Method of mapping a natural neural network and electronic neural network device

Abstract

A method for mapping natural neural networks and an electronic neural network device are disclosed. The method includes: constructing a neural network graph of the natural neural network based on the membrane potentials of multiple biological neurons, wherein the membrane potentials correspond to at least two different forms of membrane potentials; and mapping the neural network graph to an electronic neural network device. Constructing and mapping the neural network graph involves learning the electronic neural network device. The method may further include: obtaining an input or stimulus; activating the learned electronic neural network device under the condition of the obtained input or stimulus to perform neural network operations; and generating a neural network result of the obtained input or stimulus based on the result of the activated learned electronic neural network device.

Description

[0001] This application claims the benefit of Korean Patent Application No. 10-2020-0150527, filed on November 11, 2020, with the Korean Intellectual Property Office, and Korean Patent Application No. 10-2021-0108472, filed on August 18, 2021, the entire disclosure of which is incorporated herein by reference for all purposes. Technical Field

[0002] The following description relates to methods for mapping natural neural networks to electronic neural network devices, as well as electronic neural network devices. Background Technology

[0003] Neuromorphic engineering involves attempting to mimic the network operations of biological nervous systems.

[0004] For example, various approaches to neuromorphic electronics can be categorized into natural efforts and unnatural efforts. Natural efforts attempt to accurately reproduce or mimic the structure, operation, and function of natural neural networks (NNNs), while unnatural efforts are based on mathematical models trained through machine learning to realize artificial neural networks (ANNs) with artificial structures. Natural efforts typically require individual consideration of a limited number (e.g., 10) of target biological neurons to identify natural neural networks (e.g., by using voltage or patch-clamp methods applied to selected biological neurons), or require extracellular macroscopic measurements of the firing of multiple action potentials (APs) of biological neurons in natural neural networks by using extracellular electrodes to collectively observe these firings. Extracellular electrodes generate noisy extracellular measurements of in vitro (dissociated cell culture) preparations or ex vivo (tissue section) preparations. Typically, extracellular macroscopic measurements cannot accurately measure or distinguish other synaptic potentials (such as post-synaptic potentials, PSPs, etc.) due to low sensitivity, poor registration, mixed signals and signal distortion, and / or due to the non-proximity arrangement of extracellular electrodes relative to individual neurons. Therefore, it is difficult to map individual connections between large numbers of biological neurons, and even more difficult to map the individual strength of such connections. Summary of the Invention

[0005] This summary is provided to introduce, in a simplified form, the selection of concepts further described below in the detailed description. This summary is not intended to identify key or essential features of the claimed subject matter, nor is it intended to help determine the scope of the claimed subject matter.

[0006] In one general aspect, a method for mapping a natural neural network to an electronic neural network device includes: constructing a neural network graph of the natural neural network based on the membrane potentials of a plurality of biological neurons of the natural neural network, wherein the membrane potentials correspond to at least two different forms of membrane potentials; and mapping the neural network graph to an electronic neural network device.

[0007] The construction and mapping of neural network graphs can be achieved based on the interaction of information from a first measurement of membrane potentials and information from a second measurement of membrane potentials regarding the presynaptic / postsynaptic relationships between presynaptic and postsynaptic biological neurons in a natural neural network. The first measurement of membrane potentials can correspond to a first form of membrane potential among the at least two different forms of membrane potentials, and the second measurement of membrane potentials can correspond to a different second form of membrane potential among the at least two different forms of membrane potentials.

[0008] The construction steps may include: identifying the connection structures between the plurality of biological neurons; and estimating the synaptic weights of the connections between the plurality of biological neurons.

[0009] The synaptic weights can be estimated based on the results of identifying the connection structure.

[0010] The method may further include: measuring the membrane potential of the plurality of biological neurons over time; extracting the action potential of the plurality of biological neurons from the action potential (AP) result of measuring the membrane potential; and extracting the postsynaptic potential of the plurality of biological neurons from the postsynaptic potential (PSP) result of measuring the membrane potential.

[0011] The step of measuring the membrane potential of the plurality of biological neurons may include: measuring the intracellular membrane potential of the plurality of biological neurons using intracellular electrodes.

[0012] The step of identifying the connection structure may include: identifying the connection structure between the multiple biological neurons based on the various time series of AP and PSP.

[0013] The step of identifying the connection structure may include: determining the presynaptic / postsynaptic relationship between the presynaptic and postsynaptic neurons of the plurality of biological neurons.

[0014] The step of estimating the synaptic weights may include estimating the synaptic weights of the connection between the presynaptic neuron and the postsynaptic neuron based on the individual PSPs of the postsynaptic neuron and the individual APs of the presynaptic neuron.

[0015] The mapping steps may include: mapping the plurality of biological neurons to the circuit layer of the electronic neural network device; and mapping the synaptic weights and corresponding connectivity between the plurality of biological neurons to the memory layer of the electronic neural network device.

[0016] The method of constructing and mapping a neural network graph enables the learning of an electronic neural network device, wherein the method may further include: obtaining an input or stimulus; activating the learned electronic neural network device under the condition of the obtained input or stimulus to perform neural network operations; and generating a neural network result of the obtained input or stimulus based on the result of the activated learned electronic neural network device.

[0017] In one general aspect, a non-transitory computer-readable storage medium stores instructions that, when executed by a processor, cause the processor to implement or perform one or more of the operations and / or methods described herein.

[0018] In one general aspect, a method for generating neural network results using an electronic device employing a learned electronic neural network device having learned synaptic connections and synaptic weights, the learned synaptic connections and synaptic weights possessing the characteristics of the learned electronic neural network device, the learned electronic neural network device having been mapped from the natural neural network based on the interaction of various information of measured action potentials (APs) and measured postsynaptic potentials (PSPs) between various presynaptic / postsynaptic biological neurons for the natural neural network, wherein the method may correspond to: receiving an input or stimulus; activating the learned electronic neural network device under the condition of the received input or stimulus to perform neural network operations; and generating a neural network result of the received input or stimulus based on the result of the activated learned electronic neural network device.

[0019] The method may further include: measuring AP using a first plurality of electrodes; measuring PSP using a second plurality of electrodes; and performing learning by the electronic neural network device by constructing a neural network graph of a natural neural network based on the interaction between the various information of the measured PSP and the various information of the measured AP using corresponding cross links of cross switches.

[0020] In each first time interval, the first plurality of electrodes may be different from the second plurality of electrodes, and in each different second time interval, some of the first plurality of electrodes may be the same as some of the second plurality of electrodes, in order to measure the additional AP or measure the additional PSP.

[0021] In one general aspect, a method for mapping a natural neural network to an electronic neural network device may include: using a plurality of neuronal modules of the electronic neural network device, considering at least two different forms of membrane potential measured from a plurality of biological neurons of the natural neural network; and based on said consideration, constructing a neural network graph in the electronic neural network device such that the electronic neural network device mimics the natural neural network.

[0022] The steps to be considered may include: considering the individual presynaptic / postsynaptic relationships between presynaptic and postsynaptic biological neurons in natural neural networks, and considering the interaction between the individual information of the measured action potentials (APs) and the individual information of the measured postsynaptic potentials (PSPs).

[0023] The construction steps may include: identifying the connection structure between the plurality of neuronal modules; and updating the synaptic weights of the connectivity between different neuronal modules in the plurality of neuronal modules.

[0024] In one general aspect, an electronic neural network device may correspond to: one or more memory layers configured to: store a neural network graph of a natural neural network for a plurality of neuron modules of the electronic neural network device; one or more circuit layers configured to: activate each of the plurality of neuron modules in response to a stimulus or input signal to the electronic neural network device, and perform signal transmission between the plurality of neuron modules; and a connector configured to connect the memory layers and the circuit layers.

[0025] The neural network result of the stored natural neural network graph can be generated based on the execution of the signal transmission.

[0026] When the electronic neural network device is a learned electronic neural network device, the information in the one or more memory layers and the information in the one or more circuit layers may have the characteristics of the electronic neural network device, which are mapped from the natural neural network based on the interaction of the various information of the measured action potentials (APs) and the various information of the measured postsynaptic potentials (PSPs) between the presynaptic and postsynaptic biological neurons of the natural neural network.

[0027] The connector may include at least one of the following: a through-silicon via (TSV) penetrating each of the one or more memory layers and each of the one or more circuit layers; and a microbump connecting the memory layers to the circuit layers.

[0028] The neural network result of the stored natural neural network graph can be generated based on the execution of the signal transmission, and the circuit layer can also be configured to: in response to a stimulus or input signal, activate the corresponding neuron module by reading the synaptic weights corresponding to the connectivity between the corresponding neuron module from the memory layer to generate the neural network result.

[0029] The one or more memory layers may be one or more cross switch arrays, and the synaptic weights in the neural network graph may be stored at the intersections of the one or more cross switch arrays.

[0030] The one or more memory layers and the one or more circuit layers can be stacked in three dimensions.

[0031] In one general aspect, an electronic device includes: a processor configured to: construct a neural network graph of a natural neural network based on the membrane potentials of a plurality of biological neurons in a natural neural network, wherein the membrane potentials correspond to at least two different forms of membrane potentials; and an electronic neural network device for mapping the neural network graph to the electronic device.

[0032] The processor can also be configured to: identify the connection structures between the plurality of biological neurons and estimate the synaptic weights of the connections between a plurality of biological neurons, respectively.

[0033] The processor can also be configured to map the plurality of biological neurons to the circuit layer of the electronic neural network device and to map the synaptic weights to the memory layer of the electronic neural network device.

[0034] The device may further include: electrodes for measuring the membrane potentials of the plurality of biological neurons over time, wherein the processor may further be configured to: extract the action potentials of the plurality of biological neurons from the action potentials (AP) of the measured membrane potentials; and extract the postsynaptic potentials of the plurality of biological neurons from the postsynaptic potentials (PSP) of the measured membrane potentials.

[0035] The construction and mapping of neural network graphs can be achieved based on the interaction of information from a first measurement of membrane potentials and information from a second measurement of membrane potentials regarding the presynaptic / postsynaptic relationships between presynaptic and postsynaptic biological neurons in a natural neural network. The first measurement of membrane potentials can correspond to a first form of membrane potential among the at least two different forms of membrane potentials, and the second measurement of membrane potentials can correspond to a different second form of membrane potential among the at least two different forms of membrane potentials.

[0036] Other features and aspects will become clear from the following detailed description, the accompanying drawings, and the claims. Attached Figure Description

[0037] Figure 1 Examples of natural neural network mapping systems according to one or more embodiments are shown.

[0038] Figure 2 This is a flowchart illustrating an example of a method for mapping a natural neural network to an electronic neural network according to one or more embodiments.

[0039] Figure 3 This is a flowchart illustrating an example of a method for mapping a natural neural network to an electronic neural network according to one or more embodiments.

[0040] Figure 4 An example of the structure of an electronic neural network according to one or more embodiments is shown.

[0041] Figure 5 An example of the architecture of a crossbar array according to one or more embodiments is shown.

[0042] Throughout the accompanying drawings and detailed embodiments, unless otherwise described or provided, the same reference numerals will be understood to denote the same or similar elements, features, and structures. The drawings may not be to scale, and for clarity, illustration, and convenience, the relative dimensions, scale, and depiction of elements in the drawings may be exaggerated. Detailed Implementation

[0043] The following detailed embodiments are provided to aid the reader in gaining a comprehensive understanding of the methods, apparatus, and / or systems described herein. However, various changes, modifications, and equivalents of the methods, apparatus, and / or systems described herein will become apparent upon understanding this disclosure. For example, the order of operations described herein is merely illustrative and is not limited to those orders set forth herein, but may be changed as will become clear upon understanding this disclosure, except for operations that must occur in a specific order. Furthermore, for clarity and brevity, descriptions of features known upon understanding this disclosure may be omitted.

[0044] The features described herein may be implemented in different forms and should not be construed as limited to the examples described herein. Rather, the examples described herein have been provided merely to illustrate some of the many feasible ways of implementing the methods, apparatus, and / or systems described herein that will be clear upon understanding the disclosure of this application.

[0045] Throughout this specification, when a component is described as "connected to" or "attached to" another component, that component may be directly "connected to" or "attached to" that other component, or there may be one or more other components in between. Conversely, when an element is described as "directly connected to" or "directly attached to" another element, there may be no other elements in between. Similarly, similar expressions (e.g., "between" and "immediately between," and "adjacent to" and "closely adjacent to") should be interpreted in the same manner. As used herein, the term "and / or" includes any one of the associated listed items and any combination of any two or more.

[0046] Although terms such as “first,” “second,” and “third” may be used herein to describe various components, assemblies, regions, layers, or parts, these components, assemblies, regions, layers, or parts should not be limited by these terms. Rather, these terms are used only to distinguish one component, assembly, region, layer, or part from another. Thus, without departing from the teaching of the examples described herein, the first component, first assembly, first region, first layer, or first part referred to as the first component, first assembly, first region, first layer, or first part may also be referred to as the second component, second assembly, second region, second layer, or second part.

[0047] The terminology used herein is for the purpose of describing various examples only and is not intended to limit disclosure. Unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. The terms “comprising,” “including,” and “having” indicate the presence of the described features, quantities, operations, components, elements, and / or combinations thereof, but do not preclude the presence or addition of one or more other features, quantities, operations, components, elements, and / or combinations thereof.

[0048] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains and based on the understanding of the disclosure of this application. Unless expressly defined herein, terms (such as those defined in a general dictionary) shall be interpreted as having the same meaning as they have in the relevant field and in the context of the disclosure of this application, and shall not be interpreted in an idealized or overly formalistic manner.

[0049] Examples herein can be of various types of products (such as personal computers (PCs), laptops, tablets, smartphones, televisions (TVs), smart home appliances, smart vehicles, self-service kiosks, and wearable devices), or can be implemented in various types of products (such as personal computers (PCs), laptops, tablets, smartphones, televisions (TVs), smart home appliances, smart vehicles, self-service kiosks, and wearable devices), but this embodiment is not limited thereto.

[0050] Figure 1 Examples of natural neural network mapping systems according to one or more embodiments are shown.

[0051] Reference Figure 1 A natural neural network mapping system can generate a neural network map by recording (measuring) or by recording / measuring and computationally analyzing neural signals generated in biological neurons. The neural network map may attempt to completely mimic the structure and (one or more) functions of multiple biological neurons (e.g., biological neurons of a large-scale natural neural network 130 as a non-limiting example). For example, as a non-limiting example, the large-scale neural network may be a natural neural network of an animal or human brain (or a natural neural network included in an animal or human brain), a corresponding nervous system, or other large-scale natural neural networks. The natural neural network 130 can be used to configure a neuromorphic processor by copying the natural neural network 130 (e.g., including biological neuron connection structures and corresponding connection strengths or weights) to a neuromorphic processor. In one or more embodiments, for example, the configuration of the neuromorphic processor can be performed directly using the measurement results of the corresponding biological neurons without computationally analyzing the measured potentials to identify the individual connections between biological neurons and / or the strengths or weights of such connections. Additionally, such measurements of neural signals in / generated by biological neurons can include different types or forms of membrane potential (e.g., measuring different types or forms of membrane potential using at least intracellular electrodes, such as via intracellular electrode interfaces). In the following text, the terms "neural network map," "functional map," and "synaptic connectivity map" can be understood to have the same meaning.

[0052] The natural neural network mapping system may include a recording unit 110 and mapping devices 120-1 and / or 120-2, wherein the recording unit 110 is used, for example, to measure in real time different types or forms of membrane potentials of each of a plurality of biological neurons in an example large-scale natural neural network 130, and the mapping devices 120-1 and / or 120-2 are used to configure electronic neural networks 140-1 and / or 140-2 to have the same structure as the natural neural network 130.

[0053] Electronic neural networks 140-1 and 140-2 can respectively reproduce or mimic biological operations based on the biological neurons of natural neural network 130. For example, in an example where electronic neural networks 140-1 and / or 140-2 (which have replicated the structure of natural neural network 130) can subsequently perform based on input information or stimuli, the results or functions of such performance by electronic neural networks 140-1 and / or 140-2 can be the same or substantially the same, as if the biological neurons of natural neural network 130 had responded to the same stimuli. In one example, electronic neural networks 140-1 and / or 140-2 can receive input or stimuli, and under the conditions of the received input or stimuli, electronic neural networks 140-1 and / or 140-2 can be activated to perform neural network operations, and the neural network results of the received input or stimuli can be generated based on the results of the activated electronic neural networks 140-1 and / or 140-2. Here, for example, the reference to biological neurons is a reference to living nerve cells, not artificial neurons. Additionally, in the following text, the terms "neuron" or "multiple neurons" and "nerve cells" are to be understood to have the same meaning as such biological neurons. Furthermore, operations based on biological neurons may include, for example, synaptic connection analysis, ion channel analysis, ion channel current measurement, and / or measurement of the effects of drugs on neural network connectivity and dynamics. However, examples are not limited to these.

[0054] The recording unit 110 may include an electrode layer comprising a plurality of (e.g., N) electrodes (e.g., intracellular electrodes) that can contact biological neurons to record (or measure) neural signals generated in the biological neurons and / or to inject (or provide) stimulation signals to the biological neurons.

[0055] For example, by using multiple electrodes, the recording unit 110 can use complementary metal-oxide-semiconductor (CMOS) nanoelectrode array (CNEA) technology (e.g., in real time) to read the electrical activity 115 of all individual biological neurons of a natural neural network 130 that are in contact with at least one of the electrodes.

[0056] The electrodes of the recording unit 110 can be independently connected to a single biological neuron to simultaneously perform individual recordings and measurements of the membrane potential of each of the multiple biological neurons in the natural neural network 130.

[0057] For example, in biological neurons, the membrane potential across the neuron typically has a resting membrane potential (e.g., approximately -70 mV). The membrane potential can increase or decrease, for example, due to the individual reception / stimulation of a neurotransmitter from another neuron by the biological neuron. For instance, the individual reception of this neurotransmitter can cause or affect the exchange of ions across the neuronal membrane, which in turn leads to changes in the membrane potential. Due to the cascading change in membrane potential from the resting membrane potential, when the changing membrane potential meets a specific threshold (e.g., approximately -45 mV), the biological neuron can generate an action potential (AP) (also referred to as a "nerve impulse" or "spike"). The firing of the AP toward the axon terminal of the biological neuron can also be referred to as the biological neuron "firing." In response to the AP, the biological neuron can release the aforementioned neurotransmitter. Here, the biological neuron that releases neurotransmitters can be called a presynaptic neuron, and the subsequent neuron that receives neurotransmitters can be called a postsynaptic neuron. The reception of neurotransmitters by the postsynaptic neuron can also be reflected in the changes in the membrane potential of the postsynaptic neuron, which can be called the postsynaptic potential (PSP). Therefore, AP and PSP are different forms or types of membrane potentials. Thus, based on, for example, neurotransmitters received by the postsynaptic neuron from the presynaptic neuron and neurotransmitters received by the postsynaptic neuron from other presynaptic neurons, the membrane potential of the postsynaptic neuron can repeatedly satisfy the aforementioned threshold and generate individual APs within the postsynaptic neuron. The time series of such APs generated by biological neurons can also be called their "spike train." For example, the timing and frequency of such pulses or spikes in the AP of a presynaptic neuron can represent the strength of the AP generated by the presynaptic neuron, and the timing and frequency of such pulses or spikes in the AP of a postsynaptic neuron can represent the strength of the AP generated by the postsynaptic neuron. Therefore, as a non-limiting example, the connection weighting or strength between presynaptic and postsynaptic neurons can be represented by a defined relationship between the AP of the presynaptic neuron and the PSP of the postsynaptic neuron.Here, while the above explanation pertains to the general pre- / post-synaptic relationship of neurons relative to such different forms or types of membrane potentials (e.g., AP and PSP neuronal signals), the above discussion is merely illustrative, for example, since the disclosure herein also applies to other neuronal types having different operations regarding the connection between (one or more) pre-synaptic neurons and (one or more) post-synaptic neurons for such information sharing between pre-synaptic and post-synaptic neurons that can be measured by intercellular electrodes. In one example, within each first time interval, the first plurality of electrodes differ from the second plurality of electrodes, and within each different second time interval, some of the first plurality of electrodes are the same as some of the second plurality of electrodes, for measuring additional AP or measuring additional PSP.

[0058] return Figure 1 A large amount of measurement data can be used to construct a neural map (e.g., by separate signal processing and analysis by mapping device 120-1), and the neural map can be mapped / copied to electronic neural network 140-1 (e.g., mapped / copied to electronic neural network 140-2) so as to have the same synaptic connections and synaptic weights as the natural neural network 130.

[0059] Optionally, a large amount of measurement data can be directly acquired, transmitted, provided, or received by the mapping device 120-2 in real time, for example. The mapping device 120-2 is configured to directly map / copy the synaptic connections and synaptic weights of the natural neural network 130 to the electronic neural network 140-2 based on the natural electrical activity between adjacent (i.e., presynaptic / postsynaptic) biological neurons of the natural neural network 130.

[0060] In the following text, reference will be made to Figure 2 This describes an example operation of constructing a neural network graph of a natural neural network through separate signal processing and replicating the constructed neural network graph to an electronic neural network 140-1, with reference to... Figure 3 This describes example operations of directly sending or providing extracted / measured neural signals, such as those from a natural neural network, to an electronic neural network 140-2 and using the electronic neural network 140-2 to construct and map / replicate neural network graphs from the natural neural network itself.

[0061] Figure 2 This is a flowchart illustrating, as a non-limiting example, a method for mapping a natural neural network to an electronic neural network (e.g., to a solid-state electronic memory network and circuit system) according to one or more embodiments.

[0062] Natural neural network mapping methods can be referenced from the above. Figure 1 The mapping device 120-1 described is executed. The mapping device 120-1 may be implemented by one or more hardware components, by one or more processors configured to implement the mapping method based on instructions executed by one or more processors, or by a combination thereof. Furthermore, the mapping device 120-1 may be included in an example electronic device (electronic neural network device 140-1) having an electronic neural network 140-1, or may be a separate external device (e.g., a personal computer) including or separate from the electronic neural network 140-1. The example electronic device may also include or exclude the following references. Figure 3 Further detailed discussion includes mapping device 120-2 and electronic neural network 140-2 (electronic neural network device 140-2). Additionally, the example electronic device may optionally include mapping device 120-2 and electronic neural network 140-2, but not mapping device 120-1 and electronic neural network 140-1. Furthermore, the example includes an electronic device comprising any one or both of electronic neural networks 140-1 and 140-2, the electronic device performing such neural network mapping of the respective electronic neural networks 140-1 and 140-2 and / or configured to perform any one or both of electronic neural networks 140-1 and 140-2 including neural networks with respect to input information, to artificially perform the natural operation and function of the corresponding mapped natural neural network for the same input information. The reference to electronic neural networks can also correspond to electronic devices that may also include one or more recording units and / or mapping devices, as well as remaining additional hardware components configured to perform one or more or all of the functions of various types of electronic devices such as those described above (e.g., personal computers (PCs), laptops, tablets, smartphones, televisions (TVs), smart home appliances, smart vehicles, self-service kiosks, and wearable devices).

[0063] return Figure 2 The mapping device 120-1 can construct a neural network graph by analyzing the data collected from the natural neural network and subsequently mapping the electronic neural network 140-1 to have the same configuration as the natural neural network. If the natural neural network graph is accurately mapped to the electronic neural network 140-1, the individual weight values ​​or connection strengths of the electronic neural network 140-1 can accurately represent the corresponding natural weights or connection strengths of the natural connections between the biological neurons of the natural neural network.

[0064] Reference Figure 2In operation 210, the mapping device 120-1 constructs a neural network graph of the natural neural network based on the membrane potentials of multiple biological neurons of the natural neural network. The mapping device 120-1 can extract action potentials (APs) and postsynaptic potentials (PSPs) (i.e., neuronal signals as various different forms (types) of the corresponding biological neurons) from the membrane potentials. The extraction of APs and PSPs can also or optionally be performed before operation 210 of the mapping device 120-1 (e.g., via the example circuit system of the recording unit 110).

[0065] The mapping device 120-1 can first identify individual connection structures between any(one or more) neurons and any(one or more) other neurons in a plurality of biological neurons based on separately received / measured membrane potentials. Identifying individual connection structures may include identifying presynaptic / postsynaptic relationships (i.e., corresponding presynaptic and postsynaptic neurons) between biological neurons. For example, connection structures between multiple biological neurons can be identified based on individual timing sequences of the AP and PSP of the biological neurons.

[0066] More specifically, the mapping device 120-1 can identify adjacent cells by analyzing the relationship between the PSP and AP of the measured biological neurons. For example, when the time interval between the AP of the first biological neuron and the PSP of the second biological neuron respectively meets a threshold interval (e.g., less than or equal to a threshold interval), and such time intervals meeting the threshold interval occur consecutively with a predetermined or higher frequency, the mapping device 120-1 can determine that the first and second biological neurons are matched and therefore have a presynaptic / postsynaptic relationship. Thus, the first biological neuron can be considered a presynaptic neuron, and the matched second biological neuron can be considered a postsynaptic neuron in this presynaptic / postsynaptic relationship. The first biological neuron may also have one or more other corresponding presynaptic / postsynaptic relationships in which the first biological neuron can be considered a presynaptic neuron, and other matched biological neurons can be considered corresponding postsynaptic neurons. Similarly, the second biological neuron may also have one or more other corresponding presynaptic / postsynaptic relationships, in which the second biological neuron can be considered a postsynaptic neuron, and other matching biological neurons can be considered corresponding presynaptic neurons. The first biological neuron may also be identified as a postsynaptic neuron with respect to the corresponding presynaptic / postsynaptic relationship having one or more matching presynaptic neurons, and the second biological neuron may also be identified as a presynaptic neuron with respect to the corresponding presynaptic / postsynaptic relationship having one or more matching postsynaptic neurons. In short, although regarding... Figure 2The operations (e.g., in the context of mapping device 120-1 and electronic neural network 140-1) discussed these neighboring cells and corresponding potential presynaptic / postsynaptic relationships between biological neurons in natural neural networks, but such discussion also applies to the following regarding Figure 3 The operation (e.g., in the context of mapping device 120-2 and corresponding electronic neural network 140-2) discusses the discrimination of the execution / realization of adjacent cells between biological neurons of the natural neural network or another natural neural network and the corresponding execution / realization of presynaptic / postsynaptic relationships, as well as the corresponding learning of presynaptic / postsynaptic relationships and corresponding synaptic weights.

[0067] After identifying the connection structure between biological neurons in each or more such matching presynaptic / postsynaptic relationships of the corresponding natural neural network, the mapping device 120-1 can estimate the respective individual synaptic connection strengths or weights (here referred to as individual synaptic weights) between each biological neuron match (i.e., between each determined presynaptic / postsynaptic relationship).

[0068] For example, mapping device 120-1 may be configured to reference postsynaptic neurons, and when the PSP of the reference postsynaptic neuron occurs (is measured), it estimates the synaptic weights between one or more defined presynaptic neurons that have a presynaptic / postsynaptic relationship with the reference postsynaptic neuron. The estimation of these synaptic weights can be performed by analyzing the interrelationships between the PSPs of one or more presynaptic neurons and one or more APs. For example, the analysis may include considering the size of the PSP of the presynaptic neuron and the strength (magnitude) of the AP. The intensity of the AP of each biological neuron can be a relative unit of measurement depending on the timing and frequency of the AP of the biological neuron (e.g., a first AP of a first biological neuron may have a determined intensity higher than a determined intensity of a second AP of a second biological neuron (corresponding to a delayed timing and / or lower frequency of the second AP) (corresponding to an earlier timing and / or faster frequency of the first AP). As another example, the individual intensities of the APs of presynaptic biological neurons (i.e., the APs of the respective spike queues) can be determined by summing the individual APs measured from each presynaptic biological neuron, wherein a higher summed AP result of one presynaptic biological neuron connected to a postsynaptic biological neuron will represent a higher intensity of the corresponding presynaptic / postsynaptic relationship, and another presynaptic... A lower summation of the AP of a biological neuron will represent a lower strength of the corresponding presynaptic / postsynaptic relationship. Here, a reference to the size of the AP of a biological neuron can be understood as representing any or all such considerations regarding the strength of the AP of the biological neuron. For example, mapping device 120-1 can estimate synaptic weights based on "the size of the PSP of the reference postsynaptic biological neuron" and "the individual sizes of the APs of the n presynaptic biological neurons connected to the reference postsynaptic biological neuron". Therefore, a neural network graph of a natural neural network can be generated based on determined presynaptic / postsynaptic relationships in the natural neural network and can include separately estimated synaptic weights for one or more or each of the biological neuron connections in these presynaptic / postsynaptic relationships in the natural neural network.

[0069] In operation 220, mapping device 120-1 maps the neural network graph to electronic neural network 140-1. Mapping device 120-1 can combine the neural network graph constructed in operation 210 based on the membrane potential of biological neurons in electronic neural network 140-1, which has the same configuration as natural neural networks.

[0070] As will be described in further detail below, the electronic neural network 140-1 may include one or more memory layers and one or more circuit layers. The memory layers are used to store mapped synaptic weights, and the circuit layers are used to perform operations on the biological neurons for each corresponding defined presynaptic / postsynaptic connection using the appropriately mapped synaptic weights stored in the memory layers(s). Therefore, the mapping device 120-1 can map multiple biological neurons to one or more circuit layers of the electronic neural network 140-1, and map corresponding synaptic weights and / or corresponding connectivity to one or more memory layers of the electronic neural network 140-1.

[0071] Figure 3 This is a flowchart illustrating, as a non-limiting example, a method for mapping a natural neural network to an electronic neural network (e.g., to a solid-state electronic memory network and circuit system) according to one or more embodiments.

[0072] For example, it can be referenced above. Figure 1 The described natural neural network mapping system is used to execute the natural neural network mapping method. The recording unit 110, the mapping device 120-2, and the electronic neural network 140-2 can be implemented by one or more hardware components, or can be implemented based on a combination of one or more processors and hardware components configured to implement the mapping method based on instructions executed by one or more processors, or by a combination thereof.

[0073] Mapping device 120-2 can map the synaptic weights of a natural neural network by directly sending (e.g., real-time measured) or providing the measured / read membrane potentials of the natural neural network to an electronic neural network 140-2. The electronic neural network 140-2 can learn presynaptic / postsynaptic relationships and the strength or weight of each of the presynaptic / postsynaptic biological neuronal connections based on the membrane potentials collected from the natural neural network. Based on this learning, the electronic neural network 140-2 can replicate the connection structure of the original natural neural network or mimic its behavior. In one example, the electronic neural network 140-2 mapped by mapping device 120-2 can mimic the response of the target natural neural network to one or more predetermined stimuli based on learning only time-series membrane potential information of some biological neurons in the target natural neural network measured / read from the target natural neural network (e.g., without using information related to the number of unmeasured neurons other than those measured in the target natural neural network and the connectivity between neurons). For example, since "some" biological neurons in the target natural neural network may not have such presynaptic / postsynaptic connections (e.g., the corresponding measured AP from one biological neuron may not match the measured PSP from another biological neuron, so these AP and PSP measurements will not affect the electronic neural network 140-2 learning such non-connections between such "some" biological neurons), corresponding portions of the electronic neural network 140-2 may not be learned or may not include synaptic weight information. For example, the corresponding synaptic weight information for such "some" biological neurons in the electronic neural network 140-2 may have zero values ​​at the corresponding portions of the electronic neural network 140-2 (e.g., memory elements).

[0074] Reference Figure 3 In operation 310, the mapping device 120-2 transmits or provides the measured / read membrane potentials of the plurality of biological neurons constituting the natural neural network to an electronic neural network 140-2 comprising a plurality of neuronal modules (e.g., physical or virtual neuron representations provided by hardware provided by the processor and / or corresponding circuit layers of the electronic neural network 140-2). Therefore, each of the plurality of neuronal modules in the electronic neural network 140-2 may correspond to a corresponding biological neuron in the natural neural network.

[0075] In operation 320, based on the transmitted or provided measured / read membrane potential, the electronic neural network 140-2 constructs a neural network graph during its learning process to ultimately mimic a natural neural network. For example, with Figure 2 Compared to (or except for) the operation Figure 2Apart from its operation, the electronic neural network 140-2 can operate without a separate external device (e.g., without the example of a separate mapping device 120-1) and does not need to perform [operations]. Figure 2 Operation 210 analyzes and constructs a neural network graph by identifying presynaptic / postsynaptic biological neurons and estimating the connection strength or weights between them. Specifically, the electronic neural network 140-2 may construct its own neural network graph based on individual inputs to the electronic neural network 140-2 regarding measured / readout membrane potentials. For this purpose, the electronic neural network 140-2 may include a processor or other circuitry system for constructing the neural network graph. As a non-limiting example, the processor may include a crossbar memory structure (e.g., as a memory layer of the electronic neural network 140-2). In one example, the electronic neural network 140-2 may have one or more memory layers and one or more circuitry layers. For example, a crossbar memory structure may correspond to... Figure 5 500 cross switch arrays.

[0076] Constructing a neural network graph may include mapping the connection structure between multiple neuron modules of the electronic neural network 140-2 in a processor and setting or updating synaptic weights among the multiple neuron modules. (For example, represented by one or more circuit layers of the electronic neural network 140-2) The neuron module circuitry can control the updating of synaptic weight values ​​through spike-timing-dependent plasticity (STDP) learning by the processor.

[0077] For example, the electronic neural network 140-2 may represent including a pulse converter (e.g., as one of the circuit layers 420 of the electronic neural network 140-2) and may represent including a delay converter (e.g., as one of the circuit layers 420 of the electronic neural network 140-2). The pulse converter is used to convert the AP and PSP signaling of the biological neurons into memory write pulses with fixed time intervals, and the delay converter is used to adjust each interval between the individual pulses inversely proportional to the size of the PSP. Furthermore, the extracted AP and PSP pulses can be sent / input to a processor to adjust the processor's target crossover point (e.g., ...). Figure 5 The conductance of the cross points of the 500 cross switch array.

[0078] Electronic neural networks can change or update the values ​​of synaptic weights between neuronal modules through STDP learning. In one or more embodiments, the electronic neural network can map the connection strength between two neuronal modules as increasing as the time interval between the AP and PSP of the two connected neuronal modules decreases, based on the STDP property of resistive random access memory (RRAM).

[0079] Depending on the implementation, synaptic weights can be changed or updated with predetermined values ​​via a simple comparator, or they can be selected from several values ​​based on the difference in firing timing using a lookup table (LUT) scheme with a corresponding LUT stored in any memory of an electronic neural network, neural network mapping system, or electronic device. For example, the weight updates of synaptic modules can occur independently based on the sharing of information between adjacent neuronal modules (e.g., the individual characteristics of AP and PSP neural signals based on each natural presynaptic / postsynaptic relationship). The values ​​of synaptic weights can be updated in a variety of other ways, and therefore, examples are not limited to the synaptic weight update methods described above.

[0080] Figure 4 An example of the structure of an electronic neural network according to one or more embodiments is shown.

[0081] Reference Figure 4 The electronic neural network may include one or more memory layers 410, one or more circuit layers 420, and connectors 430. (See reference...) Figures 1 to 3 The provided description can be applied to Figure 4 Examples, and Figure 4 The electronic neural network can also correspond to Figures 1 to 3 The electronic neural network is therefore, for ease of description, repeated descriptions will be omitted.

[0082] Memory layer 410 can store the neural network graph of a natural neural network. For example, each of memory layers 410 can store the synaptic weights between biological neurons in the presynaptic / postsynaptic relationship of the mapped neural network. For example, one of memory layers 410 can store the synaptic weights between the i-th biological neuron (as a presynaptic biological neuron) and the j-th biological neuron (as a correspondingly connected postsynaptic biological neuron). Similarly, each of memory layers 410 can store the corresponding synaptic weights between the respective presynaptic and postsynaptic biological neurons in the corresponding presynaptic / postsynaptic relationship of the natural neural network.

[0083] Memory layer 410 can be capable of storing all synaptic weights for each of the presynaptic / postsynaptic relationships in biological neuronal connections. As an example, in a network including N (e.g., 10...), 9 In an example natural neural network with 1000 neurons, each neuron has K (e.g., 1000) synaptic connections, and memory layer 410 can be able to store K×N / 2 (e.g., 1000×10) synaptic connections. 9 / 2) synaptic weights. The following will refer to a non-limiting example as memory layer 410. Figure 5 To describe in more detail an example architecture of a crossbar switch array capable of efficiently storing such large amounts of data.

[0084] therefore, Figure 5 An example of the architecture of a cross switch array according to one or more embodiments is shown.

[0085] like Figure 5 As shown, the memory layer (e.g., Figure 4 The memory layer 410 can be implemented in the architecture of the cross switch array 500. The cross switch array 500 may include a first electrode 510 arranged in multiple rows on a substrate, a second electrode 520 arranged in multiple rows intersecting the first electrode 510, and memory elements 530 disposed between the first electrode 510 and the second electrode 520, each memory element having a resistance that varies according to the voltage applied between the respective first electrode 510 and the second electrode 520.

[0086] The mapping device 120-1 can, for example, map multiple biological neurons constituting a natural neural network to a first electrode 510 and a second electrode 520, and map the respective synaptic weights between each biological neuron to a memory element 530. As a non-limiting example, the mapping device 120-1 can map N biological neurons constituting a natural neural network to first electrodes 510 and second electrodes 520 arranged in their respective N rows (e.g., where the two N rows have an equal number of rows corresponding to the equal number of N biological neurons). Thereafter, the synaptic weights between the i-th biological neuron and the j-th biological neuron in the generated neural network graph of the natural neural network can be stored in the corresponding memory element 530 located at the intersection of the first electrode 510 corresponding to the i-th biological neuron and the second electrode 520 corresponding to the j-th biological neuron. In this case, the mapping device 120-1 can, for example, store the synaptic weights in the memory element 530 by adjusting the variable resistance value of the memory element 530. Neural signals (e.g., AP neural signals or pulses) can be provided to sequential rows 1 to N of the first electrode 510 from 1 to N biological neurons. Similarly, different neural signals (e.g., PSP neural signals or pulses) can be provided to sequential rows 1 to N of the second electrode 520 from 1 to N biological neurons, but the embodiments are not limited thereto. For example, there may be any order in which neural signals from 1 to N biological neurons are provided to 1 to N first electrodes 510, which may be the same as or different from any order in which other neural signals from 1 to N biological neurons are provided to 1 to N second electrodes 520.

[0087] A crossbar switch array 500 with an N×N structure can be used to store N (e.g., 10) 9The cross-switch array 500, with its N×N structure, stores synaptic weights between biological neurons. However, as mentioned above, since a biological neuron can have many K (e.g., 1000) synaptic connections, many regions of the cross-switch array 500 may be unused because the corresponding biological neuron is not connected to (e.g., not at all or not sufficiently connected) another biological neuron in the natural neural network. In one or more embodiments, since the mapping device 120-1 knows the relationships between the actually connected biological neurons, for example, through the already generated neural network graph, the cross-switch array 500 can store only the synaptic weights present for the determined presynaptic / postsynaptic biological neurons, and can exclude or avoid cross-points with synaptic weights representing no or zero values. This improves the efficiency of the memory layer.

[0088] However, the architecture of memory layer 410 is not necessarily limited to crossbar switch array 500.

[0089] Return to reference Figure 4 The circuit layer 420 can activate each of the multiple neuron modules in response to a received signal and perform signal transmission / providing between the multiple neuron modules.

[0090] Circuit layer 420 may include stacked circuitry, and the stacked circuitry may be circuitry configured to perform functions such as the aforementioned neural signal measurement, signal processing, analysis, and / or any other operations discussed herein (e.g., to cooperate with one or more memory layers 410 capable of storing synaptic weights). Circuitry with the aforementioned functions may be distributed across multiple circuit layers or integrated into a single circuit layer. The circuitry may be, for example, a CMOS integrated circuit (IC). However, the examples are not limited to this. To replicate the connection structure of large-scale natural neural networks (e.g., those with multiple biological neurons), electronic neural network structures of the same or similar size as natural neural networks may be used. For example, the electronic neural network may be a 3D stacked system that can improve the degree of integration between layers.

[0091] For example, memory layer 410 may include, for example, memory layer 1, memory layer 2, ..., memory layer L. For example, memory layers may be stacked vertically on top of each other.

[0092] Similarly, circuit layer 420 may include, for example, circuit layer 1, circuit layer 2, ..., circuit layer M. For example, circuit layers may be stacked vertically on top of each other. Each circuit layer may include circuitry for performing different functions or operations. For example, circuit layer 1 may include circuitry for performing accumulation, circuit layer 2 may include circuitry for discharging, and circuit layer M may include circuitry for voltage amplification.

[0093] Optionally, each circuit layer may include circuitry for performing the same function or operation. For example, circuit layer 1 may include circuitry for accumulation and circuitry for discharge; similarly, circuit layer 2 may also include circuitry for accumulation and circuitry for discharge.

[0094] Connector 430 can connect memory layer 410 and circuit layer 420. Connector 430 can be at least one of, for example, a through silicon via (TSV, also known as a through-silicon via) passing through memory layer 410 and circuit layer 420 and a micro bump connecting memory layer 410 and circuit layer 420.

[0095] TSV (Through-Video) is a packaging technology that uses fine through-holes drilled in a chip and filled with conductive material to connect upper and lower chips, instead of using wires to connect the chips. Because TSV ensures direct electrical connection paths within the chip, it uses less space than previous non-TSV packages, allowing for smaller package sizes and shorter interconnect lengths between chips.

[0096] In response to receiving a stimulus signal, the circuit layer 420 can read the synaptic weights corresponding to the connected neuron module from the memory layer 410 which stores synaptic weights, and activate the neuron module. In this case, the connector 430 can also transmit signals between the memory layer 410 and the circuit layer 420.

[0097] According to the example, a natural neural network mapping device may include a processor or other circuit system that receives membrane potentials of multiple biological neurons constituting a natural neural network, constructs a neural network graph of the natural neural network based on the membrane potentials, and maps the neural network graph to an electronic neural network.

[0098] (For example, a processor or other circuitry system of a corresponding electronic device) can identify individual connection structures between multiple biological neurons and estimate the synaptic weights between those biological neurons for which the connection structures are identified.

[0099] (For example, a processor or other circuitry system in a corresponding electronic device) can map multiple biological neurons to circuit layers of an electronic neural network and map synaptic weights to memory layers of the electronic neural network. The electronic device can implement the mapped neurons and synaptic weights to artificially achieve the same functionality as the measured biological neurons in the original biological neural network.

[0100] Regarding Figures 1 to 5The neural network mapping systems, electronic devices, mapping devices, electronic neural networks, electronic neural network devices, recording units, membrane potential recording / measuring electrodes, signal and / or analysis processors, processors, neuromorphic processors, cross switches, memory elements, resistive random access memory, memory layers, circuit layers, circuit systems for performing accumulation, circuit systems for discharging, circuit systems for voltage amplification, CMOS integrated circuits (ICs), 3D stacked systems, 3D vertical stacked systems, neuron modules, electrodes, complementary metal-oxide-semiconductor (CMOS) nanoelectrode arrays, solid-state electronic memory networks and / or circuit systems, and other devices, apparatuses, modules, elements, and components described as non-limiting examples are implemented by hardware components. Examples of hardware components that can be used to perform the operations described in this application include, where appropriate, controllers, sensors, generators, drivers, memories, comparators, arithmetic logic units, adders, subtractors, multipliers, dividers, integrators, and any other electronic components configured to perform the operations described in this application. In other examples, one or more of the hardware components performing the operations described in this application are implemented by computing hardware (e.g., by one or more processors or computers). A processor or computer may be implemented by one or more processing elements, such as logic gate arrays, controllers and arithmetic logic units, digital signal processors, microcomputers, programmable logic controllers, field-programmable gate arrays, programmable logic arrays, microprocessors, or any other means or combination of means configured to respond to and execute instructions in a defined manner to achieve a desired result. In one example, the processor or computer includes or is connected to one or more memories storing instructions or software executed by the processor or computer. Hardware components implemented by the processor or computer may execute instructions or software (such as an operating system (OS) and one or more software applications running on the OS) for performing the operations described herein. The hardware components may also access, manipulate, process, create, and store data in response to the execution of instructions or software. For simplicity, the singular terms “processor” or “computer” may be used in the description of the examples described herein, but in other examples, multiple processors or computers may be used, or a processor or computer may include multiple processing elements or multiple types of processing elements or both. For example, a single hardware component or two or more hardware components may be implemented by a single processor, or two or more processors, or a processor and a controller. One or more hardware components may be implemented by one or more processors, or processors and controllers, and one or more other hardware components may be implemented by one or more other processors, or additional processors and additional controllers. One or more processors, or processors and controllers, may implement a single hardware component or two or more hardware components.The hardware components can have any one or more different processing configurations, examples of which include: a single processor, a discrete processor, a parallel processor, a single instruction single data (SISD) multiprocessing, a single instruction multiple data (SIMD) multiprocessing, multiple instruction single data (MISD) multiprocessing, and multiple instruction multiple data (MIMD) multiprocessing.

[0101] Perform the operations described in this application Figures 1 to 5 The method is executed by computing hardware (e.g., by one or more processors or a computer), which is implemented to execute instructions or software as described above to perform the operations performed by the method as described in this application. For example, a single operation or two or more operations may be executed by a single processor, or two or more processors, or a processor and a controller. One or more operations may be executed by one or more processors, or a processor and a controller, and one or more other operations may be executed by one or more other processors, or additional processors and additional controllers. One or more processors, or a processor and a controller, may execute a single operation or two or more operations.

[0102] Instructions or software for controlling computing hardware (e.g., one or more processors or computers) to implement hardware components and perform the methods described above can be written as computer programs, code segments, instructions, or any combination thereof to individually or collectively instruct or configure one or more processors or computers, such as machines or special-purpose computers, to perform the operations performed by the hardware components and methods described above. In one example, the instructions or software include machine code (such as machine code generated by a compiler) that is directly executed by one or more processors or computers. In another example, the instructions or software include high-level code that is executed by one or more processors or computers using an interpreter. The instructions or software can be written in any programming language based on the block diagrams and flowcharts shown in the accompanying drawings and the corresponding description used herein, which disclose algorithms for performing the operations performed by the hardware components and methods described above.

[0103] Instructions or software for controlling computing hardware (e.g., one or more processors or computers) to implement hardware components and perform the methods described above, along with any associated data, data files, and data structures, may be recorded, stored, or fixed on one or more non-transitory computer-readable storage media. Examples of non-transitory computer-readable storage media include: read-only memory (ROM), random access programmable read-only memory (PROM), electrically erasable programmable read-only memory (EEPROM), random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), flash memory, non-volatile memory, CD-ROM, CD-R, CD+R, CD-RW, CD+RW, DVD-ROM, DVD-R, DVD+R, DVD-RW, DVD+RW, DVD-RAM, BD-ROM, BD-R, BD-RLTH, BD-RE Blu-ray or optical disc storage devices, hard disk drives (HDDs), solid-state drives (SSDs), card storage devices (such as multimedia cards or microcards (e.g., Secure Digital (SD) or Extreme Digital (XD))), magnetic tape, floppy disks, magneto-optical data storage devices, optical data storage devices, hard disks, solid-state drives, and any other devices configured to store instructions or software and any associated data, data files, and data structures in a non-transitory manner and to provide the instructions or software and any associated data, data files, and data structures to one or more processors or computers, such that one or more processors or computers can execute the instructions. In one example, the instructions or software and any associated data, data files, and data structures are distributed across a networked computer system, such that the instructions and software and any associated data, data files, and data structures are stored, accessed, and executed in a distributed manner by one or more processors or computers.

[0104] While this disclosure includes specific examples, it will be clear upon understanding this disclosure that various changes in form and detail may be made in these examples without departing from the spirit and scope of the claims and their equivalents. The examples described herein should be considered descriptive only and not for limiting purposes. The description of features or aspects in each example should be considered applicable to similar features or aspects in other examples. Suitable results may be achieved if the described techniques are performed in a different order, and / or if components in the described system, architecture, apparatus, or circuit are combined in a different manner, and / or replaced or supplemented by other components or their equivalents.

Claims

1. A method for mapping a natural neural network to an electronic neural network device, the method comprising: A neural network graph of a natural neural network is constructed based on the membrane potentials of multiple biological neurons, wherein the membrane potentials correspond to at least two different forms of membrane potentials; and Mapping neural network graphs to electronic neural network devices. The construction and mapping of the neural network graph are achieved by the interaction of information from the first measurement of membrane potentials and information from the second measurement of membrane potentials regarding the presynaptic / postsynaptic relationships between presynaptic and postsynaptic biological neurons in natural neural networks. Wherein, the first measured membrane potential corresponds to the action potential among the at least two different forms of membrane potential, and the second measured membrane potential corresponds to the postsynaptic potential among the at least two different forms of membrane potential. Specifically, when the time interval between the action potential of the presynaptic neuron and the postsynaptic potential of the postsynaptic neuron satisfies a threshold interval, and the time interval satisfying the threshold interval occurs continuously with a predetermined or higher number of occurrences or frequencies, the presynaptic neuron and the postsynaptic neuron are matched and have a presynaptic / postsynaptic relationship.

2. The method according to claim 1, wherein, The construction steps include: Identify the connection structures between the plurality of biological neurons; and Estimate the synaptic weights of the connections between the multiple biological neurons.

3. The method according to claim 2, wherein, The synaptic weights are estimated based on the results of identifying the connection structure.

4. The method according to claim 2, further comprising: The membrane potential of the multiple biological neurons was measured over time; The action potentials of the multiple biological neurons are extracted from the action potential results of the measured membrane potentials. as well as The postsynaptic potentials of the plurality of biological neurons are extracted from the postsynaptic potential results of the measurement of the membrane potential.

5. The method according to claim 4, wherein, The steps for measuring the membrane potential of the plurality of biological neurons include: measuring the intracellular membrane potential of the plurality of biological neurons using intracellular electrodes.

6. The method according to claim 4, wherein, The step of identifying the connection structure includes: identifying the connection structure between the plurality of biological neurons based on the timing of each action potential and the timing of each postsynaptic potential.

7. The method according to claim 2, wherein, The step of identifying the connection structure includes: determining the presynaptic / postsynaptic relationship between the presynaptic and postsynaptic neurons of the plurality of biological neurons.

8. The method according to claim 7, wherein, The step of estimating the synaptic weights includes estimating the synaptic weights of the connection between the presynaptic neuron and the postsynaptic neuron based on the individual postsynaptic potentials of the postsynaptic neuron and the individual action potentials of the presynaptic neuron.

9. The method according to claim 2, wherein, The mapping steps include: Mapping the plurality of biological neurons to the circuit layer of an electronic neural network device; and The synaptic weights and corresponding connectivity between the multiple biological neurons are mapped to the memory layer of the electronic neural network device.

10. The method according to any one of claims 1 to 9, wherein, Constructing and mapping neural network graphs to perform learning on electronic neural network devices, and The method further includes: To receive input or stimulation; Under the condition of the received input or stimulus, the learned electronic neural network device is activated to perform neural network operations; and Based on the results of the activated, learned electronic neural network device, the neural network results of the obtained input or stimulus are generated.

11. A non-transitory computer-readable storage medium storing instructions, which, when executed by a processor, cause the processor to perform the method according to any one of claims 1 to 10.

12. A method for generating neural network results using an electronic device with learned synaptic connections and synaptic weights, the learned synaptic connections and synaptic weights possessing the characteristics of a learned electronic neural network device, the learned electronic neural network device being mapped from a natural neural network based on the interaction of measured action potentials and measured postsynaptic potentials between presynaptic and postsynaptic biological neurons in a natural neural network, the method comprising: To receive input or stimulation; Under the conditions of the input or stimulus received, the learned electronic neural network device is activated to perform neural network operations; as well as Based on the results of the activated, learned electronic neural network device, the neural network output of the acquired input or stimulus is generated. Specifically, when the time interval between the action potential of the presynaptic neuron and the postsynaptic potential of the postsynaptic neuron satisfies a threshold interval, and the time interval satisfying the threshold interval occurs continuously with a predetermined or higher number of occurrences or frequencies, the presynaptic neuron and the postsynaptic neuron are matched and have a presynaptic / postsynaptic relationship.

13. The method of claim 12, further comprising: The action potential is measured using a first plurality of electrodes; Postsynaptic potentials were measured using a second set of multiple electrodes; as well as The learning of the electronic neural network device is performed by constructing a neural network graph of a natural neural network based on the interaction of various information of the measured postsynaptic potential and various information of the measured action potential using corresponding cross links of cross switches.

14. The method according to claim 13, wherein, In each first time interval, the first plurality of electrodes are different from the second plurality of electrodes, and in each different second time interval, some of the first plurality of electrodes are the same as some of the second plurality of electrodes, in order to measure additional action potentials or measure additional postsynaptic potentials.

15. A non-transitory computer-readable storage medium storing instructions, which, when executed by a processor, cause the processor to perform the method according to any one of claims 12 to 14.

16. A method for mapping a natural neural network to an electronic neural network device, the method comprising: Using multiple neuronal modules of an electronic neural network device, considering at least two different forms of membrane potential measured from multiple biological neurons of a natural neural network; as well as Based on the aforementioned considerations, a neural network graph is constructed in the electronic neural network device to enable the device to mimic natural neural networks. This includes considering the interaction between various information from the measured action potentials and various information from the measured postsynaptic potentials in the presynaptic / postsynaptic relationships between presynaptic and postsynaptic biological neurons in natural neural networks. Specifically, when the time interval between the action potential of the presynaptic neuron and the postsynaptic potential of the postsynaptic neuron satisfies a threshold interval, and the time interval satisfying the threshold interval occurs continuously with a predetermined or higher number of occurrences or frequencies, the presynaptic neuron and the postsynaptic neuron are matched and have a presynaptic / postsynaptic relationship.

17. The method according to claim 16, wherein, The construction steps include: Identify the connection structures between the plurality of neuronal modules; and Update the synaptic weights for connectivity between different neuronal modules in the plurality of neuronal modules.

18. A non-transitory computer-readable storage medium storing instructions, which, when executed by a processor, cause the processor to perform the method according to any one of claims 16 to 17.

19. An electronic neural network device, comprising: One or more memory layers are configured to store a neural network graph of a natural neural network for a plurality of neuron modules of the electronic neural network device; One or more circuit layers are configured to: activate each of a plurality of neuron modules in response to a stimulus or input signal to the electronic neural network device, and perform signal transmission between the plurality of neuron modules; as well as Connectors are configured to connect the one or more memory layers and the one or more circuit layers. Wherein, when the electronic neural network device is a learned electronic neural network device, the information in the one or more memory layers and the information in the one or more circuit layers possess the characteristics of the electronic neural network device. These characteristics are mapped from the natural neural network based on the interaction of various information from measured action potentials and measured postsynaptic potentials between presynaptic and postsynaptic biological neurons in a natural neural network. When the time interval between the action potential of the presynaptic neuron and the postsynaptic potential of the postsynaptic neuron satisfies a threshold interval, and the time interval satisfying the threshold interval occurs continuously with a predetermined or higher number of occurrences or frequencies, the presynaptic neuron and the postsynaptic neuron are matched and have a presynaptic / postsynaptic relationship.

20. The electronic neural network device according to claim 19, wherein, The neural network result of the stored natural neural network graph is generated based on the execution of the signal transmission.

21. The electronic neural network device according to claim 19, wherein, The connector includes at least one of the following: Through-silicon vias penetrate each of the one or more memory layers and each of the one or more circuit layers; as well as Microbumps connect the various memory layers to the various circuit layers.

22. The electronic neural network device according to claim 19, in, The neural network result of the stored natural neural network graph is generated based on the execution of the signal transmission, and The one or more circuit layers are further configured to: in response to a stimulus or input signal, activate a corresponding neuron module by reading synaptic weights corresponding to the connectivity between the corresponding neuron module from the one or more memory layers to generate the neural network result.

23. The electronic neural network device according to any one of claims 19 to 22, wherein, The one or more memory layers are one or more cross-switch arrays, wherein the synaptic weights in the neural network graph are stored at the respective intersections of the one or more cross-switch arrays.

24. The electronic neural network device according to any one of claims 19 to 22, wherein, The one or more memory layers and the one or more circuit layers are stacked in three dimensions.

25. An electronic device comprising: The processor is configured as follows: A neural network graph of a natural neural network is constructed based on the membrane potentials of multiple biological neurons, wherein the membrane potentials correspond to at least two different forms of membrane potentials; and An electronic neural network device that maps neural network graphs to electronic devices. The construction and mapping of the neural network graph are achieved by the interaction of information from the first measurement of membrane potentials and information from the second measurement of membrane potentials regarding the presynaptic / postsynaptic relationships between presynaptic and postsynaptic biological neurons in natural neural networks. The first measured membrane potential corresponds to the action potential among the at least two different forms of membrane potentials, and the second measured membrane potential corresponds to the postsynaptic potential among the at least two different forms of membrane potentials. When the time interval between the action potential of the presynaptic neuron and the postsynaptic potential of the postsynaptic neuron satisfies a threshold interval, and the time interval satisfying the threshold interval occurs continuously with a predetermined or higher number of occurrences or frequencies, it is determined that the presynaptic neuron and the postsynaptic neuron are matched and have a presynaptic / postsynaptic relationship.

26. The electronic device according to claim 25, wherein, The processor is also configured to: identify the connection structures between the plurality of biological neurons, and estimate the synaptic weights of the connections between the plurality of biological neurons respectively.

27. The electronic device according to claim 26, wherein, The processor is also configured to map the plurality of biological neurons to the circuit layer of the electronic neural network device and to map the synaptic weights to the memory layer of the electronic neural network device.

28. The electronic device of claim 25, further comprising: Electrodes are used to measure the membrane potential of the multiple biological neurons over time. The processor is also configured as follows: The action potentials of the multiple biological neurons were extracted from the measured action potentials of the membrane potentials; and The postsynaptic potentials of the multiple biological neurons are extracted from the measured membrane potentials and postsynaptic potentials.

Images