Soft Condensed Matter & Physics of Living Systems Seminar

"Cognition in Tiny, Random Spaces: What You Get for Free When There is No Free Energy"

Abstract

A variety of naturally evolved systems—from highly integrated neural networks inside the animal brain to loosely integrated collective behaviors in an ant colony—show hallmarks of cognition; these systems appear to make decisions based on information accumulated from the surrounding environment. Prior investigations have shown how cognitive processes can manipulate the surrounding environment while making a decision, but little focus has been placed on understanding the latent information-processing ability in the environment itself. The earliest information-processing systems emerged out of entities interacting within physical spaces, and so understanding the evolution of information processing requires understanding the exploitable information-processing opportunities afforded by background structures and constraints. In this talk, I describe our recent findings that decision-making performance curves taken from real ants that were once thought to be a product of a deliberative, cognitive process within the ant brain can also be explained as primarily a property of the cavities that constrain the motion of stochastically interacting particles. This idea that constrained randomness can be a module for cognition can be extended to very different contexts, such as the design of deep neural networks. I demonstrate this by showing that deep networks designed for sophisticated knowledge representation and reasoning tasks can have increased performance by counterintuitively replacing training with random weighting, thereby showing that a network's representational strengths can be more a property of macroscopic structure of a dense network and not any particular "optimal" pattern of network weights. As these examples demonstrate, a true science of cognitive ecology is necessarily a physics of living systems as it requires marrying the non-living physical world with the out-of-equilibrium behaviors of those agents both constrained by and, in turn, enabled by it.

Bio

After a decade of working in Software and Systems Engineering, Theodore Pavlic received his Ph.D. in 2010 in Electrical and Computer Engineering at The Ohio State University and has progressed through postdoctoral appointments in both Computer Science and Behavioral Ecology. He currently is an Assistant Professor at Arizona State University jointly appointed in the School of Computing, Informatics, and Decision Systems Engineering and the School of Sustainability, with an adjunct appointment in the School of Life Sciences. He also serves as the Associate Director for Research at The Biomimicry Center and is affiliated with a number of ASU centers related to complexity and trans-disciplinary thinking. His interdisciplinary laboratory includes students from a range of programs stretching across Computer Science, Industrial Engineering, Animal Behavior, Biology, and Applied Math for the Life and Social Sciences. Consequently, work in the lab spans empirical work with natural systems, such as social-insect colonies in both the field and lab, artificial intelligence and machine learning work related to computational sustainability, and multi-robot systems work. Furthermore, his lab participates in conferences, publishing venues, and professional organizations across several more traditional disciplines in biological sciences and engineering. The common thread that goes through all of his lab's work is a better understanding of how autonomous systems can make good decisions especially amongst a background of long-term autonomy in changing environments. 

Event Details

Date/Time:

  • Date: 
    Tuesday, March 10, 2020 - 3:00pm to 4:00pm

Location:
Howey School of Physics N202

For More Information Contact

Prof. Dan Goldman