Cover First, Disagree Softly: Rethinking Mismatch-First Active Learning for Frame-Level Audio Classification

Sound event detection (SED) requires expensive frame-level strong labels. Active learning (AL) aims to reduce this cost by selecting the most informative audio segments for annotation. This paper investigates the prevailing 'mismatch-first' acquisition strategies, which combine disagreement-based selection with geometric diversity, and identifies why they often fail to outperform simpler methods at low annotation budgets.

The authors factorize existing mismatch-first farthest-traversal (MFFT) strategies into two design axes: the geometric backbone (how the budget is spent) and the disagreement mechanism (how model uncertainty is used). They identify two primary failure modes: hard gating, which causes the model to oversample redundant segments in high-disagreement clusters, and farthest traversal, which tends to select uninformative outliers at low budgets. To address these, they introduce Mismatch-Weighted Facility Location (MW-FL), which replaces hard gating with a submodular coverage objective that uses disagreement signals as soft weights. This approach penalizes redundancy by shrinking the marginal gain of segments similar to those already selected.

Experiments across two multi-label datasets show that the geometric backbone is the dominant factor in performance, with facility location consistently outperforming farthest traversal. The authors find that hard disagreement gating is universally harmful, as it forces the selection of redundant segments. In contrast, soft disagreement weighting effectively modulates coverage, allowing MW-FL to achieve the best area under the learning curve (AULC) across all tested configurations. The study concludes that disagreement signals should be used to tilt coverage rather than override it through hard selection rules.

This work provides a clear diagnosis of why common hybrid active learning strategies fail in audio tasks. By shifting from hard-gated selection to a submodular, hyperparameter-free framework, researchers can improve the efficiency of data labeling without the risk of oversampling redundant or uninformative audio segments.