Transductive Feature Space Regularization for Few-shot Bioacoustic Event Detection

In few-shot bioacoustic event detection, besides interested target events, background noises and various uninterested sound events lead to complex decision boundaries, which require regularized feature distributions in feature space. Due to the low label availability of uncertain noise events, existing few-shot learning methods with entropy-based regularizers suffer from overfitting during optimization. In this paper, we propose a transductive inference model with a prior knowledge based regularizer (PKR) to overcome the overfitting problem. We use a task-adaptive feature extractor to reconstruct a regularized feature space. A PKR is proposed to minimize the divergence between the original and reconstructed feature space. The development set of DCASE 2022 Task 5 is adopted as the experimental dataset. With the increasing iterations, the proposed model performs with long-lasting results around 55.43 F-measure, and well solves the overfitting problem in transductive inference.