Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Ternary Spike-based Neuromorphic Signal Processing System

Shuai Wang, Dehao Zhang, Ammar Belatreche, Yichen Xiao, Hongyu Qing, Wenjie We, Malu Zhang, Yang Yang

TL;DR

The paper tackles the resource constraints of deploying deep neural networks on edge devices by introducing a ternary spike-based neuromorphic signal processing system. It combines Threshold-Adaptive Encoding (TAE) to convert time-varying analog signals into sparse ternary spikes and a Dual-Scaling Factor Quantization ternary SNN (QT-SNN) to reduce memory and energy by quantizing membrane potentials and weights. The approach achieves state-of-the-art performance on speech and EEG tasks while delivering substantial memory (up to 94%) and energy (up to 7.5x) savings relative to comparable methods. These findings underscore the practical potential of neuromorphic, energy-efficient signal processing for real-time edge applications and point to future chip-based deployments.

Abstract

Deep Neural Networks (DNNs) have been successfully implemented across various signal processing fields, resulting in significant enhancements in performance. However, DNNs generally require substantial computational resources, leading to significant economic costs and posing challenges for their deployment on resource-constrained edge devices. In this study, we take advantage of spiking neural networks (SNNs) and quantization technologies to develop an energy-efficient and lightweight neuromorphic signal processing system. Our system is characterized by two principal innovations: a threshold-adaptive encoding (TAE) method and a quantized ternary SNN (QT-SNN). The TAE method can efficiently encode time-varying analog signals into sparse ternary spike trains, thereby reducing energy and memory demands for signal processing. QT-SNN, compatible with ternary spike trains from the TAE method, quantifies both membrane potentials and synaptic weights to reduce memory requirements while maintaining performance. Extensive experiments are conducted on two typical signal-processing tasks: speech and electroencephalogram recognition. The results demonstrate that our neuromorphic signal processing system achieves state-of-the-art (SOTA) performance with a 94% reduced memory requirement. Furthermore, through theoretical energy consumption analysis, our system shows 7.5x energy saving compared to other SNN works. The efficiency and efficacy of the proposed system highlight its potential as a promising avenue for energy-efficient signal processing.

Ternary Spike-based Neuromorphic Signal Processing System

TL;DR

The paper tackles the resource constraints of deploying deep neural networks on edge devices by introducing a ternary spike-based neuromorphic signal processing system. It combines Threshold-Adaptive Encoding (TAE) to convert time-varying analog signals into sparse ternary spikes and a Dual-Scaling Factor Quantization ternary SNN (QT-SNN) to reduce memory and energy by quantizing membrane potentials and weights. The approach achieves state-of-the-art performance on speech and EEG tasks while delivering substantial memory (up to 94%) and energy (up to 7.5x) savings relative to comparable methods. These findings underscore the practical potential of neuromorphic, energy-efficient signal processing for real-time edge applications and point to future chip-based deployments.

Abstract

Deep Neural Networks (DNNs) have been successfully implemented across various signal processing fields, resulting in significant enhancements in performance. However, DNNs generally require substantial computational resources, leading to significant economic costs and posing challenges for their deployment on resource-constrained edge devices. In this study, we take advantage of spiking neural networks (SNNs) and quantization technologies to develop an energy-efficient and lightweight neuromorphic signal processing system. Our system is characterized by two principal innovations: a threshold-adaptive encoding (TAE) method and a quantized ternary SNN (QT-SNN). The TAE method can efficiently encode time-varying analog signals into sparse ternary spike trains, thereby reducing energy and memory demands for signal processing. QT-SNN, compatible with ternary spike trains from the TAE method, quantifies both membrane potentials and synaptic weights to reduce memory requirements while maintaining performance. Extensive experiments are conducted on two typical signal-processing tasks: speech and electroencephalogram recognition. The results demonstrate that our neuromorphic signal processing system achieves state-of-the-art (SOTA) performance with a 94% reduced memory requirement. Furthermore, through theoretical energy consumption analysis, our system shows 7.5x energy saving compared to other SNN works. The efficiency and efficacy of the proposed system highlight its potential as a promising avenue for energy-efficient signal processing.
Paper Structure (17 sections, 13 equations, 8 figures, 4 tables, 3 algorithms)

This paper contains 17 sections, 13 equations, 8 figures, 4 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Neural Encoding for Signals
  4. Spiking Neural Networks
  5. Preliminaries
  6. Threshold-based Encoding for Raw Signals
  7. Ternary Spiking Neural network
  8. Method
  9. Threshold-Adaptive Encoding
  10. Dual-scaling Factor Quantization Ternary Spiking Neural Networks
  11. Experiment
  12. Datasets
  13. Performance of TAE Method
  14. Performance of QT-SNN
  15. Performance of Ternary Spike-based Neuromorphic Signal Processing System
  16. ...and 2 more sections

Figures (8)

  • Figure 1: Ternary spike-based neuromorphic signal processing system. (a) Threshold-adaptive encoding method. It can efficiently encode analog signal data into ternary spike trains, thereby reducing memory footprints and energy consumption on signal transmission. (b,c) Dual-scaling quantization ternary spiking neural network. It enhances the SNNs' performance and quantizes both synaptic weights and membrane potential to lower bit-width, significantly reducing the network's memory and computational resource requirements.
  • Figure 2: The threshold-based encoding method for raw signals. (a)The threshold-based encoding methods transform raw signals into ternary spike trains consisting of $\left\{-1,0,1\right\}$. (b) A larger threshold may result in notable reconstruction fluctuations in smooth signal regions. (c) Conversely, a smaller threshold may result in significant peak information loss..
  • Figure 3: The comparison between SF and TAE methods. (a) In SF coding, the fixed threshold results in a linear increase and decrease of the $base$ (indicated in red). (b) The TAE method employs an adaptive thresholding technique, allowing the base to conform to any given curve (red cave).
  • Figure 4: The distribution of weights and membrane potential. (a) In a full-precision ternary spike SNN, the membrane potential and weights follow normal distributions with a mean of zero and differing standard deviations. (b) The MINT method quantizes the membrane potential and weights into 4 bits using nonlinear mapping function $\tanh$, disrupting the normal distribution characteristics of the membrane potential.
  • Figure 5: A comparative performance analysis of different encoding methods on radar and GSC datasets. (a-c) depict the encoding and reconstruction capabilities of MW, SF, and our TAE method for a radar signal segment. (e-g) illustrate these encoding strategies applied to a segment of speech signal. (d) and (h) include statistical analysis of the reconstruction-MAE for the overall datasets.
  • ...and 3 more figures