Spiking Neural Network Multi-Synapse Projection for Temporal Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Spiking Neural Networks (SNNs) face challenges in achieving high accuracy and efficient hardware resource utilization, particularly in implementing temporal convolution, which leads to a gap in performance compared to traditional Artificial Neural Networks (ANNs) and increased hardware resource consumption.
Innovation Solution
The implementation of a computing device with a spiking neural network that uses multi-synapse projections with different synaptic time constants and transmission delays to reduce hardware resource consumption and improve accuracy, enabling efficient temporal convolution processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional Artificial Neural Networks (ANNs) are used to achieve high accuracy, then classification accuracy is improved, but energy consumption and storage space consumption increase significantly
Solution Approach 1:
The patent replaces traditional ANN mechanical computing systems with a Spiking Neural Network (SNN) system that mimics biological neural mechanisms. The SNN uses event-driven spike transmission instead of continuous computation, where neurons only activate when receiving sufficient input signals. This substitution reduces energy consumption by eliminating redundant computations while maintaining classification accuracy through temporal coding and spike timing mechanisms.
Solution Approach 2:
The patent changes the fundamental operating parameters from continuous analog values in ANN to discrete spike events in SNN. By using spike timing, frequency, and pattern encoding instead of continuous activation values, the system achieves the same information processing with significantly reduced energy consumption and storage requirements, as spikes are sparse and event-driven rather than continuous.
2Use of energy by moving object
If Spiking Neural Networks (SNNs) are used to reduce energy consumption, then energy efficiency is improved, but accuracy and hardware resource utilization deteriorate
Solution Approach 1:
The patent adds the temporal dimension to SNN computation by utilizing spike timing information. Instead of relying solely on spatial neuron connections, the system encodes information in the timing, frequency, and sequence of spikes. This temporal coding allows the SNN to achieve higher accuracy by capturing dynamic patterns and temporal relationships in the data, bridging the accuracy gap with ANN while maintaining energy efficiency.
Solution Approach 2:
The patent implements preprocessing and feature extraction stages that prepare data in formats optimized for SNN processing. By pre-processing input data to highlight temporal patterns and converting them into spike train representations before SNN inference, the system maximizes the accuracy potential of SNN while maintaining its energy efficiency advantage.
3Use of energy by moving object
If Spiking Neural Networks (SNNs) are used to reduce energy consumption, then energy efficiency is improved, but hardware resource consumption increases
Solution Approach 1:
The patent segments the SNN system into modular functional units including spike generation modules, synaptic integration modules, and spike transmission modules. Each module is independently optimized and can be selectively activated based on computational needs. This segmentation reduces hardware resource consumption by avoiding the need to maintain all neural components in active states continuously, allowing the system to achieve energy efficiency through selective module activation while managing hardware complexity.
Data Source
AI summary
A computing device includes computing modules, each of which comprises a plurality of neuron populations. The computing module is configured to project the input spike train which is weighted by a first weight matrix to a first neuron population through a multi-synapse projection. The multi-synapse projection has at least two different and positive synaptic time constants, or at least two different synaptic transmission delays. In order to realize time-domain convolution in a spiking neural network with low hardware resource consumption, a multi-synaptic projection technology means with different synaptic time constants is proposed. On this basis, a waveform-aware spike neural network for time-domain signal processing characterized by residual connections and skip connections is further proposed. Through these technical means, the performance gap between SNN and ANN is bridged, and SNN with performance reaching or close to ANN is obtained.