An Efficient Strategy for Catastrophic Forgetting Reduction in Incremental Learning

Abstract

Deep neural networks (DNNs) have made outstanding achievements in a wide variety of domains. For deep learning tasks, large enough datasets are required for training efficient DNN models. However, big datasets are not always available, and they are costly to build. Therefore, balanced solutions for DNN model efficiency and training data size have caught the attention of researchers recently. Transfer learning techniques are the most common for this. In transfer learning, a DNN model is pre-trained on a large enough dataset and then applied to a new task with modest data. This fine-tuning process yields another challenge, named catastrophic forgetting. However, it can be reduced using a reasonable strategy for data argumentation in incremental learning. In this paper, we propose an efficient solution for the random selection of samples from the old task to be incrementally stored for learning a sequence of new tasks. In addition, a loss combination strategy is also proposed for optimizing incremental learning. The proposed solutions are evaluated on standard datasets with two scenarios of incremental fine-tuning: (1) New Class (NC) dataset; (2) New Class and new Instance (NCI) dataset. The experimental results show that our proposed solution achieves outstanding results compared with other SOTA rehearsal methods, as well as traditional fine-tuning solutions, ranging from 1% to 16% in recognition accuracy.

CL3: Generalization of Contrastive Loss for Lifelong Learning

“…To mitigate catastrophic forgetting, the use of a memory buffer is studied [6,7] in the literature. This family of L3 methods stores a subset of samples from previous tasks in a memory buffer and replays interleaved with new samples.…”