Neuromorphic-based optimal network path search method, system, and storage medium

Abstract

The application discloses a morphological optimal network path searching method and system and a storage medium. The method comprises the following steps: constructing a pulse neural network and a morphological computing core; determining a starting neuron and a target neuron of the pulse neural network according to a problem to be solved; sending an excitation signal conforming to a SFA rule through the morphological computing core to make the starting neuron emit a plurality of clock pulses to another neuron connected to the starting neuron, so that a transmission waveform characterized by an attractor wave peak is transmitted in the pulse neural network, and the synapses of the neurons on the transmission path are enhanced or weakened according to the STDP rule; when the transmission waveform reaches the target neuron, a first network path passed by the transmission waveform is determined, and then an optimal network path is selected from a plurality of first network paths. The application improves the efficiency and accuracy of the optimal network path searching, and has a wider application range, and can be widely applied to the field of artificial intelligence technology.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a neuromorphic optimal network path search method, system, and storage medium. Background Technology

[0002] From autonomous driving finding the best path to autonomous vehicles finding parking spaces, from robots avoiding obstacles to navigation on handheld devices based on landmark images (such as Google Street View Time Machine), and then to accessing and searching on internet social platforms, the guarantee of their functions and performance is based on modeling the environment as network relationships, traversing and searching the network, and then using the results to solve problems related to product performance or internet services. Products using these functions are required to be low in power consumption, fast in time, and provide the best results.

[0003] Most solutions to the aforementioned problems rely on current von Neumann architecture computers and software search methods based on these platforms. These methods suffer from high power consumption, and their speed and service experience are not entirely satisfactory. While neuromorphic computing technologies, particularly those based on brain-like intelligence, could leverage the energy efficiency and parallel computing advantages of such technologies, current solutions are largely limited to their specific application or professional domains. Finding a more unified approach to implementing these applications would be of great significance.

[0004] The first to achieve a breakthrough in web search based on a non-von Neumann architecture were major international companies using neuromorphic computing technology. Their solutions to the search problem were inspired by the "neural computation" of "biological space" and "map-aware networks".

[0005] As early as 1940, Edward Chace Tolman hypothesized the existence of a "map-like" information storage mechanism in the brain, used to guide human spatial cognition and searching, and to establish instantaneous connections between "different locations" and "event occurrences." The foundation of cognitive maps is place cells, first discovered by Keefe et al. using electrophysiological methods in the hippocampus. Redish and Hopfield et al. further discovered that the hippocampus in rodents possesses spatial navigation and search functions. In the theoretical research on the construction and application of cognitive maps, there are two main neural network models: the first is the place map, used to represent the agent's position; the second is the target vector map, used to represent the target's pointing location.

[0006] Wilson et al. experimentally recorded the firing patterns of animal place cells during spatial exploration, clearly demonstrating how place cells express their positions in a two-dimensional plane, thus proving the effectiveness of group coding. This theory emphasizes that concentrated firing of place codes is connected through interactions between closely packed neurons. Other experimental studies have further demonstrated that group coding can be reorganized according to environmental changes. Building on this work, Muller et al. further proposed that simultaneously firing place cells can form a network, allowing interacting place cells to form tight synaptic connections. Inspired by this, in July 2019, Intel announced that its neuromorphic research chip, Loihi, can perform dedicated tasks 1000 times faster and 10000 times more efficiently than ordinary CPUs. Its network search algorithm, based on synaptic temporal plasticity, performs more than 100 times faster than the Intel Xeon 6136 CPU in small-world networks with 100,000 nodes, demonstrating exceptional search performance. However, the publicly available method only completes a simplified, single-layer network. The synaptic triggering pulses used for network search fail to abstract and represent the generation and propagation of "bump attractors" (also known as bell-shaped pulses in some Chinese literature) in the neurodynamics of the human brain within the model. Therefore, the method suffers from a lack of broad applicability to objective modeling issues. Furthermore, if the encoding patterns of group neurons cannot be better abstracted during synaptic pulse transmission, the method is susceptible to environmental interference, and its reliability needs improvement. In addition, the application scope of this method is limited to single-layer network search, while other similar domestic methods have even narrower applicability and less universality. Summary of the Invention

[0007] The purpose of this invention is to at least partially solve one of the technical problems existing in the prior art.

[0008] Therefore, one objective of this invention is to provide a neuromorphic optimal network path search method that improves the efficiency and accuracy of optimal network path search and has a wider range of applications.

[0009] Another objective of this invention is to provide a neuromorphic optimal network path search system.

[0010] To achieve the above-mentioned technical objectives, the technical solutions adopted in the embodiments of the present invention include:

[0011] In a first aspect, embodiments of the present invention provide a neuromorphological-based optimal network path search method, comprising the following steps:

[0012] A spiking neural network and a neuromorphic computing kernel are constructed. The spiking neural network is a neuronal network layer system with multiple levels of feature expression, and there are multiple neurons at the same level. The neurons have multiple synapse types to form bidirectional connections between each neuron.

[0013] The starting and target neurons of the spiking neural network are determined based on the problem to be solved.

[0014] The neuromorphic computing kernel sends instructions to run excitation signals that conform to the neuronal synaptic adaptation (SFA) rule, causing the starting neuron to emit several bell-shaped pulses to another neuron connected to it, thereby forming a transmission waveform characterized by attractor peaks that is transmitted in the spiking neural network, and causing the neuronal synapses on the propagation path to be enhanced / weakened according to the STDP rule.

[0015] When the transmitted waveform reaches the target neuron, the first network path through which the transmitted waveform passes is determined, and then the optimal network path is selected from multiple first network paths.

[0016] Furthermore, in one embodiment of the present invention, the step of constructing a spiking neural network specifically includes:

[0017] Initialize the spiking neural network and determine several first nodes in a single-layer interconnected network whose fan-out number is greater than 1;

[0018] For the first node, a preset second node is inserted between each fan-out branch and the next node to form a new spiking neural network. The second node conforms to the definition of SNN and does not affect the overall network characteristics of the spiking neural network.

[0019] The newly formed spiking neural network is assigned network parameters according to the preset modeling requirements, and then the spiking neural network is mapped to obtain the spiking neural network.

[0020] Furthermore, in one embodiment of the present invention, the neuromorphic computing kernel includes at least one of a feature extraction network, a short-term memory network, an episodic memory network, a semantic memory network, a knowledge graph network, a deep learning network, and a recursive network.

[0021] Furthermore, in one embodiment of the present invention, the synaptic input of the dendrites of the spiking neural network includes input currents of excitatory neurons and input currents of inhibitory neurons.

[0022] Furthermore, in one embodiment of the present invention, the optimal path search method further includes the following steps:

[0023] According to the preset model requirements, a neuromodulation signal with global parameters is emitted, causing the system parameters of the neuron to change.

[0024] Furthermore, in one embodiment of the present invention, the excitation signal is generated through the following steps:

[0025] Constructing short-term memory networks to achieve short-term memory function;

[0026] Constructing an episodic memory network to realize episodic memory function;

[0027] The semantic memory network has been constructed to realize the semantic memory function;

[0028] A knowledge graph network is constructed based on the entity relationships in the environment in which the agent is located, and the knowledge graph network is driven by short-term memory output, contextual memory output, or semantic memory output to obtain the corresponding memory embedding vector.

[0029] The memory embedding vector is input into a pre-constructed recurrent network to identify the hidden states, and a feature sequence is formed based on the hidden states. Then, the excitation signal is generated based on the feature sequence.

[0030] Furthermore, in one embodiment of the present invention, the optimal network path search method further includes the step of training the short-term memory network, the contextual memory network, and the semantic memory network, specifically including:

[0031] Establish geospatial location sequences and time-series image sequences and generate corresponding input data;

[0032] Determine the hierarchical transformation relationship and loss function of the short-term memory network, the episodic memory network, and the semantic memory network;

[0033] The input data is sequentially input into the short-term memory network, the contextual memory network, and the semantic memory network. The parameters of the short-term memory network, the contextual memory network, and the semantic memory network are updated through the backpropagation algorithm to obtain the trained short-term memory network, the contextual memory network, and the semantic memory network.

[0034] Secondly, embodiments of the present invention provide a neuromorphic-based optimal network path search system, comprising:

[0035] A spiking neural network construction module is used to construct a spiking neural network and a neuromorphic computing kernel. The spiking neural network is a neuronal network layer system with multiple levels of feature expression, and there are multiple neurons at the same level. The neurons have multiple synapse types to form bidirectional connections between each neuron.

[0036] A neuron determination module is used to determine the starting neuron and target neuron of the spiking neural network based on the problem to be solved.

[0037] The waveform transmission module is used to send instructions through the neuromorphic computing kernel to run an excitation signal that conforms to the SFA rule, so that the starting neuron emits several bell-shaped pulses to another neuron connected to it, thereby forming a transmission waveform characterized by attractor peaks that is transmitted in the spiking neural network, and strengthening / weakening the neuronal synapses on the propagation path according to the STDP rule.

[0038] The path selection module is used to determine the first network path through which the transmission waveform passes when the transmission waveform reaches the target neuron, and then select the optimal network path from multiple first network paths.

[0039] Thirdly, embodiments of the present invention provide a neuromorphological-based optimal network path search device, comprising:

[0040] At least one processor;

[0041] At least one memory for storing at least one program;

[0042] When the at least one program is executed by the at least one processor, the at least one processor implements the above-described neuromorphic optimal network path search method.

[0043] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to perform the aforementioned neuromorphic-based optimal network path search method.

[0044] The advantages and beneficial effects of the present invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention:

[0045] This invention, based on the adaptive frequency of synaptic discharge pulses (SFA) in biological neural networks, rapidly and effectively promotes the exploration of wave packet pulses (bell-shaped pulses) generated by a group of neurons within the neural network. Through the collaborative excitation response of a large number of neurons, a transmission waveform characterized by attractor peaks is formed, effectively filtering out noise factors generated during propagation and improving the accuracy of optimal network path search. The use of a multi-level feature-expressed neural network layer system better represents the problem model, broadening the applicability of network path search and better meeting quantitative constraints such as space, time, and security, thus improving the efficiency and reliability of optimal network path search. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the embodiments of the present invention are described below. It should be understood that the drawings described below are only for the convenience of clearly describing some embodiments of the technical solutions of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 A flowchart illustrating the steps of a neuromorphic-based optimal network path search method provided in an embodiment of the present invention;

[0048] Figure 2 A schematic diagram of a computing system provided in an embodiment of the present invention;

[0049] Figure 3 A schematic diagram of a neuromorphic computing device provided in an embodiment of the present invention;

[0050] Figure 4 A schematic diagram of the image feature extraction process provided in an embodiment of the present invention;

[0051] Figure 5 This is a schematic diagram illustrating the correspondence between images and neurons provided in an embodiment of the present invention.

[0052] Figure 6 A schematic diagram illustrating the generation process of the excitation signal provided in an embodiment of the present invention;

[0053] Figure 7(a) is a schematic diagram of a neuron learning and propagation process provided in an embodiment of the present invention;

[0054] Figure 7(b) is a schematic diagram of another neuron learning and propagation process provided by an embodiment of the present invention;

[0055] Figure 8 This is a schematic diagram of the offset of a bell-shaped pulse wave provided in an embodiment of the present invention;

[0056] Figure 9 This is a schematic diagram of neuron connectivity provided in an embodiment of the present invention;

[0057] Figure 10 A schematic diagram of a partial portion of a network reflecting the objective world, provided in an embodiment of the present invention;

[0058] Figure 11 This is a schematic diagram of neuron signal propagation provided in an embodiment of the present invention;

[0059] Figure 12 This is a schematic diagram of neuron connectivity after the STDP rule is applied, provided in an embodiment of the present invention.

[0060] Figure 13 This is a schematic diagram illustrating the propagation of global information via neural modulation, provided in an embodiment of the present invention.

[0061] Figure 14 A schematic diagram of SNN network transformation and route search provided in an embodiment of the present invention;

[0062] Figure 15 This is a schematic diagram of path planning in a car parking scenario provided by an embodiment of the present invention;

[0063] Figure 16 This is a schematic diagram illustrating the application of an embodiment of the present invention in an Internet or cloud computing center environment;

[0064] Figure 17 A structural block diagram of a neuromorphic-based optimal network path search system provided in an embodiment of the present invention;

[0065] Figure 18 This is a structural block diagram of a neuromorphic-based optimal network path search device provided in an embodiment of the present invention. Detailed Implementation

[0066] The embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. The step numbers in the following embodiments are set only for ease of explanation, and there is no limitation on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0067] In the description of this invention, "multiple" means two or more. The use of "first" and "second" is for distinguishing technical features only and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or the order of the indicated technical features. Furthermore, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art.

[0068] This invention belongs to the field of Neuromorphic Computing and can be used in areas such as web search, autonomous vehicles, intelligent assisted driving, robotics, and surgical navigation. This invention establishes a unified model by processing spatial cognitive information, network information, and hierarchical relationships of external scene features in the hippocampus of the brain. This unified model possesses a network and the feature representation supporting the network attributes in the objective world, thereby enabling optimal path search.

[0069] Inspired by existing search technologies, this invention fully explores the capabilities of neural computing units in neuromorphic chips in computation, knowledge representation, and memory, especially episodic and semantic memory. It proposes a "method for network traversal and optimal path search using neuromorphic networks," which makes this method not only limited to searches that can purely express network relationship models, but also applicable to applications designed using ANN and / or SNN network systems, ranging from finding optimal paths in autonomous driving to finding parking spaces for autonomous vehicles, or from obstacle avoidance in robots to navigation of handheld devices based on landmark images, and even searches on internet social platforms, ensuring its functionality and performance.

[0070] Furthermore, experiments have revealed that for many types of stimuli, especially continuous variables, the neural system's encoding strategy involves a large group of neurons collaboratively encoding stimulus values. Each neuron's response covers a certain range of stimulus values, with a maximum response to a specific value (represented by a Gaussian-shaped tuning function, the Bump Attractor). The tuning function of the neuron group covers the entire parameter space. When an external stimulus is presented, the group of neurons responds, averaging out ubiquitous noise and thus enhancing the reliability of this method when applied to these products or systems.

[0071] Reference Figure 1 This invention provides a neuromorphic-based optimal network path search method, which specifically includes the following steps:

[0072] S101. Construct a spiking neural network and a neuromorphic computing kernel. The spiking neural network is a neuronal network layer system with multiple levels of feature expression, and there are multiple neurons in the same level. The neurons have multiple synapse types to form bidirectional connections between neurons.

[0073] S102. Determine the starting neuron and target neuron of the spiking neural network based on the problem to be solved;

[0074] S103. The neuromorphic computing kernel sends instructions to run an excitation signal that conforms to the SFA rule, so that the starting neuron emits several bell-shaped pulses to another neuron connected to it, thereby forming a transmission waveform characterized by attractor peaks that is transmitted in the spiking neural network, and the neuronal synapses on the propagation path are strengthened / weakened according to the STDP rule.

[0075] S104. When the transmitted waveform reaches the target neuron, determine the first network path through which the transmitted waveform passes, and then select the optimal network path from multiple first network paths.

[0076] Specifically, embodiments of the present invention at least include a unit or device capable of storing and executing computer instructions. These instructions can define and implement a Spiking Neural Network (SNN) and a feature extraction network on the unit or device. The defined SNN can be a neural network layer system capable of expressing multiple levels of feature representations required by a specific network model for a problem to be solved. Each level has multiple neurons, and each neuron has multiple synapse types to form bidirectional connections between neurons. Following the Spike Frequency Adaptation (SFA) rule, instructions can be sent and initiated to execute signals conforming to the SFA rule, activating the first neuron in the network. Subsequently, the first neuron will emit one or more pulses (bell-shaped pulses) to propagate to neurons connected to it in the network, forming a transmission waveform characterized by attractor peaks propagating in the network. Neuronal synapses along the propagation path will be strengthened / weakened according to the STDP rule. Neuromodulation signals with global parameter requirements can be emitted according to model requirements to change the system parameters of neurons. This device can determine the path from the target neuron to the fastest-reaching starting neuron, thereby determining the optimal network path.

[0077] Embodiments of the present invention include a computing system, such as Figure 2 The diagram shows a computing system provided in an embodiment of the present invention. The instructions of this computing system can run and define a neuromorphic computing network based on SNN, and / or can define and run an ANN (Artificial Neural Network). Figure 3 The diagram shows a neuromorphic computing device provided in an embodiment of the present invention. The neuromorphic computing device is composed of a neural computing core that connects the internal units of the computing core through an internal network bus.

[0078] In this embodiment of the invention, the neurons expressed by the computing device can run the STDP learning rules that satisfy the definition of Hebbian Learning. The simplest STDP rule is: if the requirement that the time of the pulse arriving at the pre-neuron precedes the time of the post-neuron can be met, then the weight of the synapse of the adjacent neurons will be enhanced; otherwise, it will be weakened.

[0079] As an optional implementation, the step of constructing a spiking neural network specifically includes:

[0080] Initialize the spiking neural network and determine several first nodes in a single-layer interconnected network whose fan-out number is greater than 1;

[0081] For the first node, a pre-defined second node is inserted between each fan-out branch and the next node to form a new spiking neural network. The second node conforms to the definition of SNN and does not affect the overall network characteristics of the spiking neural network.

[0082] The newly formed spiking neural network is assigned network parameters according to the preset modeling requirements, and then mapped to obtain the spiking neural network.

[0083] Specifically, the application types of optimal path search are divided into computer network problems where "nodes are clearly defined" and problems involving hierarchical networks. The former is a single-layer network, while the latter is a hierarchical network. The attribute features of the supporting network nodes of the latter network need to be extracted through neural networks.

[0084] For a single-layer interconnected network, the following step-by-step method is used to generate a network that conforms to the characteristics of a biological SNN network:

[0085] 1.1 Distinguish between nodes with a fan-out number greater than 1 and nodes with a fan-out number equal to 1 in a single-layer interconnected network;

[0086] 1.2 For nodes with a fan-out number greater than 1, insert a new node that conforms to the SNN definition (delay, weight) but does not affect the overall network characteristics of the computing node between each fan-out path and the next node;

[0087] 1.3 Integrate steps 1.1 and 1.2 into a new network; assign parameters to the connection strength between network nodes according to the parameters (delay, weight) required by the SNN model.

[0088] 1.4. Map the network from step 1.3 to a network that conforms to the characteristics of an SNN.

[0089] For problems with hierarchical features, a hierarchical network is constructed. The construction of the top-level network conforms to steps 1.1 to 1.4 above. It supports the attributes of the nodes of the top-level support network. A neural network needs to be selected, and deep learning methods can generally be used to extract features of external perception information. These features can be feature vectors or knowledge graph networks. Considering security and real-time requirements, sometimes the network processes the on-site conditions of external objects online. After the knowledge graph network integrates historical spatiotemporal characteristics, its features or the reasoning relationship of the knowledge network are used to generate conditions that meet the input requirements of the SNN network.

[0090] As an optional implementation, the neuromorphic computing kernel includes at least one of the following: feature extraction network, short-term memory network, episodic memory network, semantic memory network, knowledge graph network, deep learning network, and recurrent network.

[0091] Specifically, in this embodiment of the invention, the network that calculates external objects and performs feature extraction can be a feature extraction network of neuromorphic computing, a short-term memory network, an episodic memory network, or a semantic memory network. Alternatively, it can be a knowledge graph network, a deep learning network, or a recursive network that extracts features or knowledge based on ANN processing technology and performs processing. These networks ultimately generate the activation signal of the SNN.

[0092] As a further optional implementation, the synaptic inputs of the dendrites of the spiking neural network include input currents to excitatory neurons and input currents to inhibitory neurons.

[0093] Specifically, the synaptic inputs of the dendrites of an SNN network include currents that excite neurons and currents that inhibit neurons. These constitute the characteristics of a hierarchical network, including networks that support spatial location awareness, networks that extract features from image markers, and knowledge graph networks that extract entity knowledge driven by a combination of short-term memory, episodic memory, and semantic memory. The generation characteristics of these three supporting networks include the following steps:

[0094] 2.1 The neurons and their parameters in the supporting network of location cell neurons, and the corresponding SNN network neuron activation signal generation method;

[0095] 2.1.1 Divide the space into clusters of stripes. If the stripes advance forward, they are represented by three directions: 60 degrees, 120 degrees, and 180 degrees. The wave vectors in these three directions are represented by k. i (i = 1, 2, 3) represents the space, and the wavelength of space is λ, then the wave number is: k = 2π / λ. The space vector is represented by r, and the initial position is r0;

[0096] 2.1.2 The superposition of three stripe waves with a 60-degree phase difference corresponds to the firing pulse activity of the grid neurons, generating precise grid positioning in space.

[0097] 2.1.3 If the spatial wavelength of the periodic wave generated by the grid neuron is λ, the orientation is Θ, and the initial spatial phase is r0 = (x0, y0), it is uniformly expressed by the parameter w = (λ, Θ, r0). A function representing the nth grid cell. C represents its corresponding weight. inh Let represent the balance constant designed to balance the excitation inputs of the grid neurons. Then, the input to the position neurons is:

[0098]

[0099] [z] + =z for z≥0 and 0 otherwise.

[0100] The constructor for each grid neuron is as follows:

[0101]

[0102] Wherein, the wave vector is k, with a spatial azimuth distribution ranging from 60 degrees to 120 degrees, and g w max Its peak value.

[0103] 2.2. Networks based on feature extraction from biomarker images and the method by which the network outputs activation signals to the corresponding neurons of the SNN;

[0104] 2.2.1, such as Figure 4 As shown, the feature extraction signal of the marker image can be generated by using an ANN network through convolution, pooling and stacking to produce an output vector. Then, based on the conversion algorithm between ANN and SNN, the feature vector and activation function are extracted.

[0105] 2.2.2. Feature extraction signals from marker images can also be performed using an SNN approach, such as... Figure 5 As shown, it directly generates the correspondence between images and neurons, completing the learning task of the SNN; in Figure 5 In this model, the input neurons correspond to images with geographic identifiers, while the activation neurons are those that assign different geographic location attributes.

[0106] As an optional implementation, the optimal path search method further includes the following steps:

[0107] According to the preset model requirements, neuromodulation signals with global parameters are emitted, causing changes in the system parameters of neurons.

[0108] Specifically, the membrane potential and current model of neurons in the SNN network are established according to the following formula:

[0109]

[0110]

[0111] Where, τ m =C m R m It is the membrane potential time constant, i sens (t) is the sum of the sensor input currents, i syn (t) is the sum of the input currents of the activated neurons, i ns (t) is the background current, typically represented as Gaussian distributed noise with a mean of 0 and a variance of 5nA, i inh(t) represents the total current of the inhibitory neurons, which serves to generate synapse frequency adaptation (SFA) operations, thereby producing bell-shaped pulses. Driven by the attractors of the neural network, the bell-shaped pulses generate wavefronts with time-phase characteristics.

[0112] The neurons mentioned above are interconnected to form an SNN network. The formation of the network is based on the topological relationships between spaces, the topological relationships between problem features, or the topological relationships between networks. Once a network is established, the neurons need to be assigned initial configuration parameters and state parameters that satisfy the pre-inference conditions. The state parameters of the network can be obtained through network transformation or through learning.

[0113] As a further optional implementation, the excitation signal is generated through the following steps:

[0114] Constructing short-term memory networks to achieve short-term memory function;

[0115] Constructing an episodic memory network to realize episodic memory function;

[0116] The semantic memory network has been constructed to realize the semantic memory function;

[0117] A knowledge graph network is constructed based on the entity relationships in the environment in which the agent is located, and the knowledge graph network is driven by short-term memory output, contextual memory output, or semantic memory output to obtain the corresponding memory embedding vector;

[0118] The memory embedding vector is input into a pre-built recurrent network to identify the hidden states, and a feature sequence is formed based on the hidden states. Then, an excitation signal is generated based on the feature sequence.

[0119] Specifically, such as Figure 6 The diagram illustrates the generation process of the excitation signal. An excitation signal is generated by a knowledge graph network based on entity knowledge extraction driven by a combination of short-term memory, episodic memory, and semantic memory. This excitation signal is then input into the features of the SNN network. The generation of the excitation signal includes the following steps:

[0120] 3.1. Establish a short-term memory network to complete the short-term memory function;

[0121] 3.2 Establish an episodic memory network to realize the episodic memory function;

[0122] 3.3 Establish a semantic memory network to realize semantic memory function;

[0123] 3.4. Establish an entity relationship knowledge graph based on the environment in which the agent is located;

[0124] 3.5. Knowledge graph networks are driven by short-term memory, episodic memory, or semantic memory to obtain the corresponding memory embedding vectors;

[0125] 3.6. Input the vector from 3.5 into the LSTM network to identify the hidden states;

[0126] 3.7 The other state information from the input external sensors, together with the hidden state information output by the LSTM, form a new feature queue;

[0127] 3.8. Transform the queue output from 3.7 through the network to output the excitation signal required by the SNN.

[0128] As an optional implementation, the optimal network path search method further includes the steps of training a short-term memory network, an episodic memory network, and a semantic memory network, specifically including:

[0129] Establish geospatial location sequences and time-series image sequences and generate corresponding input data;

[0130] Determine the hierarchical transformation relationships and loss functions of short-term memory networks, episodic memory networks, and semantic memory networks;

[0131] The input data is sequentially fed into the short-term memory network, the episodic memory network, and the semantic memory network. The parameters of the short-term memory network, the episodic memory network, and the semantic memory network are updated through the backpropagation algorithm to obtain the trained short-term memory network, the episodic memory network, and the semantic memory network.

[0132] Specifically, for a single-layer network with a clear topological relationship and satisfying the search relationship, a compilation technique can be used to map it into an SNN network. Given that the starting node neuron and the target node neuron are clear, since the connection strength of the synapses between the neurons in this network has already been given the temporal characteristics of the network connection during compilation, the steps are an arrangement of inference and search. When the SFA first acts on the target neuron, the wavefront of the bell-shaped pulse first reaches the starting neuron in the path. Based on the state path of the neuron traversed by the bell-shaped pulse, the optimal path is obtained.

[0133] For an SNN network that reflects the spatial positional relationships of an agent, the learning steps for the spatial cognitive map of the network's spatial positional relationships and the learning steps for the spatial target vector relationship map are as follows.

[0134] 4.1 Learning steps for spatial cognitive maps: using r i (s) represents the firing activity of neuron i, while x represents the firing activity of neuron i. i If we represent its location as the cell center, and use vector p to represent the entire neural network activity, then p can be represented as:

[0135]

[0136] After repeated exploration, stable bidirectional balanced synaptic weights are formed between the synapses where neurons fire, creating a stable weight matrix that serves as long-term memory.

[0137] 4.2 Learning steps for spatial target vector relationship maps:

[0138] 4.2.1 Let the function H(t') represent the inductive strength of LTP enhancement when the presynaptic neuron precedes the postsynaptic neuron by t'. When LTP occurs along path X(t), the formula for the enhancement of synaptic strength between location cells j and i is obtained as follows:

[0139]

[0140] In the formula: To be in position The firing rate of time j, and the firing rate of cell i at position X:

[0141]

[0142] Thus, the encoding position is:

[0143]

[0144] S l The position center of cell i is given by the displacement encoding vector:

[0145]

[0146] 4.2.2 Formation of synaptic vector field:

[0147] As the wavefront propagates forward, neurons, according to the STDP learning rule, form asymmetric synaptic connections between bidirectionally connected synapses. These asymmetric connections constitute the neuron's vector field, and the relationship of this vector field is as follows:

[0148]

[0149] In the formula, r i For n i The neuron is located at position x i The location of the target neuron is the vector (x) j r i In a two-dimensional network, a vector field is formed from the starting point to the target point based on the asymmetry of the connection strength between neurons.

[0150] When an agent traverses a path, it generates a corresponding network path pointing to the target. If there are multiple experiences of exploring and learning the environment, these traces will reflect the agent's history.

[0151] For a series of images A geographically identified sequence, if the corresponding location is Given the image features of these images, the steps to complete the SNN image map are as follows:

[0152] 5.1 Extract image features according to the aforementioned image feature extraction steps;

[0153] 5.2 Complete the task using vertices as... The edge is ε={c ij} i≠j The network, where the vertices are... Series, with edges as The expression of the interconnected relationships;

[0154] 5.3 Form an SNN network according to the steps described above for generating an SNN network.

[0155] For a scene memory network (SNN) with specific spatiotemporal requirements, the learning steps are as follows:

[0156] 6.1 Establish the spatiotemporal geographic location and temporal image sequence in the scene to create the corresponding input;

[0157] 6.2. A recurrent neural network is selected, and reinforcement learning is used to learn the hierarchical transformation relationship between short-term memory, episodic memory, and semantic memory. The loss function is:

[0158] L(θ) = HuberLoss(Q) target Q(s, a; θ))

[0159]

[0160] Where s, a, s0, a0, r, and γ represent the state, action, next state, next action, rate of return, and discount rate, respectively; θ target To obtain the network parameters for reinforcement learning after performing the K-step;

[0161] 6.3 Complete the learning and training of short-term memory, episodic memory and semantic memory.

[0162] In the formed SNN network, the synapses of neurons, after bell-shaped pulses pass through adjacent neurons, exhibit synaptic changes that satisfy the STDP learning rule. The learning method performed by the STDP learning rule is as follows: the change in connection strength between neurons is:

[0163]

[0164] Among them, w ji (t) represents the connection strength from neuron i to neuron j, a < 0, Si(t) and Sj(t) represent the impulse activity of the preneuron and postneuron, respectively, S(t) = Σ f δ(t f -t) indicates the Spike Train, t f It is the moment of the f-th pulse. ij (s) and a ji (s) is an integral and a function, where This is the discharge time difference between the preceding and following synapses. Based on simulation results, the desired bell-shaped pulse wave propagation can be generated using the following kernel function. The STDP learning rule is:

[0165]

[0166] According to the Spike Frequency Adaptation (SPA) principle, a signal conforming to the SFA principle is sent to activate the first neuron in the network. Subsequently, the first neuron emits one or more pulses (bell-shaped pulses) to propagate to the neurons connected to it in the network, forming a transmission waveform characterized by attractor peaks that propagates through the network. Synapses along the propagation path are strengthened / weakened according to the STDP rule. Neuromodulation signals with global parameter requirements can be emitted according to model requirements to change the system parameters of the neurons. This device can determine the shortest network path by tracing the path from the target neuron to the fastest-reaching originating neuron.

[0167] The method steps of the embodiments of the present invention have been described above. The present invention will be further explained below with reference to specific embodiments.

[0168] After defining the neuron and network models of a neural network, the connection strength of synapses between neurons is learned and adaptively modified using the STDP rule. On a computing system with a sufficient number of neurons capable of neuromorphic computation, Figures 7(a) and 7(b) illustrate the learning and propagation process of a portion of the neurons in one dimension. There are two STDP learning rules between neurons, which are described in detail below. Figure 1 and Figure 2 ,exist Figure 1In the diagram, neurons 10001, 10002, 10003, 10004, 10005, 10006, and 10007 have bidirectional synapses connected as (10011, 10021), (10012, 10022), (10013, 10023), (10014, 10024), (10015, 10025), and (10016, 10026). Each neuron can be configured with the required neural membrane potential threshold parameters, initial synaptic parameters, and charging / discharging delay parameters. Its excitation direction can be from 1004 to 1003 to 1002 and then to 1001, or from 1004 to 1005 to 1006 and then to 1007. When this propagation occurs, it creates asymmetry in the synapses in the opposite direction. That is, due to the propagation of the discharge pulse wave, there is enhancement in a certain direction (thick arrow in the figure) and weakening in a certain direction (thin arrow in the figure).

[0169] like Figure 8 As shown, due to the asymmetry of the connections between neurons, the physical center corresponding to the center of the bell-shaped pulse wave will deviate from the center of the neuron's activity. This forms a directional vector between the peak value of the actual sensing signal and the peak value of the actual neuron's activity. In fact, it is the vector direction that occurs when the agent is at a certain point in space, but the activity of the nerve is pointing to a bias. This vector direction is formed by the agent's learning process in history and is reflected in the nonlinear attractor of the network.

[0170] The spatial relationships within each environment constitute a two-dimensional network of interconnected locations. In the two-dimensional context, this forms a two-dimensional wave packet, a deflection vector pointing towards the target along the two-dimensional space. When neuromorphic computing devices express such spatial relationships, they can represent both spatial cognitive relationships and the asymmetric synaptic weight relationships formed during initial movement and exploration. This results in rapid propagation of target orientation in two-dimensional space, leading to the discovery of the optimal path. Once the environment corresponds to the network that responds to it, if SFA conditions are applied to the neurons to be activated, the final generated trajectory is the optimal path guided by the vector field.

[0171] like Figure 9 As shown, in the network diagram to be represented by this invention, the connection between the two points shown in Figure 2011 is a bidirectional connection, as shown in Figure 2012.

[0172] like Figure 10 The diagram shows a partial representation of a network in the real world. Each node represents a neuron, and the connections between nodes represent bidirectional connections between neurons. These connections can be modeled using synaptic parameters. To find the optimal connection between two nodes, the relationship between the node and the connection... Figure 10 The expressions are the same.

[0173] For such a localized network, in a learning process, the initialization of the SNN network parameters can have several options. Initially, the parameters of each neuron can be set to be the same, or the parameters can be randomly initialized. However, if it is a network capable of inference, then it means that the parameters of its neurons and the network parameters between neurons meet the inference requirements. The parameters of its neurons can consider parameters required by the model, such as membrane potential and refractory period. This ensures that any neuron has a suitable time window for its output or input, during which no impulses are generated.

[0174] The process of finding the optimal path is the inference process after the network parameters have been configured. This inference process is accomplished by the propagation of bell-shaped pulse packets within the network. Figure 11 In this process, if neuron 2031 receives a activating signal SFA, the neuron's signal propagates to neighboring neurons in chronological order.

[0175] After the wavefront has propagated according to the propagation law, as Figure 12 As shown, the connection strength between neurons is enhanced under the STDP rule in the above process.

[0176] Figure 13 It is a model for the system to transmit global information through neural modulation. Depending on the needs of the model to reflect the objective problem, sometimes it is necessary to transmit the neuronal membrane potential, synaptic weights, and even the delay time after the neuron fires after a certain period of time. 2051 plays this role.

[0177] Figure 14 The left side represents a portion of a larger network, which is first transformed into a SNN-compliant network, i.e., the right side. In the right side, after the initial excitation starts from the target node, a bell-shaped pulse propagates along its adjacent neurons, causing them to transmit firing pulse signals to each other. When this signal reaches the source neuron (starting node), this path represents the optimal search path. At this point, the system's monitoring logic circuit and the path jump readout circuit read out the jump path, thus determining the optimal search route.

[0178] For a class of objective world events and scenes that are tightly coupled and require the memorization of spatiotemporal sequence features, the designed device and method can be trained and reasoned by using a neural network with short-term memory, episodic memory and semantic memory with time dimension. This enables the network to perceive the computational power with such tight constraints. Essentially, the network has the ability to construct local maps in real time and to search for the best strategy based on the understanding of the local maps in real time.

[0179] like Figure 15 In the scenario shown, what needs to be planned is a reasonable decision-making route from the starting position to the target position. At this time, the correspondence between the neurons in the device of this invention and the positions of key points on the scene can be constructed to form a local map of the scene. Moreover, since the local map of the scene is subject to time constraints, it is necessary to know the requirements of possible future spatiotemporal changes at every moment. This requires the use of short-term memory and contextual memory on the scene, as well as past experience and knowledge, to derive the spatial map of the next moment from the current moment, and then to perform optimization according to this spatial map.

[0180] like Figure 15 As shown, in this scenario, car 301 needs to be parked at the desired location 303. Based on the relationship between the destination and source locations, if there are moving cars 304 and pedestrians 305 present, it is necessary to consider different spatiotemporal environments and perform a search at specific time points. This requires injecting external information into the neurons at the corresponding spatial locations to bias the relevant neurons, such as setting a firing threshold. According to the method provided by this invention, the neurons represented by each location, initiated by the first bell pulse, propagate in parallel, ultimately forming the optimal path.

[0181] Figure 16 This invention describes an application of the method in an internet or cloud computing center environment. The neuromorphic computing system 401 can accept various source signals, including external sensor signals such as different types of sensors (408-409). These sensors can detect and measure features of the external environment; for example, they can measure the speed and direction of an object's movement, pixels in an image, video, audio, chemical odors, computer communication signals, noise, etc. The data source can also be the database (406) or signals from other datasets, enabling 401 to complete its intelligent learning, deep learning, feature extraction, and other computational tasks. The results of the neuromorphic computing can also be provided to other computing systems, such as 402, 403, 404, 405, and 407.

[0182] A neuromorphic computing system is actually a large-scale integrated circuit (VLSI). This circuit has a neural computing unit that simulates the brain, as well as a CPU for coordination and control, input / output logic, and other circuits. In some designs, the core unit of the 401 part can sometimes be embedded in the 402, 403, 404, 405, or 407.

[0183] The specific embodiments of the present invention have been described above. It can be understood that the embodiments of the present invention, based on the frequency adaptability of synapses in biological neural networks, employ bell-shaped pulse technology. Through the collaborative excitation response of a large number of neurons, a transmission waveform characterized by attractor peaks is formed for transmission, effectively filtering out noise factors generated during propagation and improving the accuracy of optimal network path search. The use of a multi-level feature-expressing neural network layer system can better represent the problem model, making the network path search more applicable and better meeting quantitative constraints such as space, time, and security, thus improving the efficiency and reliability of optimal network path search.

[0184] Inspired by the spatial cognition of the hippocampus in organisms, it was discovered that there are corresponding neurons in the hippocampus that correspond to geographical relationships in the external objective world. These neurons are interconnected to form a network that reflects geographical relationships in the objective world. However, the problem of analogy between neurons in the network is not solved. Neural activity in the biological brain is characterized by very fast speed. The transmission relationship of neurons in this characteristic has temporal phase characteristics. Hopfield used this to calculate algorithms that are difficult to implement with traditional CPUs or GPUs (NP-hard problems). Its computing mode is called the Computing with time paradigm.

[0185] By applying the inherent search capabilities of neural networks to real-world networks, employing a computing-with-time algorithm on a neural computing platform, and combining this with the advantages of neural networks in extracting problem features, we can more effectively solve the technical problems in our field of invention.

[0186] This invention proposes a hippocampus model within a neuromorphic computing system (chip and algorithm) that more closely resembles how humans or biological systems solve such problems. Based on the relationship between this model and the objective world network we hope to correspond to, a general method for finding the optimal search solution is found. It includes:

[0187] 1. Further promote the application of the laws of steady-state and metastable states in neural dynamics, break through the traditional pulse method of triggering neural networks, and adopt the method of transmitting through bell pulses (Bump Attractor) under the paradigm of neural network adaptation and computing with time, so that the network and its variables of the model can be more universal and have better reliability.

[0188] 2. Make full use of the feature extraction advantages of neural networks, extract the supporting features of the above networks, adopt hierarchical networks, better express the problem model, make the applicable scope wider, and better meet the quantitative constraints such as space, time and security.

[0189] Neuromorphic technology may be the next stage of development in high-performance computing. It can significantly improve data processing and machine learning capabilities, and bring about a significant increase in performance and power efficiency while greatly accelerating response speed. Neuromorphic computing technology is currently in the research and development reserve stage. This application proposes targeted solutions to the problems that previous invention patents have in the process of implementation, such as insufficient anti-interference ability, limited applicable scenarios and lack of specificity, insufficient quantitative description of time, space and security constraints, and inability to use spatial cognitive knowledge for effective decision-making. It can provide good application prospects for Internet search, social network search, autonomous driving, intelligent assisted driving, robotics or medical surgical navigation, etc., and has great economic value.

[0190] Compared to existing technologies, this scheme, based on the synaptic frequency adaptation of biological neural networks, employs a bell-shaped pulse technique. A large group of neurons collaboratively encodes stimulus values, with each neuron's response covering a certain range of stimulus values ​​and producing a maximum response to a specific value (represented by a Gaussian-shaped tuning function). The tuning function (bump) of the neuron group covers the entire parameter space. When an external stimulus is presented, the group of neurons responds, thus averaging out ubiquitous noise. The method described in this invention provides a very natural network implementation mechanism for this group encoding strategy: the group of neurons involved in encoding are connected to form a Bump Attractor; when there is an external input, the network enters an attractor, generating a Gaussian-shaped wave packet activity; because it is an attractor, noise is removed, and the vertex position of the wave packet gives the result of neural network decoding, thus the algorithm is more reliable.

[0191] Meanwhile, in the technical background introduction, the biological network adopts a single network, while this solution adopts a hierarchical network. The sub-networks supporting its hierarchical network can be location coding networks that reflect location fields, geographic location networks that reflect image features, or short-term memory, episodic memory, or semantic memory networks based on memory, thereby better modeling specific networks and making them more widely applicable.

[0192] Furthermore, by adopting a memory network (short-term memory, episodic memory, and semantic memory) that is more adapted to biological characteristics, the device's network uniformity is improved. This allows for faster application of features, similar to popular Google Street View Time Machine examples, and the use of neural synapses to represent information reduces memory usage.

[0193] Reference Figure 17 This invention provides a neuromorphic-based optimal network path search system, comprising:

[0194] The spiking neural network construction module is used to build spiking neural networks and neuromorphic computing kernels. The spiking neural network is a neuronal network layer system with multiple levels of feature expression, and there are multiple neurons in the same level. The neurons have multiple synapse types to form bidirectional connections between neurons.

[0195] The neuron identification module is used to determine the starting and target neurons of the spiking neural network based on the problem to be solved.

[0196] The waveform transmission module is used to send instructions through the neuromorphic computing kernel to run an excitation signal that conforms to the SFA rule, so that the starting neuron emits several bell-shaped pulses to another neuron connected to it, thereby forming a transmission waveform characterized by attractor peaks that is transmitted in the spiking neural network, and the neuronal synapses on the propagation path are strengthened / weakened according to the STDP rule.

[0197] The path selection module is used to determine the first network path through which the transmitted waveform passes when it arrives at the target neuron, and then select the optimal network path from multiple first network paths.

[0198] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0199] Reference Figure 18 This invention provides a neuromorphic-based optimal network path search device, comprising:

[0200] At least one processor;

[0201] At least one memory for storing at least one program;

[0202] When the above-mentioned at least one program is executed by the above-mentioned at least one processor, the above-mentioned at least one processor implements the above-mentioned neuromorphic optimal network path search method.

[0203] The content of the above method embodiments is applicable to the device embodiments. The specific functions implemented by the device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0204] This invention also provides a computer-readable storage medium storing a processor-executable program that, when executed by a processor, performs the aforementioned neuromorphic-based optimal network path search method.

[0205] This invention provides a computer-readable storage medium that can execute a neuromorphic-based optimal network path search method provided in the method embodiments of this invention. It can execute any combination of the implementation steps of the method embodiments and has the corresponding functions and beneficial effects of the method.

[0206] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to perform... Figure 1 The method shown.

[0207] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the aforementioned blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0208] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the aforementioned functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0209] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0210] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0211] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the aforementioned program can be printed, because the aforementioned program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or, if necessary, processing in other suitable ways, and then stored in computer memory.

[0212] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0213] In the foregoing description of this specification, references to terms such as "one embodiment," "another embodiment," or "some embodiments" indicate that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in at least one embodiment or example of the present invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0214] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0215] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A neuromorphic-based optimal network path search method for determining the optimal movement path of a vehicle / robot, characterized in that, Includes the following steps: A spiking neural network and a neuromorphic computing kernel are constructed. The spiking neural network is a neuronal network layer system with multiple levels of feature expression, and there are multiple neurons at the same level. The neurons have multiple synapse types to form bidirectional connections between each neuron. The correspondence between the neurons and the locations of key points on site is constructed to form a local map of the site; The starting and target neurons of the spiking neural network are determined based on the problem to be solved. The neuromorphic computing kernel sends instructions to run excitation signals that conform to the adaptive laws of neuronal synapses, causing the starting neuron to emit several bell-shaped pulses to another neuron connected to it, thereby forming a transmission waveform characterized by attractor peaks that is transmitted in the spiking neural network, and causing neuronal synapses on the propagation path to be enhanced / weakened according to the STDP rule. When the transmitted waveform reaches the target neuron, the first network path through which the transmitted waveform passes is determined, and then the optimal network path is selected from multiple first network paths. Based on the optimal network path, the optimal movement path of the vehicle / robot is determined.

2. The optimal network path search method based on neuromorphology according to claim 1, characterized in that, The step of constructing the spiking neural network specifically includes: Initialize the spiking neural network and determine several first nodes in a single-layer interconnected network whose fan-out number is greater than 1; For the first node, a preset second node is inserted between each fan-out branch and the next node to form a new spiking neural network. The second node conforms to the definition of SNN and does not affect the overall network characteristics of the spiking neural network. The newly formed spiking neural network is assigned network parameters according to the preset modeling requirements, and then the spiking neural network is mapped to obtain the spiking neural network.

3. The optimal network path search method based on neuromorphology according to claim 1, characterized in that: The neuromorphic computing kernel includes at least one of the following: feature extraction network, short-term memory network, episodic memory network, semantic memory network, knowledge graph network, deep learning network, and recursive network.

4. The optimal network path search method based on neuromorphology according to claim 1, characterized in that, The synaptic inputs of the dendrites of the spiking neural network include the input currents of excitatory neurons and the input currents of inhibitory neurons.

5. The optimal network path search method based on neuromorphology according to claim 1, characterized in that, The optimal network path search method further includes the following steps: According to the preset model requirements, a neuromodulation signal with global parameters is emitted, causing the system parameters of the neuron to change.

6. A neuromorphic-based optimal network path search method according to any one of claims 1 to 5, characterized in that, The excitation signal is generated through the following steps: Constructing short-term memory networks to achieve short-term memory functions; Constructing an episodic memory network to realize episodic memory function; Constructing a semantic memory network to achieve semantic memory function; A knowledge graph network is constructed based on the entity relationships in the environment in which the agent is located, and the knowledge graph network is driven by short-term memory output, contextual memory output, or semantic memory output to obtain the corresponding memory embedding vector. The memory embedding vector is input into a pre-constructed recurrent network to identify the hidden states, and a feature sequence is formed based on the hidden states. Then, the excitation signal is generated based on the feature sequence.

7. The optimal network path search method based on neuromorphology according to claim 6, characterized in that, The optimal network path search method further includes the step of training the short-term memory network, the episodic memory network, and the semantic memory network, which specifically includes: Establish geospatial location sequences and time-series image sequences and generate corresponding input data; Determine the hierarchical transformation relationship and loss function of the short-term memory network, the episodic memory network, and the semantic memory network; The input data is sequentially input into the short-term memory network, the contextual memory network, and the semantic memory network. The parameters of the short-term memory network, the contextual memory network, and the semantic memory network are updated through the backpropagation algorithm to obtain the trained short-term memory network, the contextual memory network, and the semantic memory network.

8. A neuromorphic-based optimal network path search system for determining the optimal movement path of a vehicle / robot, characterized in that, include: The spiking neural network construction module is used to construct a spiking neural network and a neuromorphic computing kernel. The spiking neural network is a neuronal network layer system with multiple levels of feature expression, and there are multiple neurons in the same level. The neurons have multiple synapse types to form bidirectional connections between each neuron. It is also used to construct the correspondence between the neurons and the location of key points on site to form a local map of the site. A neuron determination module is used to determine the starting neuron and target neuron of the spiking neural network based on the problem to be solved. The waveform transmission module is used to send instructions through the neuromorphic computing kernel to run an excitation signal that conforms to the adaptive law of neuronal synapses, so that the starting neuron emits several bell-shaped pulses to another neuron connected to it, thereby forming a transmission waveform characterized by attractor peaks that is transmitted in the spiking neural network, and the neuronal synapses on the propagation path are enhanced / weakened according to the STDP rule. The path selection module is used to determine the first network path through which the transmitted waveform passes when it reaches the target neuron, and then select the optimal network path from multiple first network paths, and determine the optimal movement path of the vehicle / robot based on the optimal network path.

9. A neuromorphic-based optimal network path search device, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements a neuromorphic optimal network path search method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a processor-executable program, characterized in that, The processor-executable program, when executed by the processor, is used to perform a neuromorphic-based optimal network path search method as described in any one of claims 1 to 7.

Images