Memory-based Self-Explainable Early Exit Networks

Abstract

Early-exit networks speed up inference by letting intermediate classifiers make predictions. This paper makes those predictions easier to inspect by using sparse attention to retrieve examples from memory. The retrieved examples help each classifier make its prediction and show which examples influenced it. In experiments, the approach preserves more of the black-box baselines’ predictive performance than the other self-explainable designs tested.

Publication
Accepted at the Workshop on On-Device Intelligence: Foundation Models under Real-World Constraints (NeurIPS 2026)
Date
Links

Related projects

An extended abstract was also presented at BayLearn 2026.