Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Sparse Spike Encoding of Channel Responses for Energy Efficient Human Activity Recognition

Eleonora Cicciarella, Riccardo Mazzieri, Jacopo Pegoraro, Michele Rossi

TL;DR

This work tackles energy-efficient human activity recognition in ISAC-enabled wireless environments by removing Doppler-domain preprocessing and learning spike-based representations of channel impulse responses (CIR). An end-to-end architecture combining a spiking convolutional autoencoder (SCAE) with a spiking neural network (SNN) classifier learns tailored spike encodings, achieving a macro-F1 around $95.75\%$ while producing highly sparse spike trains (~$81.1\%$ sparsity). The approach outperforms delta-threshold encoding and direct-CIR SNN baselines, while using a compact model (~28k parameters) and favorable inference latency, highlighting its practicality for edge devices. These results demonstrate the viability of spike-based, end-to-end learning from raw CIR data for real-time HAR with substantial energy savings, and point to future work in extending to multi-antenna and multi-user ISAC scenarios.

Abstract

ISAC enables pervasive monitoring, but modern sensing algorithms are often too complex for energy-constrained edge devices. This motivates the development of learning techniques that balance accuracy performance and energy efficiency. Spiking Neural Networks (SNNs) are a promising alternative, processing information as sparse binary spike trains and potentially reducing energy consumption by orders of magnitude. In this work, we propose a spiking convolutional autoencoder (SCAE) that learns tailored spike-encoded representations of channel impulse responses (CIR), jointly trained with an SNN for human activity recognition (HAR), thereby eliminating the need for Doppler domain preprocessing. The results show that our SCAE-SNN achieves F1 scores comparable to a hybrid approach (almost 96%), while producing substantially sparser spike encoding (81.1% sparsity). We also show that encoding CIR data prior to classification improves both HAR accuracy and efficiency. The code is available at https://github.com/ele-ciccia/SCAE-SNN-HAR.

Sparse Spike Encoding of Channel Responses for Energy Efficient Human Activity Recognition

TL;DR

This work tackles energy-efficient human activity recognition in ISAC-enabled wireless environments by removing Doppler-domain preprocessing and learning spike-based representations of channel impulse responses (CIR). An end-to-end architecture combining a spiking convolutional autoencoder (SCAE) with a spiking neural network (SNN) classifier learns tailored spike encodings, achieving a macro-F1 around while producing highly sparse spike trains (~ sparsity). The approach outperforms delta-threshold encoding and direct-CIR SNN baselines, while using a compact model (~28k parameters) and favorable inference latency, highlighting its practicality for edge devices. These results demonstrate the viability of spike-based, end-to-end learning from raw CIR data for real-time HAR with substantial energy savings, and point to future work in extending to multi-antenna and multi-user ISAC scenarios.

Abstract

ISAC enables pervasive monitoring, but modern sensing algorithms are often too complex for energy-constrained edge devices. This motivates the development of learning techniques that balance accuracy performance and energy efficiency. Spiking Neural Networks (SNNs) are a promising alternative, processing information as sparse binary spike trains and potentially reducing energy consumption by orders of magnitude. In this work, we propose a spiking convolutional autoencoder (SCAE) that learns tailored spike-encoded representations of channel impulse responses (CIR), jointly trained with an SNN for human activity recognition (HAR), thereby eliminating the need for Doppler domain preprocessing. The results show that our SCAE-SNN achieves F1 scores comparable to a hybrid approach (almost 96%), while producing substantially sparser spike encoding (81.1% sparsity). We also show that encoding CIR data prior to classification improves both HAR accuracy and efficiency. The code is available at https://github.com/ele-ciccia/SCAE-SNN-HAR.
Paper Structure (18 sections, 6 equations, 3 figures, 2 tables)

This paper contains 18 sections, 6 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Background
  4. Wireless channel modeling
  5. Spiking neuron model
  6. Dataset
  7. Models and methods
  8. SCAE-SNN architecture
  9. CAE-SNN architecture
  10. Delta thresholding encoding
  11. Direct-SNN
  12. Training method
  13. Experimental results
  14. Accuracy results
  15. Sparsity of the encoding
  16. ...and 3 more sections

Figures (3)

  • Figure 1: Block diagram of the proposed SCAE-SNN architecture.
  • Figure 2: Confusion matrix of the SCAE-SNN model on the test set.
  • Figure 3: Inference time vs. F1 score across timesteps. Shaded areas represent $\pm$ one standard deviation computed over ten seeds.