapses, i.e., in systems where each uniformly defined receptor that forms a synapse does not have a dedicated neuron associated with the synaptic train, neurons may be shared with synaptic clusters incorporating sets of receptors that result in the same final action. As a result, input signals provided to the neuromorphic array are assembled for action not by high-level circuitry, but by dedicated neurons that can provide complex functionality to allow for savings in neuron count.

[0106]

[0104] Morphological features can be considered as specific control parameters that significantly contribute to both neurosynaptic element dynamics and network activity patterns. Without loss of generality, synaptic clusters, forming in terms of size and function, can be defined according to the proposed 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 complexities, target SNRs, time constants, etc.

[0107]

[0105] This offers a paradigm shift in how to provide sufficient resources for complex inference tasks within a feasible power-performance-area design space by increasing the number of neurosynapses in simplified (or specialized) computational elements, or by increasing the computational power of (alternatively complex multifunction) elements within a reduced number of neurosynapses. Within these complex multifunction elements, separate parts for each task can be defined, i.e., synapses contain multiple receptors, where each receptor can be independently tuned for speed and sensitivity, independently adjusted, providing more opportunities for further improvement and redundancy for fault protection.

[0108] Without loss of generality, complexity can include allowing for neurosynaptic elements (groups) that nominally have similar functionality but are enabled to support multiple pathways and have subtly different properties; for example, axons and astrocytes both express activation / refractory capabilities that are facilitated through branching mechanisms. Nominally, communication and information transfer within neural networks has been the sole domain of pre- and postsynaptic connections between neurons. However, connections between astrocytes and neurons provide additional pathways for intercellular communication, and communication from neurons to astrocytes is facilitated by additional synaptic compounds released upon the arrival of a presynaptic action potential.

[0109]

[0107] Figure 5 shows a schematic diagram of a reconfigurable modular network comprising neuromorphic synapses with multiple receptors, fitness index neurons, and an astrocyte system.

[0110]

[0108] This embodiment shows the backpropagation signal pathways of AMPA 503 and / or GABA 504 receptors, NMDA receptors 505, axons 516, and dendrites 508. Their functions are the same as those introduced above, and they can function alone or in different combinations as described above for each synapse 502 connected to the input port 501.

[0111]

[0109] In this embodiment, two new structures are introduced: neuronal backpropagation signal pathway 520 and astrocyte cell / structure 530.

[0112]

[0110] The neuron 512 may thus include a neuronal backpropagation signal path 520 connecting the axon 515 of the neuron 512 to the neuronal soma. In the case of neural frequency / firing adaptation, when the spiking frequency is sufficiently high, the astrocytes will use their internal memory to reach a steady state, resulting in a persistent neural spiking frequency. The neuronal backpropagation signal path 520 may include a local receptor 518 configured to integrate the axonal signal. The local receptor 518 may be a pure integrator element and may be voltage-driven. It may function as a filter for a specific frequency range of the axonal signal. Furthermore, the neuronal backpropagation signal path 520 may include a calcium element 519 configured to set the influence of the integrated axonal signal and create a neuronal feedback signal therefrom. The calcium element 519 may be a high-speed amplifier and preferably may provide adaptive feedback. The neuronal backpropagation signal path 520 is configured to provide the neuronal feedback signal to the neuronal soma 514 of the neuron 512.

[0113]

[0111] In the cell body 514, the neuron feedback signal may be combined with signals coming from the dendrites of a group of synapses 502 (which may be all synapses connected to the neuron 512, or a subset), and / or signals coming from other segments 513 (which may have dendritic structures or may not be between the synapse 502 and the neuron 512), as in the embodiment of Figure 4.

[0114]

[0112] The neurosynaptic structure 500 may additionally or alternatively comprise an astrocyte structure 530 connected to one or more of the plurality of synaptic elements 502. The astrocyte structure 530 may be configured to receive and sum the postsynaptic signals of each of the plurality of synaptic elements and may comprise a modulation circuit configured to modulate the summed postsynaptic signal into a modulated signal.

[0115]

[0113] The astroglial structure 530 may comprise mGluR receptors 523 configured to integrate the modulated signal. The astroglial structure may then be configured to output the integrated modulated signal to each (or a subset thereof) of the synaptic elements 502. For example, each (or a subset thereof) of the synaptic elements 502 in which the postsynaptic signal was used as a modulated input to the astroglial structure may be configured to sum the integrated modulated signal received from the astroglial structure with the input signal integrated by the AMPA 503 and / or GABA 504 receptors.

[0116]

[0114] In addition to or as an alternative to this feedback from the astrocyte structure to the synaptic element, the astrocyte structure may comprise an NMDA receptor 524 configured to integrate the modulated signal coming from the modulation circuit. The NMDA receptor 524 may be similar to the NMDA receptor 505 but may be optimized differently, e.g., functioning over a different frequency domain. The astrocyte structure 530 may be configured to output an integrated modulated signal from the NMDA receptor 524 to the neuron 512 if the neuronal backpropagation signal pathway 520 is present. The integrated modulated signal coming from the NMDA receptor 524 may modulate the neuron's calcium elements to change the influence of the neuron's integrated axon signal coming from the local receptor 518 on the neuron's cell body 514.

[0117]

[0115] The modulation circuit of the astrocyte structure 530 may comprise an IP3 receptor 521 configured to integrate the summed postsynaptic signal. The IP3 receptor 521 may be a voltage-driven pure integrator. Furthermore, the modulation circuit may comprise an astrocyte calcium element 522, which may be a high-speed amplifier, preferably providing adaptive feedback. The astrocyte calcium element 522 may be configured to set the influence of the modulated signal. The IP3 receptor 521 and the astrocyte calcium element 522 may generate the modulated signal. Optionally, one or more signal restoration units 507 are present in the modulation circuit, which have the same function as the signal restoration units described above.

[0118]

[0116] The astrocyte calcium element 522 may be adaptable in that it may be modulated. By modulating the astrocyte calcium element 522, the effect of the modulated signal on the structure to which the astrocyte structure 530 provides feedback, such as a synaptic element 502 or neuron 512, may be changed.

[0119]

[0117] Presynaptic neuronal signals or postsynaptic signals coming from synaptic elements not directly connected to neuron 512 may be received at input 525 and integrated using dendritic receptor 526. Dendritic receptor 526 may be similar to dendritic receptor 509 or 511, but may also operate in a different frequency range or have other differences with respect to its operating parameters.

[0120]

[0118] The axonal error backpropagation signal may be integrated in an axonal error backpropagation receptor 527, which is similar in its operation to the axonal error backpropagation receptor 517, but here again may operate in a different frequency range or have other differences in its operating parameters. The integrated axonal error backpropagation signal may be combined with other axonal error backpropagation signals from post-synaptic neurons at input 528, or only axonal error backpropagation signals from other post-synaptic neurons may be used as input. Preferably, the time derivative of this (combined) axonal input signal is obtained by a derivation unit 529, which works similarly to the integration unit 510.

[0121]

[0119] The input from the dendritic receptors 526 and the axonal feedback can be combined, for example by a multiplication unit (if both are present, although they can also be used alone), and the resulting signal can be provided to the astrocyte calcium element 522 to modulate the astrocyte calcium element. As mentioned, this can change the effect of the modulated signal on the structure to which the astrocyte structure 530 provides feedback, such as the synaptic element 502 or neuron 512.

[0122]

[0120] Astrocyte cells used within spiking neurons in spiking astrocyte-neuron networks can perform distributed fine-grained compensation and / or self-repair; for example, if synaptic operation is interrupted, retrograde control mechanisms effectively resume control through extensive plasticity processes.

[0123]

[0121] Although astrocytes do not directly induce the propagation of action potentials as neurons do, they can communicate bidirectionally with neurons and other astrocytes. At synapses, which comprise the connections between astrocytes and several neurons, i.e., synapses exchange signals at three terminals, and when an action potential arrives, stored electrical charge is released across the synapse, causing depolarization of the postsynaptic neuron.

[0124]

[0122] Activation of astrocyte receptors results in a transient increase in intracellular astrocyte calcium levels, which represents the fundamental mode of excitation in astrocytes, and astrocyte-neuron signaling, i.e., a slow inward current that operates with much longer rise and decay times compared to typical excitatory postsynaptic currents.

[0125]

[0123] When the postsynaptic neuron is sufficiently depolarized, voltage-gated calcium channels on the dendrites allow calcium influx into the dendrite, which also enables a control mechanism to regulate back-propagated signals from the postsynaptic terminal to the presynaptic terminal. Synaptically driven calcium transients in astrocytes modulate synaptic transmission via the release of astroglial glutamate (mGluR) binding to presynaptic receptors. This then initiates the production and release of IP3 within the astrocyte, triggering a transient intracellular release of calcium and subsequent activation of NMDA receptors on the postsynaptic neuron.

[0126] Astrocytes can also support synchronization or partial synchronization of neighboring neurons, which can be strengthened by increasing inhibitory dynamics. In heterogeneous inhibitory networks with sparse random connectivity, inhibitory dynamics provide a dual role in both suppressing population activity and generating neural reactivation. Synchronization of spike transmission within a circuit occurs when the input from excitatory neurons to adjacent inhibitory neurons is sufficiently strong. Connection strength is controlled by inhibitory innervation, and when inhibitory input is present, i.e., intermediate connections, neurons spike with slight periodicity / synchrony. The effect of this coupling is illustrated in Figures 6A and 6B.

[0127]

[0125] Figures 6A and 6B show that coordinated events in source populations propagate feedforward a) excitation and b) inhibition, where a shift toward higher excitation / inhibition sets neurons to fast / slow firing behavior with sparse bursts, and for each synchrony category, a population of postsynaptic neurons is assigned, and the synchrony category is mapped to a pattern of postsynaptic activity.

[0128]

[0126] The duration selectivity curve is asymmetric to accommodate the heterogeneity of spike latencies and resulting latency curves. Initial transient dynamics established through synaptic connections drive the network into a steady-state dynamic regime. As connectivity increases, a transition occurs from an ensemble of individually firing neurons to a coherent, synchronized network. Neurons generate time-locked patterns, and due to the interplay between conductance delays and plasticity rules, the network forms a set of neuronal groups with reproducible, precise firing sequences conditioned by activation patterns. As conductance increases (resulting in net excitation to the network), firing rates increase and become more uniform with a lower coefficient of variation. The sequencing of synaptic currents, i.e., outward, inward, reduces temporal jitter in the generation of action potentials in individual neurons, thereby creating a network with increased control over synchronous activity and homeostatic regulation.

[0129]

[0127] Both graphs show positions on both axes. By the present invention, excitation (A) and inhibition (B) events are localized in space, as indicated by the black bolt dots. When these figures are overlaid, the regions in space where both excitation and inhibition occur can be used to control dynamics using both excitation and inhibition.

[0130]

[0128] It should be noted that any features of the embodiments disclosed herein may be appropriately combined. For example, the different feedback and control structures disclosed above, such as AMPA, GABA, and NMDA receptors, axonal, dendritic, and neuronal backpropagation channels, and astroglial structures, may be employed alone or appropriately combined. The same applies to dendritic, IP3, mGluR, local neuronal, and axonal backpropagation receptors, as well as weight and calcium elements, derivation units (differentiators), and signal restoration units.

[0131]

[0129] As examples, receptors can be formed using gm-C filters, neural spike generators can be inverters (positive feedback reinforced), and weight elements can be implemented through different (static or dynamic) memory elements (latches, leakage compensation capacitors, SRAM cells, etc.).

Claims

1. A neurosynaptic structure (500) for a spiking neural network, wherein the neurosynaptic structure (500) is implemented using an electronic circuit and comprises a plurality of synaptic elements (502) and neurons (512), each synaptic element (502) being One or more input ports (501), where each input port is configured to receive an input signal. An output port configured to output postsynaptic signals, A signal path connecting one or more input ports to the output port, wherein the signal path includes a weight element (506) configured to apply weights to generate the postsynaptic signal, The neuron (512) comprises a cell body (514) and an axon (515), the cell body (514) is configured to receive and combine one or more of the postsynaptic signals, and the axon (515) is configured to generate a spike output when the membrane potential of the neuron (512) reaches a predetermined threshold in response to the output of the cell body. The neurosynaptic structure comprises an astrocyte structure (530) connected to a first subset of the synaptic elements, The astrocyte structure is configured to receive and sum the postsynaptic signals of each of the synaptic elements, and includes a modulation circuit configured to modulate the summed postsynaptic signals into a modulated signal. A neurosynaptic structure (500) characterized in that the modulated signal is used to alter the behavior of one or more of the first subset of synaptic elements (502) or of the neuron (512).

2. The neurosynaptic structure (500) according to claim 1, comprising an mGluR receptor (523) configured to integrate the modulated signal, the astrocyte structure configured to output the integrated modulated signal to each of the synaptic elements in the first subset, each of the plurality of synaptic elements configured to combine the integrated modulated signal with an integrated input signal, the integrated input signal obtained by integrating the input signal using AMPA and / or GABA receptors positioned along the signal pathway, configured to integrate the input signal.

3. The neurosynaptic structure (500) according to claim 2, wherein the mGluR receptor (523) is a voltage-driven pure integrator, and preferably the mGluR receptor (523) is formed using one or more gm-C filters.

4. The neurosynaptic structure (500) according to any one of claims 1 to 3, wherein the astrocyte structure comprises an NMDA receptor (524) configured to integrate the modulated signal, and the astrocyte structure is configured to output the integrated modulated signal to the neuron (512).

5. The neuron (512) includes a neuron error backpropagation signal path (520) comprising a calcium element (519) configured to connect the axon to the cell body (514) of the neuron (512) and to set the influence of the integrated axonal signal to create a neuron feedback signal, wherein the neuron error backpropagation signal path (520) is configured to supply the neuron feedback signal to the cell body (514) of the neuron (512). The neurosynaptic structure (500) according to claim 4, wherein the integrated modulated signal modulates the calcium element (519) of the neuron (512) to alter the influence of the integrated axonal signal of the neuron (512).

6. The modulation circuit comprises an IP3 receptor (521) configured to integrate the summed postsynaptic signals and an astrocyte calcium element (522) configured to set the effect of the modulated signals, wherein the IP3 receptor (521) and the astrocyte calcium element (522) generate the modulated signals, the neurosynaptic structure (500) according to any one of claims 1 to 3.

7. The neurosynaptic structure (500) according to claim 6, wherein the IP3 receptor (521) is a voltage-driven pure integrator, preferably the IP3 receptor (521) is formed using one or more gm-C filters, and / or the astrocyte calcium element (522) is an amplifier, preferably an amplifier faster than the weight element (506) of each synaptic element, and / or the astrocyte calcium element (522) provides adaptable feedback.

8. The neurosynaptic structure according to claim 6, comprising a synaptic element not directly connected to the neuron (512), wherein the astrocyte calcium element is modulated using a presynaptic neuron signal or a postsynaptic signal integrated using a dendritic receptor (526) coming from the synaptic element not directly connected to the neuron (512).

9. The neurosynaptic structure according to claim 6, wherein the axon (515) of the neuron (512) is connected to the astrocyte structure (530), and so the axonal error backpropagation receptor of the astrocyte is configured to integrate the axonal signal of the neuron (512), and the integrated axonal signal is used to modulate the astrocyte calcium element (522).

10. The integrated axonal signal is combined with other axonal error backpropagation signals from other neurons provided in the neurosynaptic structure (500) to form a combined axonal signal, which is used to modulate the astrocyte calcium element (522), preferably the time derivative of the combined axonal input signal is obtained by a derivation unit (529), more preferably the derivation unit is a differentiator, according to the neurosynaptic structure of claim 9.

11. At least one of the synaptic elements (202) comprises a signaling pathway including both an AMPA receptor (203) and a GABA receptor (204), and at least one synaptic element is configured to combine the outputs from the AMPA receptor and the GABA receptor to generate the postsynaptic signal, preferably, The neurosynaptic structure according to any one of claims 1 to 3, wherein the AMPA receptor and / or the GABA receptor of the neurosynaptic element can be configured to be activated or deactivated at a specific moment, and / or the AMPA receptor and / or the GABA receptor of the neurosynaptic element can be configured to select a frequency range of the input signal in which the AMPA receptor and / or the GABA receptor can operate.

12. The neurosynaptic structure (500) according to any one of claims 1 to 3, wherein the AMPA receptor comprises a voltage-driven integrator configured to generate an excitatory postsynaptic current for inducing charge inflow from the membrane potential of the neuron, and / or the GABA receptor is a voltage-driven integrator configured to generate an inhibitory postsynaptic current for inducing charge outflow from the membrane potential of the neuron.

13. The neurosynaptic structure (300) further comprises an axon error backpropagation signal path (311) connected to receive axonal signals from the axon (308) of the neuron (307), The axon error backpropagation signal path (311) comprises an axon error backpropagation receptor (312) configured to integrate the axon signal, The second subset of the synaptic elements (302) is configured to receive the integrated axonal signal and to modulate the weight elements (306) of the second subset of synaptic elements (302) based on the integrated axonal signal, preferably, The neurosynaptic structure (500) according to any one of claims 1 to 3, wherein each of the synaptic elements (302) of the second subset of the synaptic elements comprises an NMDA receptor configured to integrate the input signal to form an integrated NMDA signal, and the synaptic element (302) of the second subset of the synaptic elements is configured to combine the integrated NMDA signal and the integrated axonal signal and modulate the weight element (306) of the synaptic element (302) based on the integrated signal.

14. The neurosynaptic structure further comprises a dendritic error backpropagation signal path (408) configured to receive and combine the postsynaptic signals from one or more of the synaptic elements (402), The dendritic error backpropagation signal pathway (408) comprises a dendritic receptor (409) configured to integrate the combined postsynaptic signals, The third subset of the synaptic elements (402) is configured to receive the integrated combined postsynaptic signal and to modulate the weight elements (406) of the third subset of the synaptic elements (402) based on the integrated combined postsynaptic signal, preferably, The neurosynaptic structure (500) according to any one of claims 1 to 3, wherein each of the synaptic elements (402) of the third subset of the synaptic elements comprises an NMDA receptor (405) configured to integrate the input signal to form an integrated NMDA signal, and the synaptic element (402) of the third subset of the synaptic elements is configured to combine the integrated NMDA signal and the integrated combined postsynaptic signal, and to modulate the weight element (406) of the synaptic element based on the combined signal.

15. The neuron (512) comprises a neuronal backpropagation signaling pathway (520) comprising a local receptor (518) connected from the axon (515) of the neuron (512) to the cell body (514) of the neuron (512) and configured to integrate the axonal signal, and a calcium element (519) configured to set the effect of the integrated axonal signal and thereby create a neuronal feedback signal, wherein the neuronal backpropagation signaling pathway (520) is configured to supply the neuronal feedback signal to the cell body (514) of the neuron (512), and the neuronal feedback signal is combined with one or more postsynaptic signals, the neuronal synaptic structure (500) according to any one of claims 1 to 3.

Images