Criticality as a mechanism of structural coupling: understanding cortical dynamics through the cybernetic lens

The framework of cybernetics views the action-perception cycle as a structural coupling between the information extracted from the environment by organisms and the actions that they then perform on said environment. Through this lens, the brain is viewed as an organ that expands the internal patterns of structural coupling within an organism. This provides an evolutionary advantage, since it diversifies behavior, allowing for improved adaptability to unfamiliar situations and the expanded ability of generalize previously learned associations between perception and action to new contexts. Therefore, it follows that a brain that can generate a broader gamut of activity patterns would thus optimize the behavior of the organism across different situations and tasks. That said, it has been proposed that one of the advantages of criticality in cortical networks is precisely the expansion of its activity pattern repertoire.

In this project, we will first compile research about criticality in brain networks and reinterpret it through the lens of cybernetics, particularly regarding its role in maximizing information capacity in cortical networks, which is directly related to the entropy of activity patterns such networks can generate. Then, we will apply this framework to simulated agents by using spiking networks inspired in the neocortex to couple their sensors and actuators and have them perform reward-based foraging tasks in an environment. Furthermore, we will explore the behavior of agents with different levels of network criticality, which can be tuned by varying the excitatory-inhibitory balance of their internal spiking networks. In this case, we aim to test the ability of these agents to explore new environments and perform more complex collaborative tasks (e.g. collective foraging).

Finally, we will also explore how the rich spontaneous dynamics of near-critical systems can be leveraged to consolidate information learned during behavior, by exploring a paradigm where agents intertwine periods of action with "rest" and evaluating how network dynamics during these resting periods can optimize learning, in general, and the ability to generalize previously learned behaviors to new situations, in particular.