Neural Circuit Architectural Priors for Embodied Control

Nikhil Xie Bhattasali, Anthony M. Zador, Tatiana A Engel

Keywords: neuroscience-inspired AI, robotics, motor control

Abstract: Artificial neural networks coupled with learning-based methods have enabled robots to tackle increasingly complex tasks, but often at the expense of requiring large amounts of learning experience. In nature, animals are born with highly structured connectivity in their brains and nervous systems that enables them to efficiently learn robust motor skills. Capturing some of this structure in artificial models may bring robots closer to matching animal performance and efficiency. In this paper, we present Neural Circuit Architectural Priors (NCAP), a set of reusable architectural components and design principles for deriving network architectures for embodied control from biological neural circuits. We apply this method to control a simulated agent performing a locomotion task and show that the NCAP architecture achieves comparable asymptotic performance with fully connected MLP architectures while dramatically improving data efficiency and requiring far fewer parameters. We further show through an ablation analysis that principled excitation/inhibition and initialization play significant roles in our NCAP architecture. Overall, our work suggests a way of advancing artificial intelligence and robotics research inspired by systems neuroscience.

One-sentence Summary: NCAP, a set of reusable architectural components and design principles for deriving network architectures for embodied control from biological neural circuits