192 results about "Spiking neural network" patented technology

The invention relates to the technical field of artificial intelligence, and discloses an artificial limb driving method based on a spiking neural network, which comprises the following steps: acquiring a multi-modal signal, and preprocessing the multi-modal signal; converting the preprocessed multi-mode signal into a pulse signal; inputting the pulse signal into a pulse neural network for sensing fusion processing, and outputting a control strategy; dynamically adjusting a control strategy based on pulse reinforcement learning according to the behavior feedback information through a reward function, and generating a learning result; and generating a bionic motion instruction according to the adjusted control strategy and the learning result, and driving an artificial limb execution mechanism to act according to the bionic motion instruction. According to the method, the electromyographic signals, the tactile pressure signals and the inertial measurement data are uniformly coded into the pulse sequence, the pulse neural network is utilized to realize multi-modal sensing fusion, the pulse reinforcement learning algorithm is introduced, the artificial limb control strategy is optimized in real time according to user behavior feedback, and the accurate control of the artificial limb joint is improved.

The invention relates to a method for deploying a spiking neural network to a hardware device. The method includes providing the spiking neural network, training the spiking neural network to obtain a trained spiking neural network, mapping neurons and synapses in the trained spiking neural network to corresponding components of the hardware device, simulating deployment of the trained spiking neural network on a hardware device using the obtained mapping, and deploying the trained spiking neural network to the hardware device using the mapping and the simulation. The training, mapping, simulation and deployment steps are executed by using hardware information of the hardware device; wherein the hardware information comprises at least one of hardware resource constraints, hardware connection constraints, dynamic range of hardware design parameters, characterization or statistics of neurons and synapses, reconfigurability, programmability, yield, computing resources, temporal characteristics, constraints on pre-processing, interfaces and peripheral devices, and available encoders and / or decoders.

The invention discloses a lightweight target trajectory prediction method based on a twin pulse neural network, and belongs to the field of target trajectory prediction. The method comprises the following steps: constructing a sample pair comprising a template frame and a search frame; constructing a twin spiking neural network model comprising a template branch, a search branch, an adaptive mask module, a spiking nerve Transform module and an output prediction layer, wherein the template branch and the search branch share weights; the adaptive mask module is used for aligning output features of two branches and generating a mask and feature enhancement, and the pulse nerve Transform module is used for acquiring a global dependency relationship of the enhanced features and outputting a current position and a tracking path end point of a target through an output prediction layer so as to generate a tracking plan path of the target; and training the constructed model based on the training data constructed by the sample pair to obtain a model used for obtaining the tracking plan path of the target. According to the method, key pulses are dynamically screened, the network structure is optimized, and double improvement of precision and efficiency is achieved.

The invention provides a cable testing method and a small handheld cable tester. The method comprises the following steps: carrying out wavelet transform noise reduction, Z-score standardization and time sequence dynamic regularization on original cable data; then constructing a geometric deep learning network by using a dispersed self-organizing structure of the rotating cube set, and extracting multi-modal depth features; fusing different modal features through a spiking neural network gating mechanism and a cross-modal attention mechanism; and finally, fault classification is carried out based on a differential evolution optimized neural network, and fault location is realized by combining a time sequence generative adversarial network and a dynamic probability neighborhood growth clustering algorithm. The system also evaluates the data contribution degree of each modal through information entropy and mutual information analysis, and optimizes the model performance by using adaptive weight distribution and lightweight neural network technologies. According to the invention, various fault types such as cable breakage, short circuit, insulation aging, poor contact and the like can be identified and positioned with high precision, and the efficiency and the accuracy of cable maintenance are remarkably improved.

Koopman operator theory is a widely used method to analyze, control, and predict the behavior of the states of a non-linear dynamical system using measurement functions in Hilbert space. Real time approximation of the Koopman operator is crucial in order to adapt and understand behavior of underlying non-linear dynamical system. Traditional approaches leverage matrix-based methods or artificial neural networks to approximate Koopman operator. However, such methods necessitate significant power and computational resources, therefore may not be suitable for applications that require real-time on-board processing. The problems of the conventional approaches are resolved based on a recent development of brain inspired spiking neural networks and neuromorphic computing platforms, as these offer extremely low-energy computation and real-time responses. Embodiments of the present disclosure provide implementation of a Spiking Neural Network (SNN) based architecture that efficiently approximate Koopman operator with minimal length of data and demonstrates significant computational savings.

The invention relates to a high-efficiency three-dimensional point cloud classification method and device based on a pulse neural network, and the method obtains an event-driven point cloud expression form through sparse pulse voxel coding, and achieves the efficient extraction of local features through the combination of a pulse-driven sparse convolution structure. And a pulse self-attention module driven by logical operation is utilized to construct global correlation characteristics, and finally, category judgment is completed through a pulse time integration mechanism, so that the calculation amount and energy consumption are remarkably reduced while the classification precision is kept. According to the method, sparse activation and event driving characteristics of the pulse neural network are fully utilized, efficient processing of the three-dimensional point cloud is achieved, and the method is suitable for automatic driving, robots and embedded scenes sensitive to power consumption.

The invention discloses a satellite Internet of Things uplink signal detection method and a satellite Internet of Things uplink signal detection system, which integrate traditional signal processing and advanced spiking neural network technologies, realize high-precision frame synchronization by performing two-dimensional cross-correlation matching in a time-frequency domain, and perform low-power-consumption and high-robustness symbol judgment by using an event-driven spiking neural network. Compared with a traditional method based on FFT or matched filtering, the method has higher adaptability and detection performance in a satellite Internet of Things scene with large Doppler frequency shift, low signal-to-noise ratio and dense users.

The invention belongs to the technical field of thermal management and performance regulation and control of a computer hardware system, and particularly discloses a temperature and performance coordinated regulation and control method of an MODT mainboard central processing unit. The entropy production rate of the system is calculated in real time in combination with a thermodynamic entropy production evaluation mechanism, and pre-cooling or advanced frequency reduction pulse type regulation and control are triggered when entropy production exceeds the threshold; meanwhile, a biological rhythm simulation mechanism is introduced, high-frequency and low-frequency alternating periodic oscillation control is carried out on the frequency of the processor in a millisecond-level window, the two mechanisms cooperatively act on the power management unit, and microsecond-level voltage regulation and heat dissipation driving linkage is achieved. According to the technical scheme, on the premise that the hardware cost is not increased, the energy efficiency performance, the transient heat dissipation capacity and the long-term operation reliability of the system under the high dynamic load are improved.

The application discloses a reconfigurable processor and an execution method fusing neuromorphic and general computing, and comprises a total control module and an on-chip network module, wherein the on-chip network module comprises a control module, a pulse storage module, a weight storage module, a pulse time module, a LIF neuron module and a classification comparison module. The execution method comprises the following steps: acquiring an input data set; selecting a working mode of the on-chip network module according to the input data set; and performing data recognition processing on the input data set according to the working mode, and outputting a data recognition output result. The application realizes the working of a central processing unit (CPU) and a spiking neural network (SNN) in a single core, thereby improving the utilization rate of a heterogeneous core and reducing the delay of core-to-accelerator communication. The application can be widely applied to the technical field of neural network accelerator as a reconfigurable processor and an execution method fusing neuromorphic and general computing.

This invention discloses a method for constructing a spiking neural network based on recurrent firing neurons, comprising: S1: constructing a recurrent firing spiking neuron model; S2: calculating positive and negative pulses based on the gradient function corresponding to each time step in the recurrent firing spiking neuron model to obtain cross-time-step positive and negative pulses; S3: backpropagating the cross-time-step positive and negative pulses back through the recurrent firing spiking neuron model to obtain a spiking neural network; S4: training the spiking neural network using a positive and negative membrane potential balance loss function to obtain a recurrent firing spiking neural network. This invention significantly improves the learning ability and information representation ability of the spiking neural network, while enhancing its robustness.

The disclosure relates to an artificial spiking neural network building block, a spiking neural network, a method for training a spiking neural network, an apparatus and a non-transitory computer readable media. The computer implemented method comprises assigning random weights to all connections between spiking neurons of the spiking neural network. The method comprises iteratively optimizing the weights by running a simulated annealing algorithm, using training data. The method comprises refining the weights by running a genetic algorithm.

This invention provides a task processing method and apparatus based on an AI chip. The method includes: constructing a spiking neural network model; optimizing parameters in the spiking neural network model, including at least a pulse trigger threshold and membrane potential, using a preset threshold prediction model and a leaky integrated release model; receiving a task request and determining the task type of the request; determining a target spiking neural network model based on the task type, and calling the target spiking neural network model to process the task request to obtain a processing result. This invention enables task processing using an optimized spiking neural network model within an AI chip, and improves model processing efficiency by optimizing the pulse trigger threshold and membrane potential within the spiking neural network model.

The invention discloses a spiking neural network model compiling method based on SpikingJeller framework description, and the method comprises the steps: carrying out the symbolization tracking of a model described by a SpikingJeller framework through introducing PyTorch FX according to the characteristics of time dimension recursion, pulse event driving, and PyTorch-based dynamic graph expression of the spiking neural network model described by the SpikingJeller framework, and obtaining a convertible static calculation graph; relay custom operators special for the spiking neural network are expanded in the TVM, the Relay operators are connected and bound to a brain-like processor hardware operator library, and conversion from a model described by a SpikingJeller framework to Relay intermediate representation capable of being adapted to a brain-like processor is automatically achieved.

This application discloses a controller and control system for a flexible actuator, and a biomimetic robot, belonging to the field of flexible actuator control technology. The controller includes: a pulse encoding module for converting target motion commands into pulse sequences; a spiking neural network processing unit for receiving the pulse sequences and collaboratively generating multi-channel pulse control signals based on its internal state evolution; and a multi-channel drive circuit connected to the output of the spiking neural network processing unit, where each channel receives one pulse control signal and converts it into a drive signal suitable for driving the corresponding flexible drive unit. By integrating the spiking neural network processing unit and the multi-channel drive circuit, the controller achieves end-to-end pulse domain communication from command input to drive execution, and can adaptively adjust the multi-channel control signals based on real-time feedback, thereby providing precise and compliant collaborative control for the flexible actuator.

There is provided a neurosynaptic processing core with spike time dependent plasticity (STDP) learning for a spiking neural network, including: a spiking neuron block including a pre-synaptic block and a post-synaptic block; a synapse block communicatively coupled to the spiking neuron block; a STDP learning block communicatively coupled to the spiking neuron block and the synapse block, the STDP learning block including a pre-synaptic event accumulator including a pre-synaptic spike event memory block and a pre-synaptic spike parameter modifier; a post-synaptic event accumulator including a post-synaptic spike event memory block and a post-synaptic spike parameter modifier, a weight change accumulator, and a weight change parameter modifier; a learning error modulator; and a synaptic weight modifier configured to modify a synaptic weight parameter based on a weight change parameter and a learning error corresponding to the synaptic weight parameter. There is also provided a corresponding method of operating and a corresponding method of forming the neurosynaptic processing core.

The application provides a kind of method for enhancing robustness of pulse neural network based on neural activation and connection optimization, applied to artificial intelligence and neural network technical field, the method comprises: constructing pulse neural network model, the neuron of pulse neural network model adopts burst enhancement type pulse neuron, the input of pulse neural network model is image data, and the output of pulse neural network model is the image processing result corresponding to computer vision task;When the membrane potential of burst enhancement type pulse neuron exceeds the firing threshold, the part exceeding the membrane potential is converted into the pulse output within the burst window by quantization linear mapping function, and the number of pulse output is an integer between 0 and the preset maximum burst pulse number;When training pulse neural network model, add activation perception regularization term to total loss function;Through the application, the accuracy and robustness can be simultaneously improved while maintaining the high energy efficiency advantage of pulse neural network.

The invention provides a fire detection system and method based on a pulse neural network. An image acquisition module is used for reading image frames of a video; the pulse coding module forms image pulse sequence data; the fire feature extraction module extracts related features of the image pulse sequence data layer by layer to respectively obtain a shallow fire feature map and a deep fire feature map; the fire feature extraction and fusion module fuses the high-resolution features of the shallow fire feature map and the low-resolution features of the deep fire feature map; the pulse activation module triggers a fire characteristic time sequence pulse sequence; the fire target detection module optimizes the fire characteristic time sequence pulse sequence to obtain fire characteristic core data; the pulse decoding module obtains classification and bounding box parameters of a fire target; the fire result output module outputs a final fire detection result after screening; according to the invention, high-precision fire detection performance is realized, and the advantage of low power consumption of the pulse neural network is maintained.

A method is provided and includes operations as below: training a spiking neural network (SNN) in a first device to generate multiple first weight values of M bits; calculating multiple second weight values of N bits corresponding to the first weight values according to a threshold value, the number M, and the first weight values, wherein the number N is smaller than the number M; retraining the spiking neural network with the second weight values to update the second weight values; and performing a write operation to save the updated plurality of second weight values in a memory in a second device for performing a spiking neural network operation in the second device.

This application discloses a method, apparatus, electronic device, and storage medium for recognizing facial expression images, relating to the field of image recognition technology. It includes: determining the original vector of the expression image based on its pixel values; performing principal component analysis on the original vector to obtain the target vector of the expression image; inputting the target vector into a feature extraction network to obtain feature information of the expression image; encoding the feature information using pulse frequency based on a Poisson distribution to obtain a pulse sequence of the expression image; and inputting the pulse sequence into an expression recognition model to obtain the recognition result of the expression image. The expression recognition model is a multilayer spiking neural network, where the spiking neuron model is an improvement based on the leaky integral discharge neuron (LIF) model. This application's technical solution offers better interpretability and flexibility, significantly reducing network energy consumption while improving computational power, making it suitable for more intelligent small robots such as customer service robots.

This invention provides a cable testing method and a small handheld cable tester. The method includes: performing wavelet transform noise reduction, Z-score normalization, and time-series dynamic normalization on the raw cable data; then constructing a geometric deep learning network using the dispersed self-organizing structure of a rotating cube set to extract multimodal deep features; next, fusing different modal features through a spike neural network gating mechanism and a cross-modal attention mechanism; finally, performing fault classification based on a differential evolution optimized neural network, and achieving fault location by combining a temporal generative adversarial network and a dynamic probabilistic neighborhood growing clustering algorithm. The system also evaluates the contribution of each modality of data through information entropy and mutual information analysis, and optimizes model performance using adaptive weight allocation and lightweight neural network technology. This invention can accurately identify and locate various fault types such as cable breaks, short circuits, insulation aging, and poor contact, significantly improving the efficiency and accuracy of cable maintenance.

The invention discloses an equipment repair method and device, electronic equipment and a storage medium, and belongs to the field of automatic equipment repair. The method comprises the following steps: determining a multi-mode signal as a first input current of a neuron of a spiking neural network; determining a trigger threshold value through the operation environment parameter, and determining the trigger threshold value as a second input current of the neuron; according to the first input current and the second input current, determining an abnormal critical value for avoiding discharge of the neurons in real time, and determining the abnormal critical value determined in real time as a dynamic threshold value; and when the actually measured parameter of the multi-modal signal is greater than the dynamic threshold value, executing a preset repair strategy for the to-be-repaired equipment, thereby realizing real-time and self-adaptive adjustment of a fault early warning standard by fusing the multi-modal signal of the equipment and the operating environment parameter, thoroughly solving the problems of false alarm and missing alarm caused by a traditional fixed threshold value, and improving the fault early warning efficiency. And therefore, the self-repairing can be started at a more accurate opportunity.

Disclosed herein are an asynchronous spike counting interface apparatus and an asynchronous spike counting method in a spike neural network. The asynchronous spike counting interface apparatus may include multiple counters, each counting the number of times each of neurons included in the spike neural network fires, and an address decoder for outputting an enable signal to the counter that counts the number of times the neuron corresponding to the address output from the spike neural network fires.

Disclosed is a spike neural network circuit according to an embodiment of the present disclosure, which includes a self-correcting control circuit that generates an input signal and a first control code, a bias voltage generation circuit that generates a first bias voltage based on the first control code, a synaptic circuit including a first synaptic column that performs an operation of the input signal and a first weight signal and generates a first operation signal, a neuron circuit including a first neuron that generates a first output signal based on a comparison of the first operation signal and a threshold voltage, and a spike comparison circuit that generates a first comparison signal corresponding to a difference between the first output signal and a reference number, and the self-correcting control circuit further generates a second control code for correcting the first bias voltage.

An apparatus comprises a bias current generation circuit configured to generate a bias current, a synapse circuit including a weight register storing a weight value and configured to perform a charge operation based on the input spike signal, the bias current, and the weight value, a membrane capacitor having a potential determined based on the charge operation of the synapse circuit, and a neuron circuit configured to generate an output spike signal based on a comparison between the potential of the membrane capacitor and a threshold potential. The bias current generation circuit comprises a self-correction circuit, and the self-correction circuit comprises a target input spike register storing a value related to a target number of input spikes and is configured to correct the bias current generated by the bias current generation circuit based on the value related to the target number of input spikes.

A spiking neural network (SNN) system is disclosed having spiking neurons associated with clause values associated with a satisfiability (SAT) problem. Each spiking neuron generates a voltage spike when on input voltage applied to each spiking neuron is increased above a spiking voltage threshold. The input voltage that is increased above the spiking voltage threshold is indicative that the corresponding clause value is satisfied. A controller applies the clause grid that includes clause values to a literal grid that includes literal values. The controller generates the input voltage that is applied to each spiking neuron based on each corresponding clause value associated with each spiking neuron that is applied to the literal values. The controller determines that each clause value is satisfied when the corresponding spiking neuron generates the voltage spike and that each clause value is unsatisfied when the corresponding spiking neuron fails to generate the voltage spike.

Methods for encoding and decoding data from an array-based sensing system. Example embodiments describe a computer-implemented method for training a spiking neural network to encode and decode event information obtained from a set of sensors in an array-based sensing system. Further embodiments describe a neural network training system and sensor and edge gateway devices employing such an encoding and decoding spiking neural network.

This invention provides a deep spiking neural network method for ECG classification based on attention and integer training pulse inference, comprising the following steps: acquiring raw ECG signal data and preprocessing the raw ECG signal data; performing three rounds of convolution and corresponding max pooling on the ECG data features; performing block-based local self-attention processing on the ECG data features to obtain feature associations in local regions of the ECG data features; performing global self-attention processing on the ECG data features to obtain feature associations across the entire sequence of the ECG data features; performing two rounds of convolution and corresponding max pooling on the ECG data features; classifying the integrated ECG data features using a classification head, and outputting the ECG data classification result. This invention can effectively mine long-range dependency information of ECG data and retain shallow feature information through residual fusion, avoiding the gradient decay problem in deep networks, thereby enabling deep spiking neural network learning.

This invention belongs to the field of image recognition technology, specifically relating to a continuous learning target recognition method based on spatiotemporal information fusion using a spiking neural network. It comprises two learning modules: a fast learner and a slow learner. The slow learner acquires general features that are invariant to input changes but sensitive to semantic representation through self-supervised learning. In the fast learner, whenever a new task arrives, the processor freezes the trained feature extraction modules and adds new feature extraction modules to learn features of the new category. These feature extraction modules gradually aggregate to form a joint feature representation. By fusing general representations with task-specific features across spatial and temporal scales, this method effectively mitigates task confusion and catastrophic forgetting, improving target recognition accuracy.

The invention belongs to the technical field of computer vision, and discloses a robustness detection method for an aerial target of an unmanned aerial vehicle. A pulse neural network architecture is adopted, LIF neurons serve as basic calculation units, and invalid calculation is reduced through an event-driven sparse pulse distribution mechanism. In the feature extraction stage, a pulse adversarial interactive distillation module is introduced, L2 normalized energy constraint is applied to synaptic current, amplification and propagation of adversarial disturbance in the network are inhibited, and feature stability and detection robustness of the model under the conditions of noise interference and hostile attack are enhanced. The pulse channel characteristic refining module is based on a pulse perception-analog modulation mechanism, carries out adaptive modulation on synaptic current gain by using channel and space context information on the premise of maintaining pulse binary distribution characteristics, thereby highlighting target related characteristics and inhibiting complex background interference, and improving the target tracking precision. And the accuracy and reliability of unmanned aerial vehicle aerial image target detection are improved.

A coupling strength update device is configured to comprise: a coupling strength management unit (3) that manages the coupling strength between a first neuron, which is a certain neuron among a plurality of neurons included in a spiking neural network, and a second neuron, which is a neuron that acquires a spike signal outputted from the first neuron upon firing of the first neuron; an elapsed time management unit (1) that manages the elapsed time since the spike signal was outputted from the first neuron; and a firing rate management unit (2) that manages the firing rate, which is a rate at which the first neuron fires. In addition, when the second neuron fires, the coupling strength management unit (3) of the coupling strength update device updates the coupling strength between the first neuron and the second neuron on the basis of the elapsed time managed by the elapsed time management unit (1) and the firing rate managed by the firing rate management unit (2).