Spiking Bandpass Wavelets Bridge Neuromorphic Hardware and Signal Processing
New wavelets enable spike-based encoding with reconstruction accuracy rivaling classical transforms.
A new arXiv paper by Pedersen, Lindeberg, and Gerstoft (arXiv:2605.09770) fundamentally rethinks spike-based signal encoders by recasting them as time-causal wavelet frames with rigorous quantitative bandwidths and reconstruction error bounds. Traditionally, spike encoders have been formulated probabilistically, disconnected from most signal processing literature. This work bridges that gap by showing that the spiking representation is mathematically equivalent to a wavelet decomposition—each spike corresponds to the output of a bandpass filter with known center frequency and bandwidth. The wavelets are designed to be causal and time-recursive, making them suitable for real-time streaming data.
The authors demonstrate the method on ECG signals and audio datasets, achieving a normalized root-mean-square error (RMSE) comparable to continuous wavelet transforms (CWT) while maintaining the sparsity and locality of spiking representations. Reconstruction quality is limited only by spike quantization and time discretization. Crucially, the spiking wavelets map directly to neuromorphic hardware, meaning they can run on low-power event-driven chips. This opens up applications in energy-efficient sensing for wearable devices, biomedical implants, and edge AI—anywhere that requires low-latency, low-power processing of temporal signals.
- Recasts spike encoders as time-causal wavelet frames with explicit bandwidths and reconstruction error bounds.
- Achieves normalized RMSE comparable to continuous wavelet transforms on ECG and audio datasets.
- Directly maps to neuromorphic hardware for energy-efficient, sparsity-preserving temporal signal processing.
Why It Matters
Enables rigorous, energy-efficient spike-based signal processing for neuromorphic systems and edge AI.