Title: SpurAudio: A Benchmark for Investigating Shortcut Learning in Few-Shot Audio Classification

Abstract:

Few-shot classification (FSC) is a prevalent approach for learning from scarce labeled data; however, standard evaluations typically operate under the implicit assumption that target concepts are independent of contextual cues. In practical scenarios, examples are frequently embedded within complex contexts, enabling models to leverage spurious correlations between foreground content and background signals. While this phenomenon has been examined in few-shot image classification, its impact on few-shot audio classification remains largely unexplored, with existing audio benchmarks providing insufficient control over contextual structures.

To address this gap, we introduce SpurAudio, a benchmark that capitalizes on the inherent separability of foreground events and background environments in audio. This tool facilitates a controlled, multi-level assessment of contextual shifts across both support and query sets. Our analysis reveals that numerous state-of-the-art few-shot methods experience significant performance declines when background correlations are altered, even though they maintain comparable accuracy under traditional evaluation protocols. Notably, this vulnerability is not limited by backbone capacity, as it persists even in large pretrained audio foundation models.

Furthermore, our findings demonstrate that methods appearing similar under conventional benchmarks can display vastly different sensitivities to spurious correlations. This discrepancy highlights systematic algorithmic strengths and weaknesses linked to how feature representations interact with classifier heads during inference. These insights offer a deeper understanding of few-shot method behavior in the audio domain and underscore the necessity for benchmarks that explicitly test context dependence when evaluating FSC models.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC