Emergence and Function of Abstract Representations in Self-Supervised Transformers

Human intelligence relies in part on our brains' ability to create abstract mental models that succinctly capture the hidden blueprint of our reality. Such abstract world models notably allow us to rapidly navigate novel situations by generalizing prior knowledge, a trait deep learning systems have historically struggled to replicate. However, the recent shift from supervised to self-supervised objectives, combined with expressive transformer-based architectures, have yielded powerful foundation models that appear to learn versatile representations that can support a wide range of downstream tasks. This promising development raises the intriguing possibility of such models developing in silico abstract world models. We test this hypothesis by studying the inner workings of small-scale transformers trained to reconstruct partially masked visual scenes generated from a simple blueprint. We show that the network develops intermediate abstract representations, or abstractions, that encode all semantic features of the dataset. These abstractions manifest as low-dimensional manifolds where the embeddings of semantically related tokens transiently converge, thus allowing for the generalization of downstream computations. Using precise manipulation experiments, we demonstrate that abstractions are central to the network's decision-making process. Our research also suggests that these abstractions are compositionally structured, exhibiting features like contextual independence and part-whole relationships that mirror the compositional nature of the dataset. Finally, we introduce a Language-Enhanced Architecture (LEA) designed to encourage the network to articulate its computations. We find that LEA develops an abstraction-centric language that can be easily interpreted, allowing us to more readily access and steer the network's decision-making process.