Baseline control of optimal performance in recurrent neural networks*

Changes in an animal's behavioral state, such as arousal and movements, induce opposite effects on task performance depending on the sensory modality (visual vs. auditory). Experimental studies showed that the changes in behavioral states are mediated by modulation of the baseline inputs current to populations of neurons in sensory areas. Here, we investigate the benefits of these modulations using a reservoir computing approach, modeling a sensory area as a recurrent neural network, and changes in brain state as modulations of its baseline inputs. In our brain-inspired framework for reservoir computing, we found that the dynamical phase of a recurrent neural network is controlled by modulating the mean and quenched variance of its baseline inputs. Baseline modulation unlocks a zoo of new phenomena. First, we found that baseline modulation drive a novel noise-induced enhancement of chaos. Second, baseline modulations unlocked access to a large repertoire of network phases. On top of the known fixed point and chaotic ones, we uncovered several new bistable phases, where the network activity breaks ergodicity and exhibits the simultaneous coexistence of a fixed point and chaos, of two different fixed points, or of weak and strong chaos. The bistable phases give rise to ergodicity breaking. By driving the network with adiabatic changes in the baseline statistics one can toggle between the different phases, charting a trajectory in phase space. These trajectories exhibited the new phenomenon of neural hysteresis, whereby adiabatic transitions across a phase boundary retain the memory of the adiabatic trajectory. Finally, we showed that baseline control can achieve optimal performance in a sequential memory task at a second-order phase boundary without any fine tuning of the network recurrent couplings. Our results show that 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, enabling behavioral flexibility.