Method and apparatus for electronic memory mapping and copying of biological neural networks

Abstract

To provide a method and a device for mapping a biological neural network to an electronic memory network and a circuit.SOLUTION: A method for mapping a biological neural network to an electronic memory network and a circuit, according to one embodiment includes: a step for configuring a neural network map of the biological neural network based on membrane potentials for each of plural organism neurons constituting the biological neural network; and a step for mapping the neural network map to an electronic neural network.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The following embodiments relate to an electronic memory copy mapping method and apparatus for a biological neural network.

Background Art

[0002] Neuromorphic engineering is for mimicking the network structure and functions of the biological nervous system using analog integrated circuits. For example, it is not only to completely mimic the structure and functions of the brain in an electronic system, but also to understand some operating principles of the brain and to expand into a form of realizing a system applying these principles.

[0003] Neuromorphic electronic devices are classified into attempts to most strictly reproduce the structure and functions of the brain's natural neuronal network (NNN), and attempts to realize an artificial neural network (ANN), which is a mathematical model, with low power consumption. As the former attempt, it is required to grasp the connection of a large number of neurons and the map of the natural neural network in which the strength of each connection is specified, but this is extremely difficult.

Summary of the Invention

Problems to be Solved by the Invention

[0004] An object of the present invention is to provide an electronic memory copy mapping method and apparatus for a biological neural network.

Means for Solving the Problems

[0005] A method for mapping an electronic memory network of a biological neural network to a circuit according to an embodiment includes: a method of directly mapping a neural network map of the biological neural network to an electronic memory network if the neural network map of the biological neural network is configured based on the membrane potential of each of a plurality of biological neurons constituting the biological neural network; and a method of driving the electronic memory network with a membrane potential signal without clearly configuring a map so that the electronic memory network configures a map.

[0006] The construction and mapping of the neural network map is achieved based on the information of the first measured membrane potential interacting with the information of the second measured membrane potential for each synaptic relationship between presynaptic and postsynaptic neurons in the biological neural network, wherein the first measured membrane potential can correspond to a first form of membrane potential of the at least two different forms of membrane potential, and can correspond to a second form of membrane potential of the at least two different forms of membrane potential.

[0007] The steps of constructing the neural network map may include the steps of understanding the connectivity structure between the plurality of living neurons and estimating the synaptic weight between the plurality of living neurons.

[0008] The step of estimating the synaptic load may include the step of estimating the synaptic load based on the result of understanding the connection structure.

[0009] A method for mapping a biological neural network to an electronic memory network and circuit according to one embodiment may further include the steps of measuring the membrane potential of a plurality of biological neurons over time, and extracting the action potentials (APs) and postsynaptic potentials (PSPs) of each of the plurality of biological neurons based on the membrane potential.

[0010] The step of measuring the membrane potential may include measuring the intracellular membrane potential of the plurality of living neurons using intracellular electrodes.

[0011] The step of understanding the connection structure may include the step of understanding the connection structure between the plurality of living neurons based on the time corresponding to the action potential and the time corresponding to the postsynaptic potential of each of the plurality of living neurons.

[0012] The step of understanding the aforementioned connection structure allows for the determination of the presynaptic neuron and postsynaptic neuron relationships for each of the multiple living neurons.

[0013] The step of estimating the synaptic load allows for the estimation of the synaptic load between the presynaptic neuron and the postsynaptic neuron based on the postsynaptic potential of the postsynaptic neuron and the action potential of the presynaptic neuron.

[0014] The step of mapping the neural network map to the electronic neural network may include the steps of mapping the plurality of biological neurons to the circuit layer of the electronic neural network and mapping the synaptic loads to the memory layer of the electronic neural network.

[0015] The configuration and mapping of the neural network map may include the steps of enabling learning of the electronic neural network device and receiving input or stimuli; activating the learned electronic neural network to which the input or stimuli has been provided and performing electronic neural network operation; and generating a neural network result for the input or stimuli based on the result of the activated electronic neural network.

[0016] A method for generating a neural network result in an electronic device using a learned electronic neural network having learned synaptic connections and synaptic loads having the characteristics of a learned electronic neural network

[0017] A method for generating neural network results according to one embodiment may include the steps of measuring AP using a first plurality of electrodes, measuring PSP using a second plurality of electrodes, and constructing a neural network map of a natural neural network based on the measured AP information interacting with the PSP information measured using the corresponding crosslinks of the crossbar, thereby learning the electronic neural network device.

[0018] The first plurality of electrodes may be used for measurements different from those of the second plurality of electrodes during each first timing interval, and some of the first plurality of electrodes may be used for the same measurements as some of the second plurality of electrodes to measure additional AP or additional PSP during each different second timing interval.

[0019] A method for mapping a biological neural network to an electronic memory network and circuit according to one embodiment includes the steps of considering at least two different forms of membrane potentials measured from multiple biological neurons of a biological neural network using a plurality of neuron modules of an electronic neural network device, and constructing a neural network map of the electronic neural network so that the electronic neural network can mimic the biological neural network.

[0020] The aforementioned consideration step may include considering the interaction between the measured action potential (AP) information and the measured postsynaptic potential (PSP) information for each synaptic sequence between a presynaptic neuron and a postsynaptic neuron.

[0021] The steps of constructing the neural network map may include understanding the connectivity structure between the plurality of neuronal modules and updating the synaptic loads between the plurality of neuronal modules. An electronic neural network according to an embodiment includes one or more memory layers that store a neural network map of a biological neural network, one or more circuit layers that receive signals, activate each of a plurality of neuron modules, and perform signal transmission between the plurality of neuron modules, and connectors that connect between the memory layer and the circuit layer.

[0022] The neural network results of the neural network map stored in the biological neural network can be generated by the execution of the signal transmission.

[0023] Here, when the electronic neural network is a learned electronic neural network, the information of the one or more memory layers and the information of the one or more circuit layers interact with the information of the measured action potential AP and the information of the measured postsynaptic potential (PSP) for the anteroposterior relationship of each synapse between the presynaptic neuron and the postsynaptic neuron of the biological neural network, and may have the characteristics of the electronic neural network mapped in the biological neural network.

[0024] A method for mapping an electronic memory network and a circuit of a biological neural network may include that the connector includes at least one of a through-silicon via (TSV) that penetrates the memory layer and the circuit layer and a micro bump that connects the memory layer and the circuit layer.

[0025] The neural network results of the neural network map stored in the biological neural network are generated by the execution of the signal transmission. When the circuit layer receives a stimulation signal, it can read the synaptic load corresponding to the corresponding neuron module from the memory layer and activate the neuron module.

[0026] The memory layer is realized by a crossbar array architecture, and the synaptic load of the neural network map can be stored at the cross points of the crossbar array.

[0027] The memory layer and the circuit layer may be stacked three-dimensionally.

[0028] A biological neural network mapping device according to an embodiment includes a processor that constructs a neural network map of the biological neural network based on the membrane potential of each of a plurality of biological neurons constituting the biological neural network and maps the neural network map to an electronic neural network, and the membrane potential may have at least two different membrane potential forms.

[0029] The processor can grasp the connection structure between the plurality of biological neurons and estimate the synaptic load between the plurality of biological neurons.

[0030] The processor can map the plurality of biological neurons to the circuit layer of the electronic neural network and map the synaptic load to the memory layer of the electronic neural network.

[0031] A biological neural network mapping device according to an embodiment further includes an electrode that measures the membrane potential of a plurality of biological neurons over time, and the processor can extract the action potential AP of the plurality of biological neurons from the action potential results of the measured membrane potential and extract the postsynaptic potential (PSP) of a plurality of biological neurons from the results of the measured postsynaptic potential on the membrane.

[0032] The construction and mapping of the neural network map are achieved based on each piece of information of the first measured membrane potential that interacts with each piece of information of the second measured membrane potential with respect to the precedence relationship of each synapse between the presynaptic neuron and the postsynaptic neuron of the biological neural network, the first measured membrane potential corresponds to the first form of the membrane potential of the at least two different forms of membrane potential, and the second form of the membrane potential of the at least two different forms of membrane potential.

Effects of the Invention

[0033] According to the present invention, a method and apparatus for electronic memory copy mapping of a biological neural network can be provided. [Brief explanation of the drawing]

[0034] [Figure 1] This is a diagram illustrating a biological neural network mapping system according to one embodiment. [Figure 2] This is a flowchart illustrating a method for mapping a biological nervous system to an electronic memory network and circuit according to one embodiment. [Figure 3] This is a flowchart illustrating a method for mapping a biological nervous system to an electronic memory network and circuit according to another embodiment. [Figure 4] This figure shows the structure of an electronic neural network according to one embodiment. [Figure 5] This is a diagram illustrating the architecture of a crossbar array according to one embodiment. [Modes for carrying out the invention]

[0035] The specific structural or functional descriptions disclosed herein are illustrative for the purpose of illustrating embodiments, and embodiments can be carried out in various different forms. The present invention is not limited to the embodiments described herein, and the scope of the present invention includes modifications, equivalents, or substitutions that are included in the technical ideas described in the embodiments.

[0036] Terms such as "first" or "second" may be used to describe multiple components, but such terms should be interpreted solely for the purpose of distinguishing one component from others. For example, the first component can be named the second component, and similarly, the second component can also be named the first component.

[0037] When it is mentioned that one component is “linked” or “connected” to another component, it should be understood that it is directly linked to or connected to the other component, but that other components may be present in between.

[0038] A singular expression includes plural expressions unless the context clearly indicates otherwise. In this specification, terms such as “includes” or “has” indicate the presence of features, figures, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood not to preemptively exclude the possibility of the presence or addition of one or more other features, figures, steps, actions, components, parts, or combinations thereof.

[0039] Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as those generally understood by a person of ordinary skill in the art to which this embodiment belongs. Commonly used, predefined terms should be interpreted as having the meaning consistent with their meaning in the context of the relevant art, and not as ideal or overly formal unless expressly defined herein.

[0040] The embodiments can be implemented in various forms of products such as personal computers, laptop computers, tablet computers, smartphones, televisions, smart home appliances, intelligent vehicles, kiosks, and wearable devices. The embodiments will be described in detail below with reference to the attached drawings. The same reference numerals shown in each drawing indicate the same component.

[0041] Figure 1 is a diagram illustrating a biological neural network mapping system according to one embodiment.

[0042] Referring to Figure 1, a biological neural network mapping system according to one embodiment can generate a neural network map of the brain that perfectly mimics the structure and function of the brain through recording (measuring) and / or analysis of neural signals occurring in biological neurons of a large neural network such as the brain, or can develop the structure of a new neuromorphic processor through brain copying. For example, the large neural network may be, as an unrestricted example, the brain of an animal or a human, the corresponding nervous system, or other large biological neural networks. The biological neural network 130 may be used to construct a neuromorphic processor by, for example, copying the biological neural network 130, which includes the biological neuronal connection structure and corresponding connection strengths, to the neuromorphic processor. In one embodiment, the neuromorphic processor can be constructed directly using the measurement results of the biological neurons in question, without computational analysis of the potentials measured to identify the respective connections between biological neurons and / or the strengths and weights of such connections. Furthermore, the measurement of nerve signals generated in biological neurons may include other types or forms of membrane potential measurement, such as using minimal intracellular electrodes via intracellular electrode interfaces.

[0043] In the following, the terms "neural network map," "functional map," and "synaptic connectivity map" can be understood as having the same meaning.

[0044] One embodiment of the biological neural network mapping system consists of a recoating unit 110 that measures the membrane potential of individual biological neurons in a large biological neural network 130 in real time, and mapping units 120-1 and 120-2 for constructing electronic neural networks 140-1 and 140-2 that have the same structure as the biological neural network.

[0045] In one embodiment, electronic neural networks 140-1 and 140-2 may mimic the operation of biological neurons in the biological neural network 130. For example, as an example of how electronic neural networks 140-1 and 140-2 are subsequently realized based on input information or stimuli, electronic neural networks 140-1 and 140-2 may be identical or substantially identical to how biological neurons in the natural neural network 130 respond to the same stimulus. Biological neurons mean living nerve cells that are not artificial neurons, and hereafter, the terms "neuron" and "nerve cell" are understood to have the same meaning. Herein, "operation based on biological neurons" may include, but is not limited to, synaptic connection analysis, ion channel analysis, ion channel current measurement, network connection, and measurement of the effect of drugs on mechanics.

[0046] In one embodiment, the recoating portion 110 may include an electrode layer composed of a plurality of electrodes, and can contact a living neuron via the electrodes to record (or measure) nerve signals generated in the living neuron, or to inject (or provide) a stimulating signal to the living neuron.

[0047] For example, the recoating unit 110 according to one embodiment can read out the electrical activity 151 of all individual biological neurons in the biological neural network 130 in real time using CNEA (CMOS NanoElectrode Array) technology.

[0048] In one embodiment, the electrodes of the recoating section 110 can be independently connected to individual biological neurons, allowing for simultaneous measurement of membrane potential. The large amount of data measured in one embodiment may be used to construct a neural map in the mapping section 120-1 through separate signal processing, and then mapped to the electronic neural network 140-1 in the same configuration.

[0049] For example, membrane potential is the potential across the membrane of a biological neuron, and the resting membrane potential may be approximately -70mV. Membrane potential may be induced / stimulated to increase or decrease depending on factors such as ion exchange across the neuron membrane or the reception of neurotransmitters from other biological neurons affecting it. This, in turn, leads to changes in membrane potential.

[0050] If the changing membrane potential satisfies a specific threshold (e.g., approximately -45 mV), the resting membrane potential can cause a biological neuron to generate an action potential (AP), also known as a "nerve impulse" or "spike," through a stepwise change in membrane potential. Here, the release of AP along the axis and terminally of a biological neuron may be referred to as "firing" of the biological neuron. In response to the AP, the biological neuron may release the aforementioned neurotransmitters. Here, the biological neuron that releases neurotransmitters is called the presynaptic neuron, and the subsequent neuron that receives the neurotransmitters is called the postsynaptic neuron. The reception of neurotransmitters by the postsynaptic neuron can also be reflected in the change in the membrane potential of the postsynaptic neuron, which is called the postsynaptic potential (PSP). Therefore, AP and PSP are different forms or types of membrane potentials.

[0051] Therefore, not only neurotransmitters received by postsynaptic neurons from presynaptic neurons, but also neurotransmitters received by postsynaptic neurons from other presynaptic neurons, such as the membrane potential of a postsynaptic neuron, may repeatedly satisfy the aforementioned thresholds to generate APs in the postsynaptic neuron. The temporal sequence of such APs generated by biological neurons may be referred to as a "spike train." Thus, as a non-restrictive example, the coupling weight or strength between a presynaptic neuron and a postsynaptic neuron is validated by a determined relationship between the AP of the presynaptic neuron and the PSP of the postsynaptic neuron.

[0052] Herein, the above description concerns the general neuronal synaptic relationships to different forms of membrane potential, e.g., AP and PSP neuronal signals, but the above discussion is merely illustrative. The disclosure herein, relating to the connection between presynaptic and postsynaptic neurons for such information sharing between presynaptic and postsynaptic neurons, is also applicable to other neuronal types having different operations from each other (e.g., synaptic neurons measurable by intercellular electrodes).

[0053] Referring again to Figure 1, the large amount of data measured by one embodiment may be transmitted directly to the mapping unit 120-2 in real time, and the synaptic weights of the corresponding electronic neural network 140-2 may be mapped based on the interelectrical activity between adjacent neurons.

[0054] The following describes, with reference to Figure 2, the specific operation of constructing a neural network map via a separate signal processing method according to one embodiment and copying the neural network map to an electronic neural network. The specific operation of directly transmitting the nerve signals extracted by one embodiment to an electronic neural network and constructing a neural network map within the electronic neural network itself will be described with reference to Figure 3.

[0055] Figure 2 is a flowchart illustrating a method for mapping an electronic memory network of a biological neural network to a circuit according to one embodiment.

[0056] The biological neural network program method according to one embodiment is executed by the mapping unit 120-1 described above with reference to Figure 1. The mapping unit 120-1 according to one embodiment may be implemented by one or more hardware modules, one or more software modules, or various combinations thereof. Alternatively, the mapping unit 120-1 according to one embodiment may be a separate external device (for example, a personal computer) that is separate from the electronic neural network 140-1.

[0057] In one embodiment, the mapping unit 120-1 may analyze the collected data to construct a neural network map, and then map the electronic neural network 140-1 with the same configuration. When the neural network map according to one embodiment is accurately constructed, the individual weighted values ​​of the corresponding electronic neural network 140-1 can be accurately extracted.

[0058] An electronic device according to one embodiment may or may not include a mapping unit 120-2 and an electronic neural network 140-2, which will be discussed in more detail with reference to Figure 3. Alternatively, an exemplary electronic device may include the mapping unit 120-2 and the electronic neural network 140-2, but may not include the mapping unit 120-1 and the electronic neural network 140-1.

[0059] Referring to Figure 2, in step S210, the mapping unit 120-1 according to one embodiment constructs a neural network map of the biological neural network based on the membrane potential of each of the multiple biological neurons constituting the biological neural network. The mapping unit 120-1 according to one embodiment may extract action potentials (APs) and postsynaptic potentials (PSPs) of the relevant biological neurons based on their membrane potential. Alternatively, the mapping unit 120-1 according to one embodiment may directly receive the extracted action potentials and postsynaptic potentials of the biological neurons. The extraction of APs and PSPs may, alternatively, be performed before step S210 of the mapping unit 120-1, for example, by an exemplary circuit of the recoating unit 110.

[0060] According to one embodiment, the mapping unit 120-1 can first identify the connectivity structure between multiple living neurons based on the received membrane potential. Identifying the connectivity structure between living neurons includes identifying the pre-post-relationship (presynaptic neuron and postsynaptic neuron) between the living neurons.

[0061] More specifically, the mapping unit 120-1 according to one embodiment can analyze the relationship between the action potential and postsynaptic potential of a living neuron and distinguish adjacent cells. For example, if the time interval between the action potential of the first living neuron and the postsynaptic potential of the second living neuron occurs continuously at or below a threshold level, the mapping unit 120-1 according to one embodiment determines that the first living neuron and the second living neuron are related in terms of sequence.

[0062] After understanding the connectivity structure between living neurons, the mapping unit 120-1 according to one embodiment can estimate the synaptic weight between multiple living neurons. In this embodiment, the synaptic weight represents the connectivity strength between a presynaptic neuron and a postsynaptic neuron.

[0063] In one embodiment, the mapping unit 120-1 can estimate the synaptic load between related biological neurons by analyzing the correlation between the magnitude of the postsynaptic potential (Amplitude) and the magnitude of the action potential of previous biological neurons when a postsynaptic potential is generated in the reference biological neuron. For example, the mapping unit 120-1 in one embodiment can estimate the synaptic load based on the magnitude of the postsynaptic potential of the reference biological neuron and the magnitudes of the action potentials of n previous biological neurons connected to the reference biological neuron.

[0064] In step S220, the mapping unit 120-1 according to one embodiment maps the neural network map to the electronic neural network 140-1. The mapping unit 120-1 according to one embodiment can combine a neural network map constructed based on the membrane potential of biological neurons with the electronic neural network 140-1 using the same configuration.

[0065] As will be explained in detail below, the electronic neural network 140-1 according to one embodiment may include a memory layer for storing synaptic loads and a circuit layer for performing the operations of biological neurons. The mapping unit 120-1 according to one embodiment may map a plurality of biological neurons to the circuit layer of the electronic neural network 140-1 and map the synaptic loads to the memory layer of the electronic neural network 140-1.

[0066] Figure 3 is a flowchart illustrating a method for mapping a biological neural network to an electronic memory network and circuit according to another embodiment.

[0067] A biological neural network program according to one embodiment can be executed by the biological neural network mapping system described above with reference to Figure 1. The recoating unit 110, mapping unit 120-2, and electronic neural network 140-2 according to one embodiment may be implemented by one or more hardware modules, one or more software modules, or various combinations thereof.

[0068] In one embodiment, the mapping unit 120-2 may directly transmit the membrane potential of the biological neural network, measured in real time, to the electronic neural network 140-2 to map synaptic loads. The electronic neural network 140-2 mapped via the mapping unit 120-2 in one embodiment may learn the membrane potentials collected from the biological neural network and replicate the connectivity structure of the original biological neural network or mimic its behavior. More specifically, the electronic neural network 140-2 mapped via the mapping unit 120-2 in one embodiment can directly mimic the response that the target neural network shows to a specific stimulus, using only time-series membrane potential information of some neurons measured from the biological neural network, without information on the number of neurons not measured in the target biological neural network, the degree of connectivity between each neuron, etc.

[0069] Referring to Figure 3, in step S310, the mapping unit 120-2 according to one embodiment transmits the membrane potential of each of the multiple biological neurons constituting the biological neural network to the electronic neural network 140-2, which is composed of multiple neuron modules. The neuron modules of the electronic neural network 140-2 according to one embodiment are the basic units for constituting the electronic neural network 140-2 and correspond to the biological neurons of the biological neural network.

[0070] In step S320, the electronic neural network 140-2 according to one embodiment configures a neural network map so that the electronic neural network 140-2 can mimic a biological neural network. The electronic neural network 140-2 according to one embodiment may configure the neural network map within the electronic neural network 140-2 itself without using a separate external device to configure the neural network map. For this purpose, the electronic neural network 140-2 according to one embodiment may include a processor for configuring the neural network map.

[0071] Constructing a neural network map according to one embodiment involves understanding the connectivity structure between multiple neuron modules and updating the synaptic loads between multiple neuron modules. The neuron module circuit device according to one embodiment may update the synaptic load values ​​via STDP learning.

[0072] An electronic neural network 140-2 according to one embodiment may include a pulse conversion unit that converts the action potential and postsynaptic potential of a biological neuron into memory writing pulses at fixed time intervals, and a delay conversion unit that adjusts the interval between pulses inversely proportional to the magnitude of the postsynaptic potential. Furthermore, the pulses generated by the method according to one embodiment can be transmitted to a processor to adjust the conductance of the target crosspoint.

[0073] An electronic neural network according to one embodiment can change synaptic load values ​​between neuron modules via STDP (Spike Timing Dependent Plasticity) learning. In this embodiment, the electronic neural network maps the time interval between two neuron modules to be smaller as the connection strength between the two connected neuron modules increases, and maps the connection strength according to the STDP properties of the RRAM.

[0074] In one embodiment, the amount of change in synaptic load may be changed by a value determined through a simple comparator as needed, or it may be selected from several values ​​according to the difference in firing time via a LUT method. That is, the weighting value update of a synaptic module may occur on its own through signal transmission between adjacent neuron modules. The method for updating the synaptic load value in one embodiment is not limited to the method described above, and the synaptic load value can be updated by various other methods.

[0075] Figure 4 shows the structure of an electronic neural network according to one embodiment.

[0076] Referring to Figure 4, an electronic neural network according to one embodiment includes one or more memory layers 410, one or more circuit layers 420, and a connecting element 430. The contents described with reference to Figures 1 to 3 also apply to Figure 4, and for the sake of convenience, redundant content is omitted.

[0077] In one embodiment, the memory layer 410 may store a neural network map of a biological nervous system. The memory layer 410 may also store synaptic loads between biological neurons. For example, the memory layer 410 may store the synaptic load between the i-th biological neuron and the j-th biological neuron.

[0078] In one embodiment, the memory layer 410 must store all synaptic loads between living neurons acquired via a neural network map. In a biological neural network, one nerve cell has K (e.g., 1000) synaptic connections, and N (e.g., 10 9 If there are (1000) nerve cells, then the memory layer 410 will have K × N / 2 (for example, 1000 × 10 9 The system must store 2 / 2 synaptic loads. Therefore, a memory architecture capable of efficiently storing large amounts of data is required. Below, with reference to Figure 5, we will describe a crossbar array architecture as an example of a memory layer.

[0079] Referring to Figure 5, a memory layer according to one embodiment may be realized from the architecture of a crossbar array 500. The crossbar array 500 includes a first electrode 510 formed from a plurality of rows on a substrate, a second electrode 520 composed of a plurality of rows intersecting the first electrode 510, and a memory section 530 formed between the first electrode 510 and the second electrode 520, the resistance of which changes in response to the voltage applied between the first electrode 510 and the second electrode 520.

[0080] In one embodiment, the mapping unit 120-1 maps each of the multiple biological neurons constituting the biological neural network to the first electrode 510 and the second electrode 520, and maps the synaptic loads between the biological neurons to the memory unit 530. For example, the mapping unit 120-1 may map each of the N biological neurons constituting the biological neural network to the first electrode 510 and the second electrode 520, which are formed from N rows. Subsequently, the synaptic load between the i-th biological neuron and the j-th biological neuron is stored in the memory unit 530 located at the cross point between the first electrode 510 corresponding to the i-th biological neuron and the biological neuron corresponding to the j-th biological neuron. Here, the mapping unit 120-1 can adjust the variable resistance value of the memory unit 530 and store the synaptic loads in the memory unit 530.

[0081] N pieces (for example, 10 9An N×N crossbar array 500 may be used to store synaptic loads between K (e.g., 1000) biological neurons. However, as mentioned above, since one biological neuron has K (e.g., 1000) synaptic connections, many regions of the N×N crossbar array 500 may not be used. In one embodiment, the mapping unit 120-1 knows the relationships between biological neurons that are actually connected via the neural network map, so a crossbar array with crosspoints that do not have synaptic loads excluded can be used as a memory layer. This can increase the efficiency of the memory layer.

[0082] However, the architecture of the memory layer 410 according to one embodiment is not necessarily limited to a crossbar array 500.

[0083] Referring again to Figure 4, the circuit layer 420 according to one embodiment can receive a signal, activate each of the multiple neuron modules, and transmit signals between the multiple neuron modules.

[0084] In one embodiment, the circuit layer 420 may include stacked circuits, which may perform functions such as measuring, processing, and analyzing nerve signals. The circuits with the aforementioned functions may be implemented in a distributed manner across various circuit layers, or they may be implemented in a manner integrated into any one of the circuit layers. The circuits may be, for example, CMOS (Complementary Metal-Oxide Semiconductor) integrated circuits (ICs), but are not necessarily limited to these. In order to directly copy the interconnected structure of a large-scale biological nervous network, an electronic nervous network structure of the same size as the biological nervous network is required. In one embodiment, the electronic nervous network can be configured as a 3D stacked system to increase the integration density.

[0085] More specifically, the memory layer 410 according to one embodiment may include, for example, memory layer 1, memory layer 2, ..., memory layer L. Each memory layer may be stacked perpendicular to the others.

[0086] Similarly, a circuit layer 410 according to one embodiment may include, for example, circuit layer 1, circuit layer 2, ..., circuit layer M. Each circuit layer may be stacked perpendicularly to one another. Each circuit layer may include circuits that perform different functions or operations from each other. For example, circuit layer 1 may include a circuit for accumulation, circuit layer 2 may include a circuit for fire, and circuit layer M may include a circuit for voltage amplification.

[0087] Alternatively, each circuit layer may include all circuits that perform the same function or operation as each other. For example, circuit layer 1 may include circuits for accumulation and circuits for firing, and circuit layer 2 may also include circuits for accumulation and circuits for firing, similar to circuit layer 1.

[0088] In one embodiment, the connecting element 430 may connect the memory layer 410 and the circuit layer 420. The connecting element 430 may be, for example, at least one of a through-silicon via (TSV) that penetrates the memory layer 410 and the circuit layer 420, and a microbump that connects the memory layer 410 and the circuit layer 420.

[0089] One embodiment of a silicon through-silicon (TSV) is a packaging technology that connects an upper and lower chip to an electrode by passing a fine pore (via) through the end and filling the pore with conductive material, instead of connecting the ends with a wire. Because the silicon through-silicon (TSV) does not require additional space by securing a direct electrical connection path within the end, the package size can be reduced and the length of the interconnection between the ends can be reduced.

[0090] In one embodiment, when the circuit layer 420 receives a stimulus signal, it can read the corresponding synaptic load from the memory layer 410, which stores the synaptic load corresponding to the neuron module, and activate the neuron module. Here, the coupling element 430 can transmit signals between the memory layer 410 and the circuit layer 420.

[0091] One embodiment of a biological neural network mapping device may include a processor that receives the membrane potential of each of a plurality of biological neurons constituting a biological neural network, constructs a neural network map of the biological neural network based on the membrane potential, and maps the neural network map to an electronic neural network.

[0092] A processor according to one embodiment can grasp the connection structure between multiple biological neurons and estimate the synaptic load between multiple biological neurons.

[0093] One embodiment of the processor can map multiple biological neurons to the circuit layer of an electronic neural network and map synaptic loads to the memory layer of the electronic neural network.

[0094] The embodiments described above are embodied in hardware components, software components, or combinations of hardware and software components. For example, the devices and components described in these embodiments are embodied using one or more general-purpose or special-purpose computers, such as a processor, controller, ALU (arithmetic logic unit), digital signal processor, microcomputer, FPA (field programmable array), PLU (programmable logic unit), microprocessor, or different devices that execute and respond to instructions. The processing device can run an operating system (OS) and one or more software applications run on the OS. The processing device can also access, store, manipulate, process, and generate data in response to software execution. For convenience of understanding, the processing device may sometimes be described as being used as a single unit, but a person with ordinary skill in the art will see that the processing device includes multiple processing elements and / or multiple types of processing elements. For example, the processing device may include multiple processors or one processor and one controller. Other processing configurations, such as parallel processors, are also possible.

[0095] Software includes computer programs, code, instructions, or a combination of one or more of these, which can configure a processing unit to operate as desired, or instruct the processing unit independently or in combination. Software and / or data can be permanently or temporarily embodied in any type of machine, component, physical device, virtual device, computer storage medium or device, or transmitted signal wave, for interpretation by a processing unit or for providing instructions or data to a processing unit. Software can be distributed across a network of computer systems and stored and executed in a distributed manner. Software and data can be stored on a recording medium readable by one or more computers.

[0096] The methods according to the embodiments are embodied in the form of program instructions that are implemented via various computer means and recorded on a computer-readable recording medium. The recording medium includes program instructions, data files, data structures, etc., individually or in combination. The recording medium and program instructions may be specifically designed and configured for the purposes of the present invention, or they may be known and usable by those skilled in the art who have technology in the field of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROMs and DVDs, magneto-optical media such as floppy disks, and hardware devices specifically configured to store and execute program instructions, such as ROMs, RAMs, and flash memory. Examples of program instructions include not only machine code such as that generated by a compiler, but also high-level language code that is executed by a computer using an interpreter or the like. The hardware devices described above may be configured to operate as one or more software modules to perform the operations shown in the present invention, and vice versa.

[0097] As described above, although embodiments have been illustrated with limited drawings, a person with ordinary skill in the art can apply various technical modifications and variations based on the above description. For example, the described techniques may be performed in a different order than described, and / or the components of the described systems, structures, devices, circuits, etc. may be combined or assembled in a different manner than described, or replaced or substituted with other components or equivalents, and still achieve suitable results.

[0098] Therefore, the scope of the present invention is not limited to the disclosed embodiments, but is defined by the claims and equivalents thereof. [Explanation of Symbols]

[0099] 110 Recoating section 120-1, 120-2 Mapping section 130 Large-scale biological nervous networks 140-1, 140-2 Electronic nerve 410 Memory Layers 420 Circuit Layers 430 Linking elements 510 1st electrode 520 2nd electrode 530 Memory section

Claims

1. The steps include constructing a neural network map of the biological neural network based on the membrane potential of each of the multiple biological neurons that make up the biological neural network, The steps include mapping the aforementioned neural network map onto an electronic neural network, Includes, The membrane potential has at least two different membrane potential configurations. A method for mapping biological neural networks to electronic memory networks and circuits.

2. The structure and mapping of the aforementioned neural network map are as follows: The anterior-posterior relationship of each synapse between presynaptic and postsynaptic neurons in a biological neural network is achieved based on the information of the first measured membrane potential interacting with the information of the second measured membrane potential. The first measured membrane potential corresponds to the first form of the membrane potential of the at least two different forms of membrane potential, The second measured membrane potential corresponds to the second form of the membrane potential of the at least two different forms of membrane potential, A method for mapping a biological neural network to an electronic memory network and circuit, as described in claim 1.

3. The step of constructing the neural network map is: The steps include: understanding the connection structure between the aforementioned plurality of living neurons, The steps include: estimating the synaptic weight between the plurality of biological neurons; A method for mapping a biological neural network to an electronic memory network and circuit, as described in claim 1, including the above.

4. The step of estimating the synaptic load includes the step of estimating the synaptic load based on the result of understanding the connection structure. A method for mapping a biological neural network to an electronic memory network and circuit, as described in claim 3.

5. The steps include measuring the membrane potential of the plurality of biological neurons over time, Based on the measurement results of the membrane potential, the steps include extracting the action potentials (APs) and postsynaptic potentials (PSPs) of each of the multiple living neurons, A method for mapping a biological neural network to an electronic memory network and circuit, further comprising claim 3.

6. The method for mapping a biological neural network to an electronic memory network and circuit, according to claim 5, wherein the step of measuring the membrane potential includes measuring the intracellular membrane potential of the plurality of biological neurons using intracellular electrodes.

7. The step of understanding the connection structure includes the step of understanding the connection structure between the plurality of living neurons based on the time corresponding to the action potential and the time corresponding to the postsynaptic potential of each of the plurality of living neurons. A method for mapping a biological neural network to an electronic memory network and circuit, as described in claim 5.

8. The step of understanding the aforementioned connection structure includes the step of determining the relationship between the presynaptic neuron and the postsynaptic neuron for each of the plurality of living neurons. A method for mapping a biological neural network to an electronic memory network and circuit, as described in claim 3.

9. The step of estimating the synaptic load includes estimating the synaptic load between the presynaptic neuron and the postsynaptic neuron based on the postsynaptic potential of the postsynaptic neuron and the action potential of the presynaptic neuron. A method for mapping a biological neural network to an electronic memory network and circuit, as described in claim 8.

10. The step of mapping the neural network map onto the electronic neural network is: The steps include mapping the plurality of biological neurons to the circuit layer of the electronic neural network, The steps include mapping the synaptic load and connection relationship to the memory layer of the electronic neural network, including, A method for mapping a biological neural network to an electronic memory network and circuit, as described in claim 3.

11. The structure and mapping of the aforementioned neural network map are as follows: To enable learning of electronic neural network devices, A step of receiving input or stimulus, The steps include activating the learned electronic neural network to which the input or stimulus is provided and performing electronic neural network operation, A step of generating a neural network result for the input or stimulus based on the results of the activated electronic neural network, including, A method for mapping a biological neural network to an electronic memory network and circuit, as described in claim 1.

12. A computer program stored in a medium for use in conjunction with hardware to perform the method described in claim 1.

13. A method for generating a neural network result by an electronic device using a learned electronic neural network device having learned synaptic connections and synaptic loads, wherein the characteristics of the learned electronic neural network device are mapped from the natural neural network based on information of measured action potentials (APs) and information of measured postsynaptic potentials (PSPs) interacting with the action potentials in each presynaptic relationship between presynaptic and postsynaptic neurons of the natural neural network, and the method is: A step of receiving input or stimulus, The steps include activating the learned electronic neural network in response to the provision of the input or stimulus to perform electronic neural network operation, A step of generating a neural network result for the input or stimulus based on the results of the activated electronic neural network, A method for generating neural network results, including the above.

14. The first step is to measure AP using multiple electrodes, A second step involves measuring PSP using multiple electrodes, The steps include: constructing a neural network map of the natural neural network and learning the electronic neural network device based on the measured AP information that interacts with each PSP information measured using the corresponding crosslink of the crossbar; A method for generating neural network results according to claim 13, further comprising:

15. The first plurality of electrodes are used for different measurements from the second plurality of electrodes during their respective first timing intervals. Some of the first plurality of electrodes are used for the same measurements as some of the second plurality of electrodes to measure additional AP and / or additional PSP during different second timing intervals. A method for generating neural network results according to claim 14.

16. A computer program stored in a medium for use in conjunction with hardware to perform the method described in claim 13.

17. The steps include considering at least two distinct forms of membrane potential measured from multiple living neurons of a biological neural network using multiple neuronal modules of an electronic neural network device, The steps include: constructing a neural network map of the electronic neural network so that the electronic neural network can mimic the biological neural network; A method for mapping biological neural networks to electronic memory networks and circuits, including the above.

18. The method for mapping a biological neural network to an electronic memory network and circuit, according to claim 17, wherein the aforementioned consideration step includes considering the interaction between each piece of information on action potentials (APs) measured for each synaptic sequence between a presynaptic neuron and a postsynaptic neuron, and the measured postsynaptic potential (PSP) information.

19. The step of constructing the neural network map is: The steps include: understanding the connection structure between the plurality of neuronal modules, The steps include updating the synaptic load between the aforementioned multiple neuronal modules, A method for mapping a biological neural network to an electronic memory network and circuit, as described in claim 17, including the above.

20. A computer program stored in a medium for use in conjunction with hardware to perform the method described in claim 17.

21. One or more memory layers that store neural network maps of biological neural networks, One or more circuit layers that receive signals, activate each of the multiple neuron modules, and transmit signals between the multiple neuron modules, A connecting element that connects the memory layer and the circuit layer, An electronic nervous system, including an electronic network.

22. The electronic neural network according to claim 21, wherein the neural network result of the neural network map stored in the biological neural network is generated by the execution of the signal transmission.

23. The electronic neural network according to claim 21, wherein, if the electronic neural network is a learned electronic neural network, the information of one or more memory layers and the information of one or more circuit layers have the characteristics of the electronic neural network mapped in the biological neural network, based on the measured information of postsynaptic potentials (PSPs) for each synaptic relationship between a presynaptic neuron and a postsynaptic neuron of the biological neural network and the measured information of action potentials (APs) interacting with the measured information of postsynaptic potentials (PSPs) for each synaptic relationship between a presynaptic neuron and a postsynaptic neuron of the biological neural network.

24. The aforementioned connecting element is A through-silicon via (TSV) penetrates the memory layer and the circuit layer, Microbumps connecting the memory layer and the circuit layer, The electronic neural network according to claim 21, comprising at least one of the above.

25. The neural network result of the neural network map stored in the biological neural network is generated by the execution of the signal transmission, The electronic neural network according to claim 21, wherein when the circuit layer receives a stimulus signal, it reads from the memory layer the synaptic load corresponding to the neuron module corresponding to the stimulus signal and activates the corresponding neuron module.

26. The electronic neural network according to claim 21, wherein the memory layer is implemented in a crossbar array architecture, and the synaptic loads of the neural network map are stored at the cross points of the crossbar array.

27. The electronic neural network according to claim 21, wherein the memory layer and the circuit layer are stacked in three dimensions.

28. The processor includes a processor that constructs a neural network map of a biological neural network based on the membrane potential of each of several biological neurons that make up the biological neural network, and maps the neural network map to an electronic neural network. The aforementioned membrane potential is a biological neural network mapping device having at least two different membrane potential morphologies.

29. The biological neural network mapping device according to claim 28, wherein the processor grasps the connection structure between the plurality of biological neurons and estimates the synaptic load between the plurality of biological neurons.

30. The biological neural network mapping device according to claim 29, wherein the processor maps the plurality of biological neurons to the circuit layer of the electronic neural network and maps the synaptic loads to the memory layer of the electronic neural network.

31. The system further includes electrodes for measuring the membrane potential of multiple living neurons over time, The biological neural network mapping device according to claim 28, wherein the processor extracts action potentials AP of the plurality of biological neurons from the measured action potential results of the membrane potential, and extracts postsynaptic potentials (PSPs) of the plurality of biological neurons from the measured postsynaptic potential results on the membrane.

32. The structure and mapping of the aforementioned neural network map are as follows: The presynaptic relationship between presynaptic and postsynaptic neurons in a biological neural network is determined based on the information from the first measured membrane potential interacting with the information from the second measured membrane potential. The biological neural network mapping device according to claim 28, wherein the first measured membrane potential corresponds to a first form of membrane potential of the at least two different forms of membrane potential, and corresponds to a second form of membrane potential of the at least two different forms of membrane potential.

Images