Relationship between behavioral output and internal dynamics in biological and recurrent neural networks

Abstract

An organism controls its state in the environment through behavior; behavior, in turn, influences neuronal dynamics through feedback to the brain. Despite this apparent connection of behavior and neuronal activity, the exact nature of the relationship between neuronal population dynamics and behavioral dynamics remains unresolved. Establishing such a relationship is inherently difficult with traditional neuroscience approaches as the diverse ranges of behaviors and neuronal activation patterns emerging due to non-linear interactions, recurrences, and feedback loops cannot be explained from the function of individual network components. In contrast, treating the network and its inputs as a single integrated complex system can reveal relationships between emergent phenomena at different organizational scales, such as between neuronal population dynamics and behavioral dynamics. However, complex network studies often focus on either internal processes or behavioral outcomes and few explore the (statistical) relationship between them. This thesis approaches the relationship between internal neuronal dynamics and behavioral output dynamics from a complex systems perspective and explores the possible connection of this relationship to the critical brain hypothesis, which proposes that the brain is tuned to a critical state which optimizes its information processing capabilities. The analysis of neuronal dynamics, behavioral output, and changes in network connectivity in two example systems—a biological and an artificial recurrent neural network—revealed that changes in functional and structural connectivity at the level of individual connections may or may not occur alongside changes in behavioral dynamics, while signatures of criticality in the internal dynamics remain robust to such changes. Specifically, neuronal avalanche size distributions in both the retrosplenial cortex of mice and in the recurrent neural network were robust to changes in connectivity; in the artificial network, this robustness occurred despite alterations in behavioral dynamics, whereas in the mice, behavioral statistics remained invariant to connectivity changes. The findings suggest that the relationship between neural activity and behavior is not one-to-one or trivial in complex networks. Uncovering the mechanisms underlying these findings remains a challenge for the future.