Real-time representations of neural dynamics: towards artificial and hybrid models

Bradly Alicea, M.I.N.D. Lab, Michigan State University

Numerous studies (Bassett et al, 2006; Freeman, 1999; Eguiluz et al, 2005; Salvador et al, 2005) have used EEG, MEG, and fMRI to uncover small-world networks and nonlinear dynamics in the functional anatomy of brains engaged in an experimental activity. How this relates to operational contexts is somewhat unclear. In this submission, I will lay out a model of real-time functional neuroanatomy for artificial and hybrid systems inspired by recent neuroimaging findings and neural systems.

In humans and animals, attentional mechanisms are the central feature regulating the input and integration of sensory features gathered from the environment (Posner et al, 1987). Evidence from humans and animals has suggested that attentional mechanisms have a finite capacity; the limits of this capacity are regulated by the number of sources in the environment and the automated nature of the process. I will argue that attentional networks existing anatomically in the brain function in a way analogous to the Bak-Sneppen sandpile model. In a dynamical context, activity within this network becomes patterned in ways that allow for phase transitions between controlled, automatic, and disrupted automatic states. This relates to traditional models of capacity that are robust to finite amounts of distraction and levels of focus.

Complex network mechanisms may also mitigate these dynamics. Connectivity within and between discrete brain regions can be selective and determine how processes are controlled and break down over time. In addition, specific portions of the network, such as the prefrontal areas in Primates and the superior colliculus in several mammal species, might act to guide the transition from state to state. In complex networks parlance, these areas may act as hubs of activity. When these hubs change their activity, they affect the entire network in ways that create a large-scale change in activity.

This type of model can be applied to both augmented cognition models and robotic systems. In one sense, the model can be used to predict real-time shifts in behavior due to variable feedback from environmental features. In the case of robotic systems, this type of model can be used to produce spontaneous but conditional shifts in attention that are not be possible using traditional artificial intelligence methods.