Project Details
Projekt
Charting diversity in computational system dynamics and behavioral strategies in cognitive flexibility
Applicant
Professor Dr. Daniel Durstewitz
Subject Area
Experimental and Theoretical Network Neuroscience
Term
since 2026
Project identifier
Deutsche Forschungsgemeinschaft (DFG) - Project number 437610067
Different behavioral strategies and computational algorithms may be used to solve the same cognitive task, and previous research has shown that employed strategies indeed vary both within and across animals. In theoretical neuroscience, computational processes in the nervous system, and associated behavioral strategies, are often conceptualized in terms of dynamical systems theory, which provides a mathematical framework for linking the physiological and computational levels. Recent years have seen tremendous progress in machine learning and AI tools for recovering dynamical surrogate models directly from experimental data, called dynamical systems reconstruction (DSR). Based on the results and methodological developments from the first funding period, this TP will use DSR methods and computational theory to study the diversity in behavioral and computational strategies animals use to accomplish different cognitive flexibility tasks. Specifically, we will train generative DSR models based on interpretable recurrent neural networks (RNNs), embedded within a hierarchical multimodal foundation model architecture, on diverse physiological and behavioral datasets acquired within this consortium. This pre-trained and evaluated DSR foundation model architecture will serve as a basis for several of the theoretical projects in this consortium. Besides subject-level models of the computational system dynamics, this hierarchical approach also delivers a feature space in which subjects or sessions/ trials with similar neuronal dynamics cluster. Based on this, we will investigate how variation in detected behavioral strategies both within and across animals is reflected in the underlying neural dynamics implementing those strategies. We will furthermore develop a computational theory based on topological concepts and symbolic dynamics to characterize the diversity in computational algorithms implemented by data-trained RNNs. At the same time, RNNs trained on the same behavioral tasks as used for the animal subjects will be used to obtain a baseline for the variation in computational and dynamical strategies that could be expected given the constraints imposed by the task structure. In our final work package, we will then link the diversity in computational algorithms to that in behavioral strategies and to the underlying neural dynamics. Thus, the present project will deliver a formal characterization of the diversity in behavioral strategies animals use to solve different flexibility tasks engaged within this consortium in neuro-dynamical and computational terms, thus contributing to the core aims of this cluster of uncovering the sources of behavioral variability.
DFG Programme
Research Units
