Neuromodulators generate multiple context-relevant behaviors in a recurrent neural network by shifting activity hypertubes

Mood, arousal, and other internal states can drastically alter behavior, even in identical external circumstances — a cold glass of water when you are thirsty is much more desirable than when you are sated. Neuromodulators are critical controllers of such neural states, with dysfunctions linked to various neuropsychiatric disorders. Although biological aspects of neuromodulation have been well studied, the computational principles underlying how large-scale neuromodulation of distributed neural populations shifts brain states remain unclear. We use recurrent neural networks to model how synaptic weight modulation — an important function of neuromodulators — can achieve nuanced alterations in neural computation, even in a highly simplified form. We find that under structural constraints like those in brains, this provides a fundamental mechanism that can increase the computational capability and flexibility of a neural network by enabling overlapping storage of synaptic memories able to generate diverse, even diametrically opposed, behaviors. Our findings help explain how neuromodulators “unlock” specific behaviors by creating task-specific hypertubes in the space of neural activities and motivate more flexible, compact and capable machine learning architectures.