Transferability of Minimal Audio Sufficiencies Across Models

Determine whether a minimal sufficient audio signal—that is, a minimal subset of an input audio signal that alone yields the same top-1 class on a given audio classifier—is transferable across other audio classification models in the sense that those models assign the same class to that minimal signal.

Background

The paper introduces transferability analysis for audio classification and studies whether minimal sufficient signals identified on one model are accepted by other models as yielding the same class. While analogous questions have been explored in other domains (e.g., images), the transferability of minimal audio sufficiencies had not been established.

This question is motivated by the possibility that different models may rely on different cues in high-dimensional data such as audio, implying that minimal signals sufficient for one model’s decision may or may not generalize across architectures and training regimes.