Informing deep neural networks by multiscale principles of neuromodulatory systems. Issue 3 (March 2022)

• Record Type: Journal Article
• Authors: Mei, Jie; Muller, Eilif; Ramaswamy, Srikanth
• Is Part Of: Trends in neurosciences. Volume 45:Issue 3(2022)
• Page Start: 237
• Page End: 250
• Publication Date: 2022-03
• Subjects: acetylcholine -- noradrenaline -- serotonin -- dopamine -- multiscale organization -- adaptive learning; Neurology -- Periodicals; Neurophysiology -- Periodicals; Neurobiology -- Periodicals; 612.8
• DOI: 10.1016/j.tins.2021.12.008
• Languages: English
• ISSNs: 0166-2236

Abstract

Abstract : Our brains have evolved the ability to configure and adapt their processing states to match the unique challenges of acting and learning in diverse environments and behavioral contexts. In biological nervous systems, such state specification and adaptation arise in part from neuromodulators, including acetylcholine, noradrenaline, serotonin, and dopamine, whose diffuse release fine-tunes neuronal and synaptic dynamics and plasticity to complement the behavioral context in real-time. Despite the demonstrated effectiveness of deep neural networks for specific tasks, they remain relatively inflexible at generalizing across tasks or adapting to ever-changing behavioral demands. In this article, we provide an overview of neuromodulatory systems and their relationship to emerging pertinent principles in deep neural networks. We further outline opportunities for the integration of neuromodulatory principles into deep neural networks, towards endowing artificial intelligence with a key ingredient underlying the flexibility and learning capability of biological systems. Highlights: Neuromodulatory systems in biological brains play an important role in learning and adaptive behavior. Different neuromodulators operate on a spectrum of spatio-temporal scales in tandem and opposition to reconfigure functions of local neural networks and to regulate global cognitive states. Although neuromodulators are important for general cognition, their phenomenology is yet to be fully realized in deep neural networks (DNNs) and has so far been mostly limited to implementations of reinforcement learning models. We highlight the organizing principles of neuromodulators in adaptive cognition, explore their diverse functional settings to inspire new approaches of DNNs, and offer suggestions on how these approaches could be evaluated.