Date of Award

2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

Josh Bongard

Abstract

Generative AI and deep learning technologies have captured societal interest in recent years due to their increasingly adaptive capabilities, but, we have yet to see similarly useful intelligent technologies working alongside humans in the physical world. Despite training of deep learners on massive amounts of data, they remain susceptible to incorrect and unexpected behaviors when presented with out-of-distribution situations. Such a shortcoming is considerably more problematic when these technologies are embedded in physical machines where consequences can be much more severe. Conversely, biological organisms are able to adapt to unforeseen circumstances due to many semi-independent elements (cells, tissues, organs) which self-organize to produce reliable global outcomes. The abstract dynamics, or rules of interaction, between elements that confer this universal plasticity are unknown, making building artificial systems with such a structure a challenging prospect. Because of this, typical approaches involve optimizing a system to robustly perform some task without constraining how its components coordinate to achieve it. However, there is no guarantee that this confers the same adaptability as seen in nature as the particular task may be achieved via sub-optimal component coordination.

Given these issues, this thesis instead explores task-agnostic information-theoretic modes of interaction in artificial and biological substrates while they confront homeostatic challenges. A series of experiments are reported in which computational models of cell collectives are optimized to grow into and maintain a pre-specified shape. Addition of auxiliary objectives to the optimization process that select for particular information dynamics between cells were found to facilitate optimization of homeostatic behaviors. Further, I found that similar information flows are present in a unique biological substrate - an organoid system derived from embryonic frog epithelial cells - that grapples with a homeostatic challenge. Taken together, these findings suggest there may exist many other kinds of information flows that promote plasticity in many-agent collectives, and that such flows may in the future be automatically discovered and distilled from biological systems, leading to ever more adaptive and useful multi-agent machines.

Language

en

Number of Pages

124 p.

Download

Share

COinS