Abstract
Spiking neural networks (SNNs) process time-series data via internalevent-driven neural dynamics whose energy consumption depends on the number ofspikes exchanged between neurons over the course of the input presentation. Intypical implementations of an SNN classifier, decisions are produced after theentire input sequence has been processed, resulting in latency and energyconsumption levels that are fairly uniform across inputs. Recently introduceddelay-adaptive SNNs tailor the inference latency -- and, with it, the energyconsumption -- to the difficulty of each example, by producing an earlydecision when the SNN model is sufficiently ``confident''. In this paper, westart by observing that, as an SNN processes input samples, its classificationdecisions tend to be first under-confident and then over-confident with respectto the decision's ground-truth, unknown, test accuracy. This makes it difficultto determine a stopping time that ensures a desired level of accuracy. Toaddress this problem, we introduce a novel delay-adaptive SNN-based inferencemethodology that, wrapping around any pre-trained SNN classifier, providesguaranteed reliability for the decisions produced at input-dependent stoppingtimes. The approach entails minimal added complexity as compared to theunderlying SNN, requiring only thresholding and counting operations at runtime, and it leverages tools from conformal prediction (CP).