SNAP-UQ: Self-supervised Next-Activation Prediction for Single-Pass Uncertainty in TinyML

ICLR 2026 Conference Submission480 Authors

Keywords: TinyML, uncertainty quantification, single-pass inference, depth-wise next-activation prediction, selective prediction, on-device monitoring

TL;DR: SNAP-UQ is a single-pass, label-free TinyML UQ: int8 heads predict next-layer stats and a monotone mapper yields risk; no buffers/exits/ensembles; +few tens of KB, ~40–60% smaller and ~25–35% faster, better drop detection and ~0.9 AUROC.

Abstract: We introduce SNAP-UQ, a single-pass, label-free uncertainty method for TinyML that estimates risk from depth-wise next-activation prediction: tiny int8 heads forecast the statistics of the next layer from a compressed view of the previous one, and a lightweight monotone mapper turns the resulting surprisal into an actionable score. The design requires no temporal buffers, auxiliary exits, or repeated forward passes, and adds only a few tens of kilobytes to MCU deployments. Across vision and audio backbones, SNAP-UQ consistently reduces flash and latency relative to early-exit and deep ensembles (typically $\sim$40--60\% smaller and $\sim$25--35\% faster), with competing methods of similar accuracy often exceeding memory limits. In corrupted streams it improves accuracy-drop detection by several AUPRC points and maintains strong failure detection (AUROC $\approx$0.9) in a single pass. Grounding uncertainty in layer-to-layer dynamics yields a practical, resource-efficient basis for on-device monitoring in TinyML.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Submission Number: 480