🧬 biology

Encoding and Decoding Temporal Signals with Spiking Bandpass Wavelets

This paper bridges the gap between probabilistic spike-based encodings and signal processing by recasting them as time-causal wavelet frames with quantitative bandwidths and reconstruction error bounds, demonstrating their ability to accurately reconstruct signals like ECG and audio while mapping directly to energy-efficient neuromorphic hardware.

Original authors: Jens Egholm Pedersen, Tony Lindeberg, Peter Gerstoft

Published 2026-05-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jens Egholm Pedersen, Tony Lindeberg, Peter Gerstoft

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to record a conversation. The standard way computers do this (like your phone or a digital recorder) is to take a snapshot of the sound thousands of times every second, whether anyone is speaking or not. This is like taking a photo of a room every second, even when nothing is moving. It works, but it wastes a lot of energy and storage space on "silence."

Nature, however, is smarter. Your eyes and ears don't record everything constantly; they only send a signal when something changes. If a bird flies by, your brain gets a "blip." If the room is quiet, your brain stays silent. This is how "spiking" works: it's an event-driven system that only fires when there is news to report.

This paper introduces a new mathematical tool that lets computers do the same thing: Spiking Bandpass Wavelets.

Here is the breakdown of how it works, using simple analogies:

1. The Problem: The "Snapshot" vs. The "Event"

Standard digital recording is like a security camera that takes a photo every second. It captures the empty hallway just as clearly as the person walking through it.

The Paper's Goal: Create a system that only records the movement, not the empty hallway. This saves massive amounts of energy and data.

2. The Solution: The "Change Detector"

The authors created a new type of filter (a mathematical sieve) that acts like a highly sensitive motion detector.

The Wavelet: Think of a wavelet as a specific "shape" of a filter. The paper uses two special shapes called DoE (Difference of Exponentials) and DoT (Difference of Time-Causal Limit kernels).
The "Bandpass" Magic: Imagine you have a radio. A "bandpass" filter only lets you hear a specific range of voices (like a specific musical note) and blocks everything else. The authors' filters break a complex signal (like a heartbeat or a voice) into different "notes" or speeds.
- Some channels listen for fast changes (high pitch).
- Some listen for slow changes (low pitch).
- Crucially, they ignore the "silence" (the constant background hum).

3. The "Spiking" Mechanism: The Light Switch

Once the signal is broken down into these different "notes," the system decides whether to send a message.

The Threshold: Imagine a light switch that only clicks on if the room gets brighter than a certain level.
The Spike: If the "change" in the signal is strong enough to hit that threshold, the system sends a single, tiny electrical "spike" (a blip). If the change is too small, it stays silent.
The Result: Instead of a long stream of numbers, you get a sparse list of "blips" with timestamps. It's like sending a text message saying "Bird flew by at 2:03 PM" instead of sending a video of the empty room for an hour.

4. The "Time-Causal" Feature: No Time Travel

Many mathematical tools that analyze signals are "non-causal," meaning they cheat by looking at the future to understand the present (like a movie editor who sees the whole scene before cutting it).

The Paper's Claim: Their system is Time-Causal. It only looks at the past and the present. It reacts to changes as they happen, just like a real biological neuron. This makes it perfect for real-time hardware that can't wait for the future.

5. Rebuilding the Signal: The Puzzle

The big question is: If we only have a list of "blips," can we rebuild the original sound or heartbeat?

The Analogy: Imagine you have a puzzle where you only have the pieces that moved. The paper proves mathematically that you can reconstruct the original picture almost perfectly.
The Method: They use a "decoder" that takes those "blips" and spreads them back out, filling in the gaps based on the known rules of the filters.
The Result: They tested this on ECG (heartbeats) and Audio (speech). The reconstructed sound was nearly as clear as the original, with errors comparable to standard, heavy-duty digital methods, but using a much sparser, more efficient code.

6. Why This Matters (According to the Paper)

Efficiency: It mimics the brain's energy efficiency.
Hardware: Because the math relies on simple "leaky integrators" (like a bucket slowly filling and leaking), it maps directly onto neuromorphic hardware (chips designed to act like brains).
Guarantees: The authors didn't just guess; they provided mathematical proofs (bounds) showing that no matter what signal you throw at it, the reconstruction error will stay within a predictable, small limit.

In Summary:
The paper presents a new way to turn continuous signals (like sound or heartbeats) into a sparse list of "events" (spikes) that only happen when the signal changes. It does this using mathematically proven filters that respect the flow of time (causality). The result is a system that is as accurate as traditional digital recording but is designed to be incredibly energy-efficient and compatible with future "brain-like" computer chips.

Technical Summary: Encoding and Decoding Temporal Signals with Spiking Bandpass Wavelets

Problem Statement

Analog-to-digital conversion (ADC) in digital signal processing typically relies on uniform Nyquist-Shannon sampling, which indiscriminately samples signals regardless of their activity level, leading to inefficiency. In contrast, biological sensory systems utilize sparse, event-driven spike representations that are emitted only when input changes occur. While neuromorphic hardware aims to exploit this energy efficiency, a principled signal-processing framework for spike-based encoding is currently missing. Existing approaches, such as neuroscience-driven probabilistic models or single-channel sigma-delta modulators, often lack formal signal processing guarantees, multi-scale capabilities, or stable reconstruction bounds. There is a specific need for a wavelet frame where coefficients are emitted as spike events rather than continuous amplitudes, preserving the scale covariance and multi-resolution properties of classical wavelet theory while producing sparse, event-driven outputs.

Methodology

The authors propose recasting spike encoders as time-causal wavelet frames with quantitative bandwidths and reconstruction error bounds. The methodology involves two primary components:

Time-Causal Wavelet Families: The paper introduces two families of spiking wavelets based on the difference of scale-space representations:
- Difference of Truncated Exponentials (DoE): Constructed from a single leaky integrator per scale.
- Difference of Time-Causal Limit Kernels (DoT): Constructed by cascading leaky integrators to recover scale covariance and the cascade-smoothing properties of the Gaussian while maintaining causality.
- Both families are derived by taking adjacent differences of scale-space representations ( $L(t; \sigma_k) - L(t; \sigma_{k-1})$ ), ensuring the resulting bandpass kernels have zero mean and satisfy wavelet admissibility criteria.
Spiking Encoding and Decoding:
- Encoding: The continuous bandpass wavelet responses are quantized into signed spike trains using a Leaky Integrate-and-Fire (LIF) neuron model. To capture polarity, the integration is split into positive and negative parts, generating spikes when the membrane potential crosses a threshold $\theta_{thr}$ .
- Decoding: Reconstruction is achieved by passing the sparse spike trains through reconstruction kernels (leaky integrators) and summing the channel outputs. Since the dual frame operator has no closed-form solution for spiking implementations, the authors employ a least-squares approach to learn optimal weights for each channel, fitting the reconstruction to the original channel analysis output.

Key Contributions

The paper makes three primary contributions:

Two Spiking Wavelet Families: The definition of DoE and DoT wavelets as time-causal, scale-covariant encoders. The paper provides a theoretical analysis of the underlying non-spiking representations, proving they form overcomplete, non-tight frames.
Formal Bounds: The derivation of frame bounds ( $A$ and $B$ ) for both analog and spiking variants. Crucially, the authors derive a closed-form reconstruction error bound (Equation 32) for the spiking case, demonstrating that the error is linear in the spike threshold $\theta_{thr}$ and converges geometrically with the number of channels.
Experimental Validation: Reconstruction experiments on ECG (MIT-BIH) and audio (LibriSpeech) datasets. The results show that the spiking wavelets achieve normalized Root Mean Square Error (nRMSE) comparable to continuous wavelet transforms (like Morlet) and discrete wavelet transforms, despite the additional quantization error inherent to spiking.

Results

Reconstruction Accuracy: On the MIT-BIH ECG dataset, the spiking DoT wavelet achieved an nRMSE of 0.058, outperforming the non-causal Morlet baseline (0.060). On the LibriSpeech audio dataset, the spiking DoT achieved 0.064, matching the Morlet baseline.
Error Characteristics: The reconstruction error was observed to be linearly dependent on the firing threshold $\theta_{thr}$ , confirming the theoretical error bounds.
Causality vs. Performance: The causal spiking wavelets (DoE and DoT) performed competitively against non-causal continuous wavelets (Morlet) and significantly better than the causal Szu wavelet, which suffered a 3–6× performance penalty due to causality constraints in traditional frameworks.
Channel Overlap: The study noted that increasing the number of channels ( $K$ ) reduced error for non-causal DoG but increased error for causal DoE/DoT due to high channel coherence and ill-conditioning of the linear system, highlighting a trade-off between scale density and reconstruction stability.

Significance and Claims

The paper claims to establish a principled signal-processing framework for spike-based ADC that bridges the gap between biological efficiency and mathematical rigor. Its significance lies in:

Direct Neuromorphic Mapping: The operations (leaky integrators and thresholding) map directly to primitives found in neuromorphic hardware, positioning spiking wavelets as a fundamental building block for event-driven analog-to-digital conversion.
Theoretical Guarantees: Unlike previous probabilistic or information-theoretic spike models, this work provides formal frame bounds and reconstruction error guarantees that hold for any finite-energy input without assumptions on signal statistics or sparsity.
Causal Multi-Resolution: It successfully extends time-encoding machines to overcomplete, multi-scale frames while maintaining time-causality, a property often lost in classical wavelet approaches or difficult to achieve in causal signal processing.

The authors conclude that these spiking DoE and DoT wavelets offer a sparse, continuous-time encoding method with provable reconstruction guarantees, suitable for low-energy, low-latency signal processing on neuromorphic platforms.