Brown Researchers' AI Reveals How Creatures Generate Multiple Gaits

Researchers at Brown University's Carney Institute for Brain Science have developed an artificial neural network demonstrating how a four-legged creature may generate multiple distinct gait patterns. This work, published in Neural Computation, offers new insights into how the brain may process complex behaviors and holds potential for advancing quadruped robot technology, enabling more autonomous movement.

Carina Curto, a professor of applied mathematics at Brown, explained that the brain maintains and changes rhythms flexibly. The team's artificial neural network, built on the rules of attractor networks, suggests how biological brains might encode and transition between different patterns and rhythms simultaneously.

Understanding Attractor Networks

Attractor networks are mathematical constructs modeling neural activity patterns. While a type called Hopfield networks models 'static' brain behaviors, such as memory retrieval, this new research expands the framework to efficiently model dynamic behaviors.

Key Findings from the Network

The streamlined network, composed of just 24 artificial neurons, achieved remarkable results:

• It generates five distinct quadruped gaits: bounding, pacing, trotting, walking, and pronking.
• The network captures rapid transitions between these gaits, from a sudden leap to a trot-to-walk shift, without needing to adjust any model parameters.
• These findings suggest that attractor-based networks offer more flexibility and interpretability than other models, providing a unified theoretical framework for studying various brain behaviors.

Juliana Londono Alvarez, a postdoctoral researcher at Brown and the study's lead author, noted that the paper expands attractor networks to include dynamic aspects, showing how principles underlying memory encoding can also generate dynamic processes like gaits.

Potential Applications in Robotics

The researchers suggest the network could serve as inspiration for robotics. Quadruped robots currently borrow behaviors from animals but often rely on expensive, massive, and internet-connected programs. A robot inspired by the Curto lab's small and efficient neural network could operate offline, significantly enhancing its autonomy.

Londono Alvarez is reportedly in discussions with roboticists about adapting the network for their projects.

Research Support & Collaboration

The research team included collaborator Katherine Morrison from the University of Northern Colorado. This work was supported by National Institutes of Health grants R01 EB022862, NSF DMS-1951165, and DMS-1951599. Additional support was provided by National Science Foundation grant DMS-1929284 during a residency at ICERM.