LINDA: Learning Integrative Neural Dynamics with Joint Embedding Predictive Architecture

Neuroscience could benefit from foundation models that integrate fragmented neural datasets into a more coherent picture of brain function. Building such models is difficult because neural datasets are often heterogeneous and collected under specialized experimental conditions; recordings across sessions, animals, tasks, and modalities are often asynchronous. To jointly train on these data, we propose to leverage the inductive bias that diverse neural recordings can be understood as partial observations of a shared latent dynamical system---the brain---whose dynamics can depend on context. Building on Joint Embedding Predictive Architecture (JEPA), we introduce LINDA, a self-supervised latent prediction framework for learning context-conditioned dynamics. LINDA summarizes latent history into a causal context vector and uses a hypernetwork to generate an explicit local linear transition operator at each step, allowing dynamics to change within a trial. Experiments on synthetic data where ground truth latents are available show that LINDA can identify the latent dynamics underlying different observations of the same system. For behavioral decoding from real monkey reaching data, joint predictive and supervised learning improves limited-data transfer across subjects compared with purely supervised learning, supporting shared dynamical structure across animals. These results suggest that context-dependent latent prediction is a promising way to leverage heterogeneous neural datasets and an important step toward building neural foundation models.