CCN talk: November 1, 2024
Tamas Spisak
Professor of Predictive Neuroscience, Center for Translational Neuro- and Behavioral Sciences, University Medicine Essen, University Duisburg-Essen, Germany
Information integration and causal inference in the brain via large-scale attractor dynamics
Abstract: Information processing in the brain is both distributed across multiple regions and hierarchical in nature. Higher-order cognitive functions, emotions, and multifaceted experiences such as anxiety, fear, pain, or self-awareness are believed to be governed by higher-order “integrator regions” at the top of this hierarchy, which have access to all information relevant to these complex processes. However, these exactly these functions often resist empirical mapping to any single brain region. Here, we introduce a theoretical framework in which large-scale integrative information processing at the apex of the computational hierarchy is achieved through the attractor dynamics of the entire system, rather than through dedicated integrator regions. We present a prototypical mathematical formalization of this concept, which involves marginalizing a hierarchy of free-energy-minimizing integrator nodes into an equivalent recurrent network composed solely of lower-level units. In this network, higher-level units manifest as attractor states. Our framework elucidates the computational mechanisms underlying the active inference processes that drive attractor dynamics and establishes connections to Bayesian causal inference, predictive coding, and computational emergence within the brain. We propose and demonstrate practical methods for reconstructing these dynamics from empirical data, discuss existing evidence that supports this theory and explore its implications across various fields, including the formulation of testable hypotheses and predictions.