Minutes to Converage: Dataset Distillation for Rapid SNN Training on Event Streams

Shuhan Ye, Yi Yu, Qixin Zhang, Chenqi Kong, Qiangqiang Wu, Kun Wang, Dacheng Tao, Xudong Jiang

Keywords: Dynamic Vision Sensors; Dataset Distillation; Spiking Neural Networks

Abstract:

Event cameras emit sparse, polarity-signed streams that align with how spiking
neural networks compute in time, yet image-centric dataset distillation trans-
fers poorly to this regime. We present PACE (Phase-Aligned Condensation for
Events), the first event-native dataset distillation framework for SNNs, which com-
prises two core modules: ST-DSM and PEQ-N. ST-DSM densifies spikes with
residual membrane potential and aligns real and synthetic streams by matching
amplitude and phase using a characteristic-function projection in feature space
and a discrete Fourier transform along time. PEQ-N is a probabilistic quantizer
whose forward pass emits hard integer frames while a straight-through estimator
preserves gradients and keeps compatibility with standard event-frame pipelines.
We optimize only the synthetic data with a time-expanded condensation objec-
tive on frozen teacher features, which encourages causal spatiotemporal structure
and shortens convergence time. On DVS-Gestures with IPC=10 at 9.29% of the
data, PACE reaches 76.5%, about 89% of full-data performance and +20.4 points
over a strong baseline. Similar gains appear on CIFAR10-DVS and N-MNIST and
transfer across SNN backbones. PACE delivers compact, accurate surrogates that
reduce storage and wall-clock time and make minutes-to-converge training practi-
cal on neuromorphic streams while opening a path to efficient on-device learning
and reproducible distilled benchmarks.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 2427