Baseline control of optimal performance in recurrent neural networks

Abstract

Changes in behavioral state, such as arousal and movements, strongly affect neural activity in sensory areas. Recent evidence suggests that they may be mediated by top-down projections regulating the statistics of baseline input currents to sensory areas, inducing qualitatively different effects across sensory modalities. What are the computational benefits of these baseline modulations? We investigate this question within a brain-inspired framework for reservoir computing, where we vary the quenched baseline inputs to a random neural network. We found that baseline modulations control the dynamical phase of the reservoir network, unlocking a vast repertoire of network phases. We uncover a new zoo of bistable phases exhibiting the simultaneous coexistence of fixed points and chaos, of two fixed points, and of weak and strong chaos. Crucially, we discovered a host of novel phenomena, including noise-driven enhancement of chaos and ergodicity breaking; neural hysteresis, whereby transitions across phase boundary retain the memory of the initial phase. Strikingly, we found that baseline control can achieve optimal performance without any fine tuning of recurrent couplings. In summary, baseline control of network dynamics opens new directions for brain-inspired artificial intelligence and provides a new interpretation for the ubiquitously observed behavioral modulations of cortical activity.