Inferring time-varying causal interactions in behavior and neural activity

Understanding causality in complex dynamical systems like animal behavior and neural activity is challenging due to inherent nonlinearity and time-varying interactions. In this talk, we introduce Temporal Autoencoders for Causal Inference (TACI), a novel framework that quantifies time-varying causal strengths and directions using a two-headed neural network architecture combined with the Comparative Surrogate Granger Index (CSGI). The TACI framework captures causal influences across entire time series by encoding interactions between variables in a shared latent space, allowing it to detect shifts in causal structure without the need for retraining. We demonstrate TACI’s robustness in identifying causal interactions across synthetic models with known coupling dynamics and real-world data, including atmospheric datasets, electrocorticography recordings from non-human primates, and rat social behavior. Our results indicate that TACI outperforms conventional methods, providing a nuanced understanding of causality in non-stationary, nonlinear systems and showcasing its potential as a tool for analyzing complex physical and biological interactions.