The article presents a new method for class-incremental learning of potentially overlapping sounds for multi-label audio classification tasks. The proposed system learns new sound classes independently of the old ones, using a cosine similarity-based distillation loss to minimize discrepancy in the feature representations of subsequent learners. Experiments were performed on a dataset with 50 sound classes, and the proposed method achieved an average F1-score of 40.9% over five phases. The paper focuses on the challenges of incremental learning, particularly in the context of audio classification.
Publication date: 11 Jan 2024
Project Page: Not Provided
Paper: https://arxiv.org/pdf/2401.04447
