Erik Winfree⏐CalTech

Multicomponent molecular systems and neural computation
Date
Mar 3, 2025, 12:30 pm1:30 pm
Location

Details

Event Description

The concepts of energy and energy landscapes have been central to the understanding of many types of neural network computation.  As Hopfield illustrated in the 1980's, generalizing the homogenous Ising model of magnetism to allow individually tunable symmetric point-to-point interactions yields a beautiful mathematical metaphor for associative memory in model neural networks. In contrast to modern machine learning where the "energy" formulation enables the application of powerful mathematical techniques but is not physical, in this talk I will discuss several classes of molecular systems where the physical energy landscape has forms and consequences analogous to those of neural networks, leading to similar "intelligent" behavior in non-living materials. Specifically, we will consider multicomponent chemical reaction networks, multicomponent liquid phase separation, and multicomponent crystal self-assembly both in theory and with respect to their incarnation using DNA nanotechnology. Energy minimization and equilibration in these systems can correspond to associative recall and probabilistic inference, while nucleation can perform pattern recognition.