Life-Like Self-Assembly through Dissipative Adaptation
Abstract
Since the last report submission, three manuscripts have been completed and submitted for publication. 1) A simulated model of a disordered mechanical network was studied. By subjecting this system to various kinds of external powered mechanical oscillations, it was shown that this model exhibits spontaneous adaptation behavior that is specific to its input driving signal and reduces work absorption from this signal across a range of different types of drive. Moreover, the system shows a lifelike ability to find attractor states that are robust against being rearranged by energy from the source that powered the attractor's emergence, much the same way that living things digest food from specific sources but are damaged by energy introduced through unfamiliar mechanisms. This was submitted to PRX. 2) A simulated model of a random spin glass in a complexly patterned sequence of 'barcoded' external fields exhibited a spontaneous novelty detection behavior, in that its energy absorption rate tended to spike when new force fluctuations never seen previously were introduced after a period of adaptation to fixed set of forcings. This finding indicated that a disordered many-body system can learn to predict the likely future of its environment simply by experiencing force fluctuations produced by that environment, and without any deliberate optimization or training or parameters. This was submitted to Nature Communications 3) The structure of this novelty-detecting spin glass system was studied using variational autoencoders and other bottlenecked neural nets. It was found that there is a regime where energy absorption does not act as a clear signature for learning, but where a neural net can detect low-dimensionality in the spin glass that indicates it has learned the pattern of its environment. This was submitted to PRX.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 28, 2020
- Accession Number
- AD1106227
Entities
People
- Jean-jacques Slotine
Organizations
- Massachusetts Institute of Technology