💻 computer science

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

This paper proposes a spiking convolutional autoencoder (SCAE) jointly trained with a Spiking Neural Network to generate sparse, spike-encoded representations of channel impulse responses for energy-efficient human activity recognition, achieving comparable accuracy to hybrid approaches while eliminating the need for Doppler domain preprocessing.

Original authors: Eleonora Cicciarella, Riccardo Mazzieri, Jacopo Pegoraro, Michele Rossi

Published 2026-02-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Eleonora Cicciarella, Riccardo Mazzieri, Jacopo Pegoraro, Michele Rossi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to recognize what a person is doing in a room—walking, running, sitting, or waving—without ever seeing them. Instead of using a camera, you use radio waves. These waves bounce off the person and return to your device, creating a complex "echo map" called a Channel Impulse Response (CIR).

The problem is that reading these echo maps is like trying to understand a novel by reading every single letter, including the spaces and punctuation, all at once. It's too much work for small, battery-powered devices (like smart routers or sensors) to process this data quickly and efficiently.

Here is how the authors of this paper solved that problem, explained simply:

1. The Old Way: The Heavy Backpack

Traditionally, to understand these radio echoes, computers had to do a lot of heavy math first. They would translate the raw echo data into a "Doppler spectrogram" (a fancy heat map showing movement). This is like taking a raw, uncooked steak and grinding it into a fine powder before you can cook it. It takes a lot of energy and time, which drains the battery of small devices.

2. The New Idea: The "Spiking" Translator

The authors propose a new system using Spiking Neural Networks (SNNs). Think of a standard computer brain as a lightbulb that is always on, humming with electricity. An SNN is more like a Morse code operator. It only sends a signal (a "spike") when something important happens, and stays silent otherwise. This saves massive amounts of energy.

However, SNNs are picky eaters; they don't like raw, dense data. They need the data to be converted into these sparse "spikes" first.

3. The Solution: A Smart "Translator" (The SCAE)

The paper introduces a special tool called a Spiking Convolutional Autoencoder (SCAE). Think of this as a smart translator that sits between the radio waves and the brain.

What it does: It looks at the raw, messy radio echoes and instantly translates them into a sparse Morse code (spikes).
The Magic: Unlike old translators that just follow a rigid rulebook (like a simple "if the signal changes, send a spike"), this translator learns how to speak Morse code specifically for human movement. It figures out exactly which parts of the echo matter and ignores the rest.
The Result: It creates a code that is 81% empty space (sparsity). This means the "brain" only has to do work for 19% of the data, saving huge amounts of energy.

4. The Team-Up: Translator + Brain

The authors built a two-part system:

The Translator (SCAE): Converts the raw radio echoes into efficient Morse code.
The Brain (SNN): Reads the Morse code and decides, "Ah, that's someone running!" or "That's someone waving!"

They trained these two parts together. The translator learns to make the code easier for the brain to read, and the brain learns to recognize the patterns.

5. The Results: Fast, Lean, and Accurate

The paper tested this system against other methods:

Accuracy: It got the activity right about 96% of the time, which is just as good as the heavy, energy-hungry methods.
Efficiency: It was much "sparser" (emptier) than other methods. While other methods left about 70-72% of the data as zeros (empty), this new method left 81% as zeros.
Speed: It could make a decision in about 20 milliseconds (faster than a human blink).
Direct Comparison: When they tried to feed the raw radio data directly into the "brain" without the translator, the brain got confused and failed miserably (dropping accuracy to around 20-50%). This proves the translator is essential.

The Bottom Line

This paper shows that you don't need a supercomputer to recognize human activity using radio waves. By using a smart "translator" that turns complex radio echoes into efficient, sparse Morse code, you can build a system that is highly accurate but extremely energy-efficient. This makes it possible to put these smart sensors on small, battery-powered devices in homes or hospitals without worrying about them running out of power.

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

Problem Statement
Integrated Sensing and Communication (ISAC) systems offer pervasive monitoring capabilities, particularly for Human Activity Recognition (HAR), by leveraging Channel Impulse Responses (CIR). However, modern sensing algorithms, typically based on Deep Learning (DL), are often too computationally complex and energy-intensive for resource-constrained edge devices where RF data is collected. Furthermore, standard approaches often rely on heavy preprocessing steps, such as generating Doppler spectrograms or range-Doppler maps, which adds latency and computational overhead. While Spiking Neural Networks (SNNs) offer a promising alternative for energy efficiency due to their sparse, event-driven processing, existing SNN-based HAR solutions frequently rely on pre-processed radar features (e.g., micro-Doppler signatures) rather than raw CIR data, failing to fully exploit the sparsity of radio channel features.

Methodology
The authors propose an end-to-end processing pipeline that eliminates the need for Doppler-domain preprocessing by learning tailored spike-encoded representations directly from raw CIR data. The core of the methodology is a Spiking Convolutional Autoencoder (SCAE) jointly trained with a Spiking Neural Network (SNN) classifier.

Architecture:
- SCAE (Encoder-Decoder): The encoder takes raw CIR samples (comprising real and imaginary components across range bins and time windows) and processes them through 3D convolutional layers and Leaky Integrate-and-Fire (LIF) neurons. It outputs a binary spike train. The decoder attempts to reconstruct the original CIR from these spikes. This reconstruction task forces the encoder to preserve critical movement features while binarizing the data.
- SNN Classifier: The spike output from the SCAE is fed into a classifier consisting of fully connected layers of LIF neurons. The network predicts one of four human activities: walking, running, sitting/standing, and waving hands.
- Training: The system is trained end-to-end using a composite loss function: $L = L_{rec} + \gamma L_{class}$ . $L_{rec}$ is the Mean Squared Error (MSE) between the input and reconstructed CIR, while $L_{class}$ is the Cross-Entropy loss based on spike counts (rate coding). The authors employ surrogate gradients (arctangent approximation) to handle the non-differentiability of spike generation.
Comparative Baselines:
To validate the proposed approach, the authors compare it against:
- CAE-SNN: A Convolutional Autoencoder using standard ReLU neurons instead of LIF neurons, trained similarly.
- Delta Thresholding + SNN: A fixed-threshold encoding method based on temporal signal differences, fed into the SNN.
- Direct-SNN: An SNN classifier trained directly on raw CIR data without an encoding autoencoder.
- Baseline CNN: A standard CNN originally trained on Doppler spectrograms, retrained here on raw CIR data.

Key Contributions

Learned Spike Encoding for CIR: The paper introduces a learned spike encoding mechanism specifically tailored to CIR data. Unlike fixed thresholding or standard autoencoders, this method balances activity recognition accuracy with spike sparsity.
Elimination of Preprocessing: The framework processes raw CIR data directly, removing the computational burden of generating Doppler maps or spectrograms.
Joint Optimization: By jointly training the autoencoder and the classifier, the system learns a spike representation that is optimal for both signal reconstruction and classification, rather than treating encoding as a separate, static step.
Performance vs. Efficiency Trade-off: The study demonstrates that encoding CIR data prior to classification significantly improves both HAR accuracy and energy efficiency compared to direct processing or standard CNN approaches.

Experimental Results
Experiments were conducted using the DISC dataset, which contains CIR measurements from 60 GHz IEEE 802.11ay signals capturing four activities performed by seven subjects.

Accuracy: The proposed SCAE-SNN achieved an F1 score of 95.75%, comparable to the CAE-SNN (95.16%) and significantly outperforming the Delta Thresholding approach (74.59%) and Direct-SNN variants (which dropped to ~50% or lower). Notably, the baseline CNN, when retrained on raw CIR data, suffered a performance drop of approximately 10 percentage points compared to its original performance on Doppler spectrograms, highlighting the difficulty of learning directly from raw signals without the proposed architecture.
Sparsity: The SCAE produced the sparsest spike representation with 81.1% sparsity. This is substantially higher than the CAE (28.6%) and the Delta Thresholding method (71.7%). Higher sparsity implies fewer computations and lower energy consumption on neuromorphic hardware.
Efficiency: The SCAE-SNN model is compact (28k parameters) and achieves an inference time of approximately 19.8 ms. While the encoding stage adds latency, the high sparsity and event-driven nature of the SNN suggest significant energy savings in deployment.
Temporal Resolution: The authors analyzed the trade-off between the number of timesteps ( $T$ ) and F1 score, finding that performance saturates around $T=29$ , which was selected as the optimal setting for deployment.

Significance and Claims
The paper claims that its proposed SCAE-SNN architecture offers a practical solution for real-time, energy-efficient HAR on edge devices. By learning a sparse, tailored spike encoding directly from raw CIR data, the system achieves high accuracy (nearly 96% F1) while drastically reducing computational load through sparsity (81.1%). The authors emphasize that this approach overcomes the limitations of both high-complexity DL models and inefficient fixed preprocessing methods, making it suitable for the emerging ISAC paradigm. The work suggests that future research should extend this framework to multi-antenna systems and multi-user scenarios.