Forget-free subnetworks for life-long learning

Abstract

SoftNet jointly learns the regularized model weight and task-adaptive non-binary masks. The outstanding points of WSN and SoftNet were derived through the following three continuous learning scenarios. In the first scenario, Task Incremental Learning (TIL), WSN, and SoftNet are inherently immune to catastrophic forgetting as each selected subnetwork model does not infringe upon other subnetworks. Moreover, we observed that sub-networks compression in TIL minimizes the model capacity. Surprisingly, in the inference step, SoftNet generated by injecting small noises to the backgrounds of acquired WSN (holding the foregrounds of WSN) provides excellent forward transfer power for future tasks in TIL. In the second scenario, Few-shot Class Incremental Learning (FSCIL), SoftNet shows its effectiveness over WSN inregularizing parameters to tackle overfitting issues caused by a few examples. In the third scenario, Video Incremental Learning (VIL), WSN suggested that the reused weights of the subnetworks depend on the video contexts and that its video generation ability is close to the upper bounds deduced by Multi-Task Learning (MTL).

Inspired by Regularized Lottery Ticket Hypothesis (RLTH), which states that competitive smooth (non-binary) subnetworks exist within a dense network in continual learning tasks, we investigate two proposed architecture-based continual learning methods which sequentially learn and select adaptive binary-(WSN) and non-binary Soft-Subnetworks (SoftNet) for each task. WSN jointly learns the model weights and task-adaptive binary masks of subnetworks associated with each task whilst attempting to select a smallset of weights to be activated (winning ticket) by reusing weights of the prior subnetworks