Adaptive Fine-Grained Pruning via Binary Search for Efficient Environmental Sound Classification

The overparameterization of pretrained transformers limits their efficiency in Environmental Sound Classification (ESC), incurring high memory and computational overhead. Pruning redundant parameters is a common remedy, yet fine-grained pruning at the intra-layer level remains difficult due to the large combinatorial search space and the risk of accuracy degradation. In this paper, we propose Adaptive Fine-Grained Pruning (AFP), an adaptive pruning method operating at intra-layer granularity under a user-defined performance constraint. AFP estimates the importance of prunable components by accumulating gradients during fine-tuning and then performs a binary search over the importance-ranked substructure space in a coarse-to-fine manner to identify the smallest submodel that satisfies the constraint. AFP avoids post-pruning gradient recalibration and efficiently identifies the smallest submodel via binary search, without traversing the full substructure space. Experiments show that AFP adaptively achieves up to 66.56% and 45.84% parameter reductions on the standard Audio Spectrogram Transformer for ESC-50 and UrbanSound8K, respectively, within a 2% drop in accuracy, outperforming popular baselines of static pruning or layer-level pruning in both pruning effectiveness and search efficiency.