Episodic memory based continual learning without catastrophic forgetting for environmental sound classification

Said Karam, Shanq-Jang Ruan, Qazi Mazhar ul Haq, Lieber Po-Hung Li

Abstract: Convolutional neural network suffers from catastrophic forgetting during continual learning. This is one of the major obstacles for artificial intelligence, to solve new problems without forgetting the previously learned information. In this article, we propose an episodic memory technique for learning sound data incrementally. The proposed method observes tasks sequentially and successfully solves the new task without forgetting the previous task. The results show that the proposed method is able to transfer backward and forward knowledge efficiently. The performance evaluation demonstrates that the proposed method achieves better performance than other benchmarks. For ESC-50 and UrbanSound8K datasets, the proposed method obtained 96.5% and 93.1% accuracy, respectively.