Fluctuation-driven attractors

Attractor dynamics emergent in network models have tremendously shaped how we think about brain function. Fixed-point attractors, for instance, are standard biological models for pattern completion, working memory and decision making. The real brain operates in a fluctuation-driven regime, with neurons bombarded by highly variable and noisy synaptic inputs, leading to Poisson-like spiking. In biological models, such fluctuations and irregular spiking usually interfere with function and destabilize attractors. Here we challenge this idea with a model and mean-field theory for a random network in which input fluctuations support, rather than degrade attractors. In our network single units are not neurons but biologically inspired small groups of neurons of distinct cell types that compete via winner-take-all, with tunable cross-unit synaptic weight means and variances for each pair of cell types. Unlike past models, this network produces robust macroscopic attractor dynamics—including multiple fixed points, slow limit cycles, and many flexible nonlinear behaviors—all powered by chaotic microscopic fluctuations sustained via Poisson-like spiking. This shows how the brain's curious, noisy operating regime could be indispensable to function and suggests the rich computational potential of fluctuation-driven dynamical systems.