Neural competition and probabilistic representations

Perception and action are closely tied to the statistical properties of physical stimuli and likely rewards. The precise mechanisms, however, remain elusive. A straightforward, biologically realistic network model reveals that probabilistic behavior can emerge naturally from the sampling of competing responses across various situations. Remarkably, this model also provides a computational framework to explain the common observation of balanced excitation and inhibition in the brain.

In this network, recurrent connections among excitatory neurons form multiple attractor states. These states are coupled to an inhibitory neuron pool, which enforces exclusivity among the attractors. When the same stimuli are presented simultaneously, the competing attractors engage in an alternating pattern of activity. This global transition between states is driven by local, uncorrelated spiking noise, yet over time it reflects the relative strengths of the different attractors, providing a neural basis for choice behavior in the face of uncertainty.

Simple probabilistic competitive recurrent networks (PCRN) permit closed-form analysis, offering insights into the neural underpinnings of decision-making under uncertainty. More intricate systems consisting of laterally connected PCRN can collectively address the numerous local ambiguities inherent in sensory stimuli. Through rapid alternations between attractors, these systems settle into globally-consistent configurations that align with perceptual reports.

Crucially, alternations in activity are essential, and they only occur if the inhibitory pool within the PCRN is robust enough to prevent any single attractor from reaching saturation. Thus, a delicate balance between excitation and inhibition emerges as both a necessary condition and a defining characteristic of probabilistic sampling within cortical networks.