MW-FL beats MFFT: New active learning method for audio classification

Active learning aims to reduce annotation costs by selecting the most informative data to label. In sound event detection, frame-level strong labels are expensive, so efficient acquisition strategies are critical. The prevailing approach, mismatch-first farthest-traversal (MFFT), first picks entire groups of high-disagreement segments, then spreads remaining budget via farthest traversal. However, this hard sequential design often selects similar segments, especially under low budgets, leading to poor performance. In a new preprint, Shiqi Zhang and Tuomas Virtanen from Tampere University analyze MFFT's flaws on two multi-label datasets and propose a better alternative: mismatch-weighted facility location (MW-FL).

MW-FL replaces hard disagreement gating with a disagreement-weighted coverage objective that penalizes similarity across all selected segments. It uses MFFT's disagreement signal as nonnegative weights in a facility-location formulation, introducing no extra hyperparameters. Experiments across two geometric mechanisms and three ways of using disagreement reveal that coverage is the dominant factor; hard disagreement gating harms both mechanisms, while soft disagreement weighting helps on top of coverage. MW-FL consistently achieves the best area under the learning curve, offering a practical upgrade for audio classification tasks where annotation budgets are tight.