A Task-Specific Meta-Learning Framework for Few-Shot Sound Event Detection

Meta-learning is extensively used for few-shot learning. Prototypical Network (ProtoNet) has been proved to perform well for few-shot sound event detection. ProtoNet as a metalearning method consists of two stages: meta-train and metatest. During meta-train stage, an embedded network is trained on base-classes using episode training strategy. During metatest stage, the embedded network is transferred directly to unseen classes during meta-train. This style of transference is task-agnostic: the embedded network may not learn the optimal discrimination embedding features for specific tasks including unseen classes. In this paper, we propose a task-specific meta-learning framework (TSMLF) for few-shot sound event detection, which makes embedded network learn discrimination embedding features for specific tasks. TSMLF inherits the metatrain process of ProtoNet. During meta-test stage, the framework enables embedded network to learn discriminative embedding features by inter-class and intra-class differences. Concretely, we calculate the inter-class and intra-class distance that support set sound samples. Maximizing inter-class distance and minimizing intra-class distance (MIMI) are used as a criteria to fine-tune embedded network for specific tasks. In addition, due to the small-scaled support set of meta-test, similar sound samples are easily excessively clustered during fine-tuning. We set a distance constraint on intra-class distance to avoid overfitting of embedded network. The proposed framework is evaluated using few-shot dataset of Detection and Classification of Acoustic Scenes and Events challenges 2022 (DCASE2022) task 5. Extensive ablation experimental results validate that all components of TSMLF can provide positive effects on few-shot sound event detection.