Edge device low-power-consumption reasoning method and system based on neuromorphic calculation and spiking neural network

Abstract

The invention provides an edge device low-power-consumption reasoning method and system based on neuromorphic calculation and a spiking neural network, and relates to the technical field of spiking neural networks, in particular to construction of a spiking neural network model, data input and preprocessing, result output and application and low-power-consumption management. Through event-driven calculation, the spiking neural network naturally has event-driven characteristics, the edge device can trigger calculation according to the occurrence of an input event instead of performing periodic dense calculation like a traditional neural network, and through a pulse-driven calculation mode, in the spiking neural network, the calculation efficiency is improved. The neuron can generate the pulse only when the membrane potential reaches the threshold value, the pulse-driven calculation mode can reduce unnecessary energy consumption, and the calculation efficiency can be further improved and the power consumption can be further reduced by reasonably designing the model and parameters of the neuron and optimizing the generation and propagation mechanism of the pulse.

Description

Technical Field

[0001] The present invention relates to the field of spiking neural network technology, and in particular to a low-power inference method and system for edge devices based on neuromorphic computing and spiking neural networks. Background Art

[0002] In recent years, neuromorphic vision sensors have great application prospects in unmanned technology, smart industry and other fields due to their advantages such as high dynamics, low power consumption and low redundancy. Deep learning has completely changed the field of machine learning, especially in computer vision. It uses backpropagation to train deep artificial neural networks in a supervised manner, but the training process requires a large number of labeled training samples, powerful computing power and long training time. However, artificial neurons and biological neurons have completely different methods of transmitting data and calculations. The former uses real-valued signals, while the latter uses discrete pulses to calculate and transmit information. In addition to the pulse firing rate, the pulse time is also important. Therefore, spiking neural networks are easier to implement biologically, more hardware-friendly and energy-efficient, and therefore attractive to portable devices.

[0003] The existing application number is 202410240906.4, which is a low-power automatic modulation recognition method based on a pulse neural network, belonging to the field of software radio. This method uses normalized pulse coding based on quantile thresholds to normalize floating-point signals and encode them into pulse sequence signals, thereby enhancing the generalization of neural networks and reducing the computational complexity and energy consumption of pulse neural networks. It adopts a dynamic time series training and testing scheme to improve the memory capacity and convergence speed of pulse neural network neurons through long time series training methods, and reduces the network inference delay and energy consumption in actual applications through short time series testing methods. It adopts pulse neural network technology oriented to residual connections to alleviate the degradation problem of deep pulse neural networks, enhance portability, and facilitate the rapid conversion between traditional neural networks and pulse neural networks. The present invention is applicable to the field of software radio, realizing low-power, high-precision, real-time automatic modulation recognition. Another example is a low-power target detection method based on a convolutional spiking neural network, filed with application number 202210550280.8. The method includes the following steps: establishing and preprocessing a weed image dataset; constructing a spiking neural network model for weed detection based on deep convolutional neural network conversion conditions; performing neuronal information transmission on the spiking neural network model's pulse sequence, and inputting the obtained leaky integral emission model neuron accumulation into the membrane voltage; training the spiking neural network model using the weed image dataset to achieve fine-grained normalization and set model parameters; and inputting the weed image to be identified into the trained spiking neural network model to obtain an image with a weed range frame. This method addresses issues such as low pulse emission frequency after conventional conversion and difficulty representing negative and biased values. The model's real-time and low-power consumption characteristics enable its application in embedded devices, mobile devices, and unmanned machines equipped with neuromorphic chips.

[0004] In summary, existing edge devices based on neuromorphic computing and spiking neural networks have been applied in many fields. However, during the long-term use of edge devices, due to the different functions monitored, some modules in the device cannot call the corresponding processing modules according to the different processing transactions, which easily leads to energy waste. To this end, we propose a low-power inference method and system for edge devices based on neuromorphic computing and spiking neural networks to solve the above problems. Summary of the Invention

[0005] The problem to be solved by the present invention is that due to different monitored functions, some modules in the device cannot call corresponding processing modules according to different processing transactions, which easily causes energy waste.

[0006] In order to solve the above technical problems, the present invention provides a low-power inference method for edge devices based on neuromorphic computing and spiking neural networks. The specific steps of the low-power inference method are as follows: S1: Build a spiking neural network model. Based on the specific task and data characteristics, determine the network structure, such as selecting the appropriate neuron model, number of network layers, and neuron connection method. Build the spiking neural network, and after training and optimization, deploy it on the system platform. S2: Data input and preprocessing: Real-time data is acquired through sensors and other devices, and input into the neuromorphic system platform where the spiking neural network is deployed. The data is then preprocessed and converted into a format suitable for spiking neural network input. Operations such as normalization and encoding can be performed to improve network performance and training results. S3: Low-power inference. Neuromorphic hardware is based on event-driven computing, sparse connectivity and computation, and pulse-driven computing. It performs computations only when data is input or when the state changes, significantly reducing power consumption. The hardware processes and analyzes the input data, and the spiking neural network transmits and computes information based on the input pulse signals, ultimately deriving the inference results. S4: Result output and application: output the inference results to the corresponding device or system to enable or disable specific applications; S5: Low power management: When no events occur, most neurons are inactive, and the system's power consumption is very low. Only when an event occurs and triggers the corresponding calculation does the corresponding hardware unit consume energy, effectively reducing the overall power consumption of the edge device; Preferably, the construction of the spiking neural network model in S1 specifically includes hardware selection and configuration, model training, model deployment and optimization; The hardware selection is to select an appropriate neuromorphic chip or processor based on the computing requirements, power consumption constraints, etc. of the edge device; The hardware configuration involves connecting and configuring the selected neuromorphic hardware with other components of the edge device, such as sensors, storage units, and communication modules, to build a complete hardware platform and ensure normal communication and collaborative work between the components. The model training is to use the preprocessed data to train the spiking neural network, adopt an improved training algorithm such as time-based backpropagation, and adjust the network parameters so that the network can accurately classify and predict the input data. The model deployment is to convert the trained spiking neural network model into a format executable by neuromorphic hardware and deploy it on the hardware platform; The model optimization is to optimize the deployed model according to the hardware characteristics and actual application requirements, such as using model compression technology to remove redundant connections and parameters in the network to reduce storage and computing requirements.

[0007] Preferably, the specific method of data preprocessing in S2 is: S201: Data cleaning, specifically including removing noise data and processing missing values, wherein the noise data removal refers to the erroneous or invalid data generated by the sensors of the edge devices due to interference from environmental factors, and the missing value processing refers to selecting an appropriate method to fill the missing values according to the characteristics of the data and the application scenario; S202: Data format conversion, specifically converting the data format and adjusting the data dimension according to the input requirements of the spiking neural network; S203: Data normalization, specifically processing the data so that the data is distributed in a suitable range, such as [0, 1] or [-1, 1]; S204: Data encoding: Encode the data according to the specific task and network requirements. For example, for a classification task, the category label can be independently encoded and converted into a format suitable for classification by the network output layer.

[0008] Preferably, the event-driven calculation in S3 includes: Event detection: At the edge device sensor level, it detects events caused by environmental changes. In dynamic vision sensors, when the brightness change of a pixel exceeds a set threshold, an event signal is generated. This signal contains information such as the location and time of the event. Event encoding and input: Detected events are encoded and converted into pulse signal inputs that can be processed by the spiking neural network. The frequency, phase and other parameters of the pulses are determined based on factors such as the intensity and time interval of the event. These pulses are then input into the first layer of neurons in the neural network. Neuron response and calculation: After receiving a pulse input, a neuron decides whether to activate based on its own pulse triggering mechanism. Only when the input pulse causes the neuron's membrane potential to reach the firing threshold will the neuron fire a pulse in response. After the neuron fires a pulse, its output pulse is transmitted to downstream neurons based on the network's connectivity. At the same time, the calculation process takes into account factors such as synaptic weights. Information propagation and reasoning: Pulses propagate in the network in an event-driven manner. That is, the network performs calculations and information propagation only when there is a pulse input generated by an event. In this process, through the processing of multiple layers of neurons, reasoning about event-related information is ultimately achieved. Energy consumption management: Since the entire computing process is event-driven, most neurons are inactive when no events occur, and the system power consumption is very low. Only when an event occurs and triggers pulse propagation and calculation, the corresponding hardware unit will consume energy, thereby effectively reducing the overall power consumption of edge devices.

[0009] The pulse drive calculation in S3 includes: Pulse generation and input: First, the external signal is converted into a pulse sequence. The changes in signal strength are encoded as pulses. A higher signal strength may correspond to a higher pulse frequency. These pulses are sent to the input layer neurons of the spiking neural network as the input of the network. Neurons receive pulses: After receiving pulses, neurons in the input layer process them according to the characteristics of the pulses and their own parameters. Each time a pulse is received, the membrane potential of the neuron changes accordingly based on the weight of the synaptic connection. Membrane potential update and pulse emission judgment: When a neuron receives multiple pulses, it continuously updates its membrane potential. When the membrane potential reaches the neuron's emission threshold, the neuron emits a pulse. This process is similar to how biological neurons work, integrating information from input pulses to determine whether to generate an output pulse. Pulse propagation and subsequent neuronal processing: When a neuron fires a pulse, it propagates to the next layer of neurons according to the network's connections. After receiving the pulse, the next layer of neurons repeats the previous steps, i.e., updating the membrane potential and determining whether to fire a pulse. This process continues layer by layer, and the pulse propagates through the neural network, achieving information transmission and processing. Inference output and adjustment: After the pulse is propagated and processed by the multi-layer neural network, the output pulse sequence of the last layer of neurons is used for the inference task. At the same time, according to the comparison between the output result and the target result, the parameters such as the synaptic weight in the network can be adjusted to optimize the inference performance. This adjustment process is also driven by the pulse signal. Preferably, the sparse connection and calculation in S3 include: Network topology construction: During the spiking neural network design phase, the sparse connectivity between neurons is determined. This can be achieved through a variety of methods, such as a probability-based connection method, where a connection probability is assigned to each possible connection. A connection is established between two neurons only when the randomly generated value is less than this probability. Alternatively, based on prior knowledge, for example, when processing image recognition tasks, only neurons corresponding to spatially adjacent or semantically related areas are connected. For example, in the connection between retinal ganglion cells and the primary visual cortex, only those connection paths that respond to adjacent visual areas are retained; Input data processing: When input data is received, it is converted into a pulse sequence suitable for processing by a spiking neural network. These pulse signals are transmitted only through established sparse connections. On edge devices, for example, after the image information collected by a smart camera is converted into a pulse sequence, only the neurons connected to it are activated. Local computation and propagation: In a sparsely connected network, neurons only receive pulse inputs from connected neurons. After receiving the pulses, neurons perform local computations based on their own dynamic models and update their own state variables, such as membrane potential. Output and inference: After calculation by multiple layers of sparsely connected neurons, the final output spike train is used for inference tasks.

[0010] The present invention also provides a low-power inference system for edge devices based on neuromorphic computing and spiking neural networks, including: Neuromorphic chip module: Able to efficiently run spiking neural networks, it must support asynchronous and sparse computing of spiking neural networks to achieve low power consumption and high performance; The spiking neural network module is composed of a large number of interconnected spiking neurons and is the core of the system's algorithm. Its neurons encode, transmit, and process information in the form of electrical pulses, which is closer to biological neurons. It can fully utilize the low power consumption characteristics of neuromorphic chips to achieve efficient information processing. Data preprocessing module: responsible for preprocessing input data, such as normalization and filtering, to improve data quality and reduce the complexity of subsequent processing. It is usually completed on edge devices, which can reduce data transmission volume and processing time, and improve the overall efficiency of the system; Sensor module: Responsible for collecting data from the external environment and providing input information to the system, such as visual sensors and auditory sensors. The data they collect is usually time-correlated and sparse, which matches the characteristics of pulse neural networks. It can be directly input into the neuromorphic chip for processing, reducing the power consumption of data conversion and preprocessing.

[0011] Preferably, the low-power reasoning system further includes: Storage module: used to store system parameters, intermediate results, and temporary data, including on-chip memory and external memory. On-chip memory is fast but has a small capacity, while external memory has a large capacity but is relatively slow. Storage resources need to be properly managed and allocated to meet the system's low power consumption and high performance requirements; Communication module: This module enables edge devices to communicate with other devices or systems, such as data transmission and interaction with cloud servers and other edge devices. Low-power communication protocols and technologies, such as Bluetooth and Zigbee, are required to reduce system power consumption. Power management module: Responsible for managing the system's power supply, monitoring and controlling the power consumption of each module. It can dynamically adjust the power supply according to the system's operating status. For example, it can reduce the clock frequency when the system is idle and shut down unnecessary modules, etc., to extend the battery life of edge devices and improve the system's energy efficiency.

[0012] Technical effects and advantages of the present invention: The present invention achieves compression and optimization of the model through hardware selection and configuration, model training, model deployment and optimization in constructing the pulse neural network model, which is beneficial to reducing power consumption. It also uses pruning technology to remove unimportant connections and neurons in the model, reducing the number of model parameters; quantizing the model, representing weights and activation values as low-precision data types, reducing storage and computing requirements, and thus reducing functional loss.

[0013] The present invention uses event-driven computing. Pulse neural networks naturally have event-driven characteristics. Edge devices can trigger computing based on the occurrence of input events, rather than performing periodic intensive computing like traditional neural networks. In image recognition tasks, when the dynamic visual sensor detects brightness changes in the scene, the pulse neural network will be triggered for processing, which can greatly reduce the power consumption of the device.

[0014] The present invention uses sparse connections and calculations, so the connections between neurons in the pulse neural network are sparse, and pulse signals are generated and transmitted to other neurons only when neurons are activated; in hardware implementation, this sparsity can be used to reduce unnecessary calculations and data transmission, thereby reducing power consumption. A special hardware architecture is used to support the storage and calculation of sparse matrices, and only connections with pulse signal transmission are processed.

[0015] The present invention adopts a pulse-driven computing mode. In the pulse neural network, neurons will only generate pulses when the membrane potential reaches a threshold. This pulse-driven computing mode can reduce unnecessary energy consumption. By rationally designing the neuron model and parameters, and optimizing the pulse generation and propagation mechanism, computing efficiency can be further improved and power consumption can be reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of the method flow structure of the present invention; Figure 2 Schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0017] The present invention provides a low-power inference method for edge devices based on neuromorphic computing and spiking neural networks, such as Figure 1 As shown, the specific steps of the low-power inference method are: S1: Build a spiking neural network model. Based on the specific task and data characteristics, determine the network structure, such as selecting the appropriate neuron model, number of network layers, and neuron connection method. Build the spiking neural network, and after training and optimization, deploy it on the system platform. S2: Data input and preprocessing: Real-time data is acquired through sensors and other devices, and input into the neuromorphic system platform where the spiking neural network is deployed. The data is then preprocessed and converted into a format suitable for spiking neural network input. Operations such as normalization and encoding can be performed to improve network performance and training results. S3: Low-power inference. Neuromorphic hardware is based on event-driven computing, sparse connectivity and computation, and pulse-driven computing. It performs computations only when data is input or when the state changes, significantly reducing power consumption. The hardware processes and analyzes the input data, and the spiking neural network transmits and computes information based on the input pulse signals, ultimately deriving the inference results. S4: Result output and application: output the inference results to the corresponding device or system to enable or disable specific applications; S5: Low power management. When no events occur, most neurons are inactive, and the system's power consumption is very low. Only when an event occurs and triggers corresponding calculations does the corresponding hardware unit consume energy, effectively reducing the overall power consumption of edge devices. In a preferred embodiment, by constructing a pulse neural network model, data input and preprocessing, low-power inference, result output and application, low-power management and other steps, it can be achieved that when no event occurs on the edge device, most neurons are in an inactive state, the system power consumption is very low, and only when an event occurs and triggers the corresponding calculation, the corresponding hardware unit will consume energy, thereby effectively reducing the overall power consumption of the edge device.

[0018] Furthermore, building a spiking neural network model in S1 specifically includes hardware selection and configuration, model training, model deployment and optimization; Hardware selection involves selecting an appropriate neuromorphic chip or processor based on the computing requirements and power consumption constraints of edge devices. Hardware configuration involves connecting and configuring the selected neuromorphic hardware with other components of the edge device, such as sensors, storage units, and communication modules, to build a complete hardware platform and ensure normal communication and collaborative work between components. Model training uses preprocessed data to train the spiking neural network, adopts improved training algorithms such as time-based backpropagation, and adjusts network parameters so that the network can accurately classify and predict input data. Model deployment is to convert the trained spiking neural network model into a format executable by neuromorphic hardware and deploy it on the hardware platform; Model optimization: Optimize the deployed model based on hardware characteristics and actual application requirements. For example, use model compression technology to remove redundant connections and parameters in the network to reduce storage and computing requirements.

[0019] In a preferred embodiment, an improved training algorithm is adopted to reduce computational complexity and energy consumption, such as adopting a backpropagation algorithm that replaces the gradient space-time to solve the training problem of the pulse neural network, so that the network can develop more deeply, thereby improving performance, and pruning the pulse neural network to remove unimportant connections and neurons to reduce storage and computing requirements. A neuromorphic chip is used as the processor of the edge device. It is based on an event-driven architecture and consumes energy only when changes in input data are detected. There is no need to continuously process large amounts of data, which significantly reduces power consumption. It can integrate computing, storage and sensor functions into a single chip, reducing component requirements and system complexity. It is suitable for space-constrained edge devices. It adopts a storage and computing integrated architecture to integrate storage units and computing units, and performs calculations directly in the storage units, reducing data movement and energy consumption.

[0020] Furthermore, the specific method of data preprocessing in S2 is: S201: Data cleaning, specifically including removing noise data and processing missing values. Removing noise data refers to the fact that sensors on edge devices may be affected by environmental factors and generate erroneous or invalid data. Processing missing values involves selecting appropriate methods to fill missing values based on the characteristics of the data and application scenarios. S202: Data format conversion, specifically converting the data format and adjusting the data dimension according to the input requirements of the spiking neural network; S203: Data normalization, specifically processing the data so that the data is distributed in a suitable range, such as [0, 1] or [-1, 1]; S204: Data encoding: Encode the data according to the specific task and network requirements. For example, for a classification task, the category label can be independently encoded and converted into a format suitable for classification by the network output layer.

[0021] In a preferred embodiment, by preprocessing the data and converting it into a format suitable for pulse neural network input, normalization, encoding and other operations can be performed to improve network performance and training effects, thereby providing accurate data for subsequent calculations.

[0022] Furthermore, event-driven computing in S3 includes: Event detection: At the edge device sensor level, it detects events caused by environmental changes. In dynamic vision sensors, when the brightness change of a pixel exceeds a set threshold, an event signal is generated. This signal contains information such as the location and time of the event. Event encoding and input: Detected events are encoded and converted into pulse signal inputs that can be processed by the spiking neural network. The frequency, phase and other parameters of the pulses are determined based on factors such as the intensity and time interval of the event. These pulses are then input into the first layer of neurons in the neural network. Neuron response and calculation: After receiving a pulse input, a neuron decides whether to activate based on its own pulse triggering mechanism. Only when the input pulse causes the neuron's membrane potential to reach the firing threshold will the neuron fire a pulse in response. After the neuron fires a pulse, its output pulse is transmitted to downstream neurons based on the network's connectivity. At the same time, the calculation process takes into account factors such as synaptic weights. Information propagation and reasoning: Pulses propagate in the network in an event-driven manner. That is, the network performs calculations and information propagation only when there is a pulse input generated by an event. In this process, through the processing of multiple layers of neurons, reasoning about event-related information is ultimately achieved. Energy consumption management: Since the entire computing process is event-driven, most neurons are inactive when no events occur, and the system power consumption is very low. Only when an event occurs and triggers pulse propagation and calculation, the corresponding hardware unit will consume energy, thereby effectively reducing the overall power consumption of edge devices.

[0023] In a preferred embodiment, spiking neural networks inherently possess event-driven computational properties, enabling edge devices to trigger computations based on input events, rather than the periodic, intensive computations required by traditional neural networks. For example, in image recognition tasks, spiking neural networks are triggered only when a dynamic vision sensor detects a change in scene brightness, significantly reducing device power consumption.

[0024] Furthermore, the pulse drive calculation in S3 includes: Pulse generation and input: First, the external signal is converted into a pulse sequence. The changes in signal strength are encoded as pulses. A higher signal strength may correspond to a higher pulse frequency. These pulses are sent to the input layer neurons of the spiking neural network as the input of the network. Neurons receive pulses: After receiving pulses, neurons in the input layer process them according to the characteristics of the pulses and their own parameters. Each time a pulse is received, the membrane potential of the neuron changes accordingly based on the weight of the synaptic connection. Membrane potential update and pulse emission judgment: When a neuron receives multiple pulses, it continuously updates its membrane potential. When the membrane potential reaches the neuron's emission threshold, the neuron emits a pulse. This process is similar to how biological neurons work, integrating information from input pulses to determine whether to generate an output pulse. Pulse propagation and subsequent neuronal processing: When a neuron fires a pulse, it propagates to the next layer of neurons according to the network's connections. After receiving the pulse, the next layer of neurons repeats the previous steps, i.e., updating the membrane potential and determining whether to fire a pulse. This process continues layer by layer, and the pulse propagates through the neural network, achieving information transmission and processing. Inference output and adjustment: After the pulses are propagated and processed through the multi-layer neural network, the output pulse sequence of the neurons in the last layer is used for the inference task. At the same time, based on the comparison between the output results and the target results, parameters such as the synaptic weights in the network can be adjusted to optimize the inference performance. This adjustment process is also driven by pulse signals.

[0025] In a preferred embodiment, a pulse-driven computing mode is used in a spiking neural network, where neurons generate pulses only when the membrane potential reaches a threshold. This pulse-driven computing mode can reduce unnecessary energy consumption. By rationally designing the neuron model and parameters, and optimizing the generation and propagation mechanism of pulses, computing efficiency can be further improved and power consumption can be reduced. Furthermore, sparse connections and computations in S3 include: Network topology construction: During the spiking neural network design phase, the sparse connectivity between neurons is determined. This can be achieved through a variety of methods, such as a probability-based connection method, where a connection probability is assigned to each possible connection. A connection is established between two neurons only when the randomly generated value is less than this probability. Alternatively, based on prior knowledge, for example, when processing image recognition tasks, only neurons corresponding to spatially adjacent or semantically related areas are connected. For example, in the connection between retinal ganglion cells and the primary visual cortex, only those connection paths that respond to adjacent visual areas are retained; Input data processing: When input data is received, it is converted into a pulse sequence suitable for processing by a spiking neural network. These pulse signals are transmitted only through established sparse connections. On edge devices, for example, after the image information collected by a smart camera is converted into a pulse sequence, only the neurons connected to it are activated. Local computation and propagation: In a sparsely connected network, neurons only receive pulse inputs from connected neurons. After receiving the pulses, neurons perform local computations based on their own dynamic models and update their own state variables, such as membrane potential. Output and inference: After calculation by multiple layers of sparsely connected neurons, the final output spike train is used for inference tasks.

[0026] In a preferred embodiment, through sparse connections and calculations, the connections between neurons in the spiking neural network are sparse, and pulse signals are generated and transmitted to other neurons only when neurons are activated. In hardware implementation, this sparsity can be used to reduce unnecessary calculations and data transmission, thereby reducing power consumption. For example, a dedicated hardware architecture is used to support the storage and calculation of sparse matrices, and only the connections with pulse signal transmission are processed. The present invention also provides a low-power inference system for edge devices based on neuromorphic computing and spiking neural networks, such as Figure 2 Shown, including: Neuromorphic chip module: Able to efficiently run spiking neural networks, it must support asynchronous and sparse computing of spiking neural networks to achieve low power consumption and high performance; The spiking neural network module is composed of a large number of interconnected spiking neurons and is the core of the system's algorithm. Its neurons encode, transmit, and process information in the form of electrical pulses, which is closer to biological neurons. It can fully utilize the low power consumption characteristics of neuromorphic chips to achieve efficient information processing. Data preprocessing module: responsible for preprocessing input data, such as normalization and filtering, to improve data quality and reduce the complexity of subsequent processing. It is usually completed on edge devices, which can reduce data transmission volume and processing time, and improve the overall efficiency of the system; Sensor module: Responsible for collecting data from the external environment and providing input information to the system, such as visual sensors and auditory sensors. The data they collect is usually time-correlated and sparse, which matches the characteristics of spiking neural networks. They can be directly input into the neuromorphic chip for processing, reducing the power consumption of data conversion and preprocessing. Storage module: used to store system parameters, intermediate results, and temporary data, including on-chip memory and external memory. On-chip memory is fast but has a small capacity, while external memory has a large capacity but is relatively slow. Storage resources need to be properly managed and allocated to meet the system's low power consumption and high performance requirements; Communication module: This module enables edge devices to communicate with other devices or systems, such as data transmission and interaction with cloud servers and other edge devices. Low-power communication protocols and technologies, such as Bluetooth and Zigbee, are required to reduce system power consumption. Power management module: Responsible for managing the system's power supply, monitoring and controlling the power consumption of each module. It can dynamically adjust the power supply according to the system's operating status. For example, it can reduce the clock frequency when the system is idle and shut down unnecessary modules, etc., to extend the battery life of edge devices and improve the system's energy efficiency.

[0027] The working principle of the present invention is as follows: by selecting and configuring the hardware, training the model, deploying and optimizing the model in constructing the spiking neural network model, the model is compressed and optimized, which is conducive to reducing power consumption. Pruning technology is used to remove unimportant connections and neurons in the model, reducing the number of model parameters. The model is quantized, and the weights and activation values are represented as low-precision data types, which reduces storage and computing requirements and thus reduces functional loss. Through event-driven computing, spiking neural networks are naturally event-driven. Edge devices can trigger computations based on input events, rather than performing periodic, intensive computations like traditional neural networks. For example, in image recognition tasks, spiking neural networks are triggered only when dynamic vision sensors detect brightness changes in the scene, significantly reducing device power consumption. Through sparse connections and computation, the connections between neurons in a spiking neural network are sparse, and pulse signals are generated and transmitted to other neurons only when a neuron is activated. In hardware implementation, this sparsity can be exploited to reduce unnecessary computation and data transmission, thereby lowering power consumption. For example, a specialized hardware architecture can be used to support the storage and computation of sparse matrices, processing only those connections where pulse signals are transmitted. Through the pulse-driven computing mode, in the spiking neural network, neurons will only generate pulses when the membrane potential reaches the threshold. This pulse-driven computing mode can reduce unnecessary energy consumption. By rationally designing the neuron model and parameters, and optimizing the pulse generation and propagation mechanism, computing efficiency can be further improved and power consumption can be reduced.

[0028] It will be understood that the present invention is described by way of some embodiments, and it will be appreciated by those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are intended to be protected by the present invention.

Claims

1. A low-power inference method for edge devices based on neuromorphic computing and spiking neural networks, characterized by: The specific steps of the low-power reasoning method are: S1: Build a spiking neural network model. Based on the specific task and data characteristics, determine the network structure, such as selecting the appropriate neuron model, number of network layers, and neuron connection method. Build the spiking neural network, and after training and optimization, deploy it on the system platform. S2: Data input and preprocessing: Real-time data is acquired through sensors and other devices, and input into the neuromorphic system platform where the spiking neural network is deployed. The data is then preprocessed and converted into a format suitable for spiking neural network input. Operations such as normalization and encoding can be performed to improve network performance and training results. S3: Low-power inference. Neuromorphic hardware is based on event-driven computing, sparse connectivity and computation, and pulse-driven computing. It performs computations only when data is input or when the state changes, significantly reducing power consumption. The hardware processes and analyzes the input data, and the spiking neural network transmits and computes information based on the input pulse signals, ultimately deriving the inference results. S4: Result output and application: output the inference results to the corresponding device or system to enable or disable specific applications; S5: Low power management. When no events occur, most neurons are inactive and the system's power consumption is very low. Only when an event occurs and triggers corresponding calculations will the corresponding hardware units consume energy, effectively reducing the overall power consumption of edge devices.

2. The low-power inference method for edge devices based on neuromorphic computing and spiking neural networks according to claim 1, characterized in that: The construction of the spiking neural network model in S1 specifically includes hardware selection and configuration, model training, model deployment and optimization; The hardware selection is to select an appropriate neuromorphic chip or processor based on the computing requirements, power consumption constraints, etc. of the edge device; The hardware configuration is a combination of the selected neuromorphic hardware and other components of the edge device; The model training is to train the pulse neural network using the preprocessed data; The model deployment is to convert the trained spiking neural network model into a format executable by neuromorphic hardware and deploy it on the hardware platform; The model optimization is to optimize the deployed model according to the hardware characteristics and actual application requirements.

3. The low-power inference method for edge devices based on neuromorphic computing and spiking neural networks according to claim 1, characterized in that: The specific method of data preprocessing in S2 is: S201: Data cleaning, specifically including removing noise data and processing missing values, wherein the noise data removal refers to the erroneous or invalid data generated by the sensors of the edge devices due to interference from environmental factors, and the missing value processing refers to selecting an appropriate method to fill the missing values according to the characteristics of the data and the application scenario; S202: Data format conversion, specifically converting the data format and adjusting the data dimension according to the input requirements of the spiking neural network; S203: Data normalization, specifically processing the data so that the data is distributed in a suitable range, such as [0, 1] or [-1, 1]; S204: Data encoding: Encode the data according to the specific task and network requirements. For example, for a classification task, the category label can be independently encoded and converted into a format suitable for classification by the network output layer.

4. The low-power inference method for edge devices based on neuromorphic computing and spiking neural networks according to claim 1, characterized in that: The event-driven computing in S3 includes: Event detection: At the edge device sensor level, it detects events caused by environmental changes. In dynamic vision sensors, when the brightness change of a pixel exceeds a set threshold, an event signal is generated. This signal contains information such as the location and time of the event. Event encoding and input: Detected events are encoded and converted into pulse signal inputs that can be processed by the spiking neural network. The frequency, phase and other parameters of the pulses are determined based on factors such as the intensity and time interval of the event. These pulses are then input into the first layer of neurons in the neural network. Neuron response and calculation: After receiving a pulse input, a neuron decides whether to activate based on its own pulse triggering mechanism. Only when the input pulse causes the neuron's membrane potential to reach the firing threshold will the neuron fire a pulse in response. After the neuron fires a pulse, its output pulse is transmitted to downstream neurons based on the network's connectivity. At the same time, the calculation process takes into account factors such as synaptic weights. Information propagation and reasoning: Pulses propagate in the network in an event-driven manner. That is, the network performs calculations and information propagation only when there is a pulse input generated by an event. In this process, through the processing of multiple layers of neurons, reasoning about event-related information is ultimately achieved. Energy consumption management: Since the entire computing process is event-driven, most neurons are inactive when no events occur, and the system power consumption is very low. Only when an event occurs and triggers pulse propagation and calculation, the corresponding hardware unit will consume energy, thereby effectively reducing the overall power consumption of edge devices.

5. The low-power inference method for edge devices based on neuromorphic computing and spiking neural networks according to claim 1, characterized in that: The pulse drive calculation in S3 includes: Pulse generation and input: First, the external signal is converted into a pulse sequence. The changes in signal strength are encoded as pulses. A higher signal strength may correspond to a higher pulse frequency. These pulses are sent to the input layer neurons of the spiking neural network as the input of the network. Neurons receive pulses: After receiving pulses, neurons in the input layer process them according to the characteristics of the pulses and their own parameters. Each time a pulse is received, the membrane potential of the neuron changes accordingly based on the weight of the synaptic connection. Membrane potential update and pulse emission judgment: When a neuron receives multiple pulses, it continuously updates its membrane potential. When the membrane potential reaches the neuron's emission threshold, the neuron emits a pulse. This process is similar to how biological neurons work, integrating information from input pulses to determine whether to generate an output pulse. Pulse propagation and subsequent neuronal processing: When a neuron fires a pulse, it propagates to the next layer of neurons according to the network's connections. After receiving the pulse, the next layer of neurons repeats the previous steps, i.e., updating the membrane potential and determining whether to fire a pulse. This process continues layer by layer, and the pulse propagates through the neural network, achieving information transmission and processing. Inference output and adjustment: After the pulses are propagated and processed through the multi-layer neural network, the output pulse sequence of the neurons in the last layer is used for the inference task. At the same time, based on the comparison between the output results and the target results, parameters such as the synaptic weights in the network can be adjusted to optimize the inference performance. This adjustment process is also driven by pulse signals.

6. The low-power inference method for edge devices based on neuromorphic computing and spiking neural networks according to claim 1, characterized in that: The sparse connection and calculation in S3 include: Network topology construction: During the spiking neural network design phase, determine the sparse connectivity between neurons; Input data processing: When input data is present, the data is converted into a spike train suitable for spiking neural network processing, and these spike signals are propagated only through established sparse connections; Local computation and propagation: In a sparsely connected network, neurons only receive pulse inputs from connected neurons. After receiving the pulses, neurons perform local computations based on their own dynamic models and update their own state variables, such as membrane potential. Output and inference: After calculation by multiple layers of sparsely connected neurons, the final output spike train is used for inference tasks.

7. A low-power inference system for edge devices based on neuromorphic computing and spiking neural networks, characterized by: include: Neuromorphic chip module: Able to efficiently run spiking neural networks, it must support asynchronous and sparse computing of spiking neural networks to achieve low power consumption and high performance; The spiking neural network module is composed of a large number of interconnected spiking neurons and is the core of the system's algorithm. Its neurons encode, transmit, and process information in the form of electrical pulses, which is closer to biological neurons. It can fully utilize the low power consumption characteristics of neuromorphic chips to achieve efficient information processing. Data preprocessing module: responsible for preprocessing input data; Sensor module: responsible for collecting data from the external environment and providing input information for the system.

8. The low-power inference system for edge devices based on neuromorphic computing and spiking neural networks according to claim 7, characterized in that: The low-power inference system further includes: Storage module: used to store system parameters, intermediate results, and temporary data, including on-chip memory and external memory. On-chip memory is fast but has a small capacity, while external memory has a large capacity but is relatively slow. Storage resources need to be properly managed and allocated to meet the system's low power consumption and high performance requirements; Communication module: enables communication between edge devices and other devices or systems; Power management module: responsible for managing the system's power supply, monitoring and controlling the power consumption of each module, and dynamically adjusting the power supply according to the system's working status.

Images