Inspired by the Well-initialized Lottery Ticket Hypothesis (WLTH), which provides suboptimal fine-tuning solutions, we propose a novel fully fine-tuned continual learning (CL) method referred to as Soft-TransFormers (Soft-TF). Soft-TF sequentially learns and selects an optimal soft-network for each task. During sequential training in CL, a well-initialized Soft-TF mask optimizes the weights of sparse layers to obtain task-adaptive soft (real-valued) networks, while keeping the well-pre-trained layer parameters frozen. In inference, the identified task-adaptive network of Soft-TF masks the parameters of the pre-trained network, mapping to an optimal solution for each task and minimizing Catastrophic Forgetting (CF) - the soft-masking preserves the knowledge of the pre-trained network. Extensive experiments on the Vision Transformer (ViT) and the Language Transformer (Bert) demonstrate the effectiveness of Soft-TF, achieving state-of-the-art performance across Vision and Language Class Incremental Learning (CIL) scenarios.
View on arXiv