Multitasking via baseline control in recurrent neural networks

Area of Science:

• Computational Neuroscience
• Artificial Intelligence

Background:

• Behavioral states, including arousal and movement, significantly influence neural activity in sensory regions.
• These modulations can be conceptualized as long-range projections that regulate the mean and variance of baseline input currents.

Purpose of the Study:

• To explore the computational advantages of baseline input modulations.
• To investigate how these modulations impact neural network dynamics and function within a brain-inspired reservoir computing framework.

Main Methods:

• Utilized a recurrent neural network with random couplings in a reservoir computing setup.
• Systematically varied quenched baseline inputs to analyze their effect on network dynamics.
• Investigated network phases, including bistable states and phenomena like noise-driven chaos and neural hysteresis.

Main Results:

• Baseline modulations were found to control the dynamical phase of the reservoir network, revealing diverse network phases.
• Identified bistable phases characterized by coexistence of fixed points and chaos, or varying degrees of chaos.
• Observed phenomena such as noise-enhanced chaos, ergodicity breaking, and neural hysteresis.
• Demonstrated that different bistable phases facilitate distinct binary decision-making tasks.
• Showcased fast task switching controlled by adjusting baseline input mean and variance.
• Determined optimal memory performance occurs at first-order phase boundaries.

Conclusions:

• Baseline input control allows for multitasking capabilities in neural networks without altering network couplings.
• Provides a framework for understanding behavioral modulations of cortical activity.
• Suggests new directions for developing brain-inspired artificial intelligence systems.