Continual Learning with Node-Importance based Adaptive Group Sparse Regularization

We propose a novel regularization-based continual learning method, dubbed as Group Sparsity based Continual Learning (AGS-CL), using two group-based penalties. Our method selectively employs the two penalties when each node based its the importance, which is adaptively updated after each new task. By utilizing the proximal gradient descent method for, the exact sparsity and freezing of the model is guaranteed, and thus, learner can explicitly control the model capacity as the learning. Furthermore, as a critical detail, we re-initialize the weights with unimportant nodes after learning each task in order to prevent negative transfer that causes the catastrophic forgetting and facilitate learning of new tasks. Throughout the extensive experimental results, show that our AGS-CL uses much less additional memory space for storing the parameters, and it significantly outperforms several-of-the-art baselines on representative continual learning benchmarks for supervised and reinforcement learning tasks.