The role of Disentanglement in Generalisation

Combinatorial generalization — the ability to understand and produce novel combinations of familiar elements — is considered a core capacity of humans that modern AI systems currently struggle at. Recently, it has been suggested that learning disentangled representations may help address this problem. It is claimed that such representations should be able to capture the compositional structure of the world which can then be combined to produce novel representations. That is, compositional representations are thought to support combinatorial generalisation. In this study, we systematically tested how the degree of disentanglement affects various forms of generalisation, including two forms of combinatorial generalisation that varied in difficulty. We trained variational autoencoders (VAEs) with different levels of disentanglement on an unsupervised task by excluding combinations of generative factors during training. At test time we ask the models to reconstruct the missing combinations in order to measure generalisation performance. Irrespective of the degree of disentanglement, we found that the models supported only weak combinatorial generalisation. Next, we tested our model in a more complex task which explicitly required independent generative factors to be controlled. Although this added pressure improved the level of disentanglement compared to normal VAEs, generalisation did not improve. While learning disentangled representations does improve interpretability and sample efficiency in downstream tasks, our results suggest that they are not sufficient for supporting more difficult forms of generalisation.