HYBRID ATTENTION-BASED PROTOTYPICAL NETWORKS FOR FEW-SHOT SOUND CLASSIFICATION

In recent years, prototypical networks have been widely usedin many few-shot learning scenarios. However, as a metric-based learning method, their performance often degrades inthe presence of bad or noisy embedded features, and outliersin support instances. In this paper, we introduce a hybrid at-tention module and combine it with prototypical networks forfew-shot sound classification. This hybrid attention moduleconsists of two blocks: a feature-level attention block, andan instance-level attention block. These two attention mech-anism can highlight key embedded features and emphasizecrucial support instances respectively. The performance ofour model was evaluated using the ESC-50 dataset and thenoiseESC-50 dataset. The model was trained in a 10-way5-shot scenario and tested in four few-shot cases, namely 5-way 1-shot, 5-way 5-shot, 10-way 1-shot, and 10-way 5-shot.The results demonstrate that by adding the hybrid attentionmodule, our model outperforms the baseline prototypical net-works in all four scenarios.