Event-driven signal reconstruction through neuromorphic compressive sensing

Conventional compressive sensing (CS) relies on dense, fixed-length measurements and continuous-valued matrix operations, which impose high computational and energy costs on resource-constrained edge devices. This paper asks whether signal sparsity can be leveraged more effectively by aligning it with the spatiotemporal sparsity of spiking neural networks (SNNs) across the entire sensing, transmission, and reconstruction pipeline.

The authors propose a neuromorphic CS framework centered on a spike-driven learned iterative shrinkage-thresholding algorithm (S-LISTA). At the transmitter, an SNN projects source observations into low-dimensional measurement spikes. At the receiver, S-LISTA reconstructs the original signal by accumulating these discrete spike updates across a fixed number of unfolded layers. The framework employs joint training of the measurement network, reconstruction network, and dictionary, with task supervision applied directly to the sparse code to ensure the retention of information relevant to downstream inference tasks.

The S-LISTA framework demonstrates substantial resource savings. On neuromorphic datasets like DVS-Gesture and Spiking Heidelberg Digits, it reduces transmitted data volume by over 96% compared to artificial neural network (ANN) baselines using 8-bit quantized measurements. Furthermore, at matched unfolded depths, the event-driven reconstruction requires significantly less estimated energy—up to 82 times less on certain datasets—while maintaining competitive reconstruction quality and downstream task performance. The framework also exhibits notable robustness under challenging channel conditions, such as unseen Rayleigh fading, where it often outperforms traditional ANN-based reconstruction methods.

This work establishes a direct correspondence between signal sparsity and computational sparsity, providing a viable paradigm for signal processing on edge devices with limited bandwidth and energy. By replacing dense matrix operations with event-driven spike accumulation, the framework offers a scalable solution for applications ranging from medical imaging to event-based vision, where efficiency is critical.