Abstract
Sound event detection relies on frame-level strong labels, which are expensive to annotate. Active learning addresses this issue by selecting audio segments that most assist the classifier. One prevalent acquisition strategy, mismatch-first farthest-traversal (MFFT), combines disagreement between two classifiers and the diversity of selected segments through hard sequential decisions. It selects entire groups of high-disagreement segments first, distributing the remaining budget through farthest traversal.
We demonstrate on two multi-label datasets that this design is blind to the similarity among selected segments and performs poorly under low budgets, with every mismatch-first variant ending below the basic geometric strategy it builds upon. We propose mismatch-weighted facility location (MW-FL), which utilizes the entire budget through a disagreement-weighted coverage objective that penalizes similarity among selected segments. The disagreement signal from MFFT is employed to derive the nonnegative weights of this facility-location objective without introducing hyperparameters.
Experiments across two geometric mechanisms reveal that coverage of selected segments is the dominant factor, hard disagreement gating of selection is detrimental in both mechanisms, and soft disagreement weighting enhances performance on top of coverage. MW-FL achieves the best area under the learning curve on both datasets.
Blogger's Review: This paper introduces MW-FL, which optimizes traditional active learning strategies by effectively utilizing disagreement signals. It highlights how to enhance audio classification accuracy under budget constraints while considering segment similarity, indicating significant practical implications.