SpikeLite: Lightweight Spiking Neural Networks for Time-Series Forecasting
SpikeLite uses frequency-selective spiking encoding and sparse channel attention for energy-efficient time-series forecasting.
SpikeLite is a lightweight spiking neural network framework for multivariate and long-term time-series forecasting. A Frequency-Selective Spiking Encoder uses LIF low-pass dynamics to reorganize sequences into frequency-sensitive components, while Sparse Spiking Channel Attention applies a learned binary mask inside spike-driven self-attention. Under the SeqSNN and SpikF protocols it reports the best aggregate scores, including average R² of 0.790 and RSE of 0.440, and the lowest long-term MSE/MAE of 0.343/0.345. On the ECL dataset it also reports the lowest energy consumption among compared methods.
- FSSE uses LIF low-pass behavior for frequency-sensitive temporal encoding.
- SSCA applies a binary mask to limit cross-channel spike attention.
- Best aggregate R² 0.790 and RSE 0.440 under both protocols.
- Lowest long-term MSE/MAE of 0.343/0.345; lowest ECL energy use.
From the original article
1 min read at arxiv.orgSpiking neural networks (SNNs) offer an energy-efficient paradigm for time-series forecasting through spike-driven computation. However, recent SNN forecasters often pursue higher accuracy through increasingly complex attention mechanisms, or specialized neuronal dynamics, weakening the lightweight motivation of SNNs.
By Bang Hu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.