Continual Learning From Unlabeled Data Via Deep Clustering

Continual learning, a promising future learning strategy, aims to learn new tasks incrementally using less computation and memory resources instead of retraining the model from scratch whenever new task arrives. However, existing approaches are designed in supervised fashion assuming all data from new tasks have been annotated, which are not practical for many real-life applications. In this work, we introduce a new framework to make continual learning feasible in unsupervised mode by using pseudo label obtained from cluster assignments to update model. We focus on image classification task under class-incremental setting and assume no class label is provided for training in each incremental learning step. For illustration purpose, we apply k-means clustering, knowledge distillation loss and exemplar set as our baseline solution, which achieves competitive results even compared with supervised approaches on both challenging CIFAR-100 and ImageNet (ILSVRC) datasets. We also demonstrate that the performance of our baseline solution can be further improved by incorporating recently developed supervised continual learning techniques, showing great potential for our framework to minimize the gap between supervised and unsupervised continual learning.