Continual Learning for Unsupervised Concept Bottleneck Discovery

In the context of continual learning, little attention is dedicated to the problem of developing a layer of “concepts”, also known as “concept bottleneck”, to support the discrimination of higher-level task information, especially when concepts are not supervised. Concept bottleneck discovery in an unsupervised setting is thus largely unexplored, and this paper aims to move a step forward in such direction. We consider a neural network that faces a stream of binary tasks, with no further infor- mation on the relationships among them, i.e., no supervisions at the level of concepts. The learning of the concept bottleneck layer is driven by means of a triplet-based criterion, which is instantiated in conjunction with a specifically designed experience replay (concept replay). Such a novel crite- rion exploits fuzzy Hamming distances to treat vectors of concept probabilities as fuzzy bitstrings, encouraging different concept activations across different tasks, while also adding a regularization effect which pushes probabilities towards crisp values. Despite the lack of concept supervisions, we found that continually learning the streamed tasks in a progressive manner yields the develop- ment of inner concepts that are significantly better correlated with the higher-level tasks, compared to the case of joint-offline learning. This result is showcased in an extended experimental activity involving different architectures and newly created (and shared) datasets that are also well-suited to support further investigation of continual learning in concept-based models