Incremental Learning for Audio Classification using Pre-Trained Models

Abstract

Deep learning models have long been used for audio classification tasks, but when it comes to training for incremental tasks, they frequently forget what they have already learned. Catastrophic forgetting is a phenomenon that makes incremental learning tasks more difficult. In order to minimize the substantial loss of previously learned information, this bachelor's thesis investigates the use of pre-trained models, such as PANNs, for incremental learning in audio classification. The effectiveness of pre-trained models for audio classification in incremental learning is first reviewed in the thesis. In order to assess the models, it then creates a series of classification tasks using the ESC-50 dataset. The trade-offs between knowledge retention and task adaptability are examined using the results. Performance degradation on previously learned tasks persists even though this method reduces catastrophic forgetting when compared to traditional methods. The results point to possible directions for further study and aid in the development of more effective incremental learning techniques for audio classification applications.