Seq-JEPA: Autoregressive Predictive Learning of Invariant-Equivariant World Models

Joint-embedding predictive architecture (JEPA) is a self-supervised learning (SSL) paradigm with the capacity of world modelling via action-conditioned prediction. Previously, JEPA world models have been shown to learn action-invariant or action-equivariant representations by predicting one view of an image from another. Unlike JEPA and similar SSL paradigms, animals, including humans, learn to recognize new objects through a sequence of active interactions. To introduce \emph{sequential} interactions, we propose \textit{seq-JEPA}, a novel SSL world model equipped with an autoregressive memory module. Seq-JEPA aggregates a sequence of action-conditioned observations to produce a global representation of them. This global representation, conditioned on the next action, is used to predict the latent representation of the next observation. We empirically show the advantages of this sequence of action-conditioned observations. We examine our sequential modeling paradigm in two settings: (1) \emph{predictive learning across saccades}; a method inspired by the role of eye movements in embodied vision. This approach learns self-supervised image representations by processing a sequence of low-resolution visual patches sampled from image saliency, without any hand-crafted data augmentations. (2) \emph{invariance-equivariance trade-off}; seq-JEPA's architecture results in automatic separation of invariant and equivariant representations, with the aggregated autoregressor outputs being mostly action-invariant and the encoder output being equivariant. This is in contrast with many equivariant SSL methods that expect a single representational space to contain both invariant and equivariant features, potentially creating a trade-off between the two. Empirically, seq-JEPA achieves competitive performance on both invariance and equivariance-related benchmarks compared to existing methods. Importantly, both invariance and equivariance-related downstream performances increase as the number of available observations increases.