Lecturer:  Dr. Itay Griniasty
 Abstract:
Systems composed of many interacting elements that collaboratively generate a function, such
as meta-material robots, proteins, and neural networks are notoriously difficult to design.
Such systems elude traditional explicit design methodologies, which rely on composing
individual components with specific subfunctions, such as cogs, springs and shafts, to achieve
complex functionality. In part the problem stems from the fact that there are few principled
approaches to the design of emergent functionality. In this talk I will describe progress towards
creating such paradigms for two canonical systems: I will first describe how bifurcations of the
system dynamics can be used as an organizing principle for the design of functionality in protein
like machines with magnetic interactions. I will then introduce a computational microscope that
we have developed to analyze emergent functionality, and its application to machine learning.
There we uncovered compelling evidence that the training of neural networks is inherently low
dimensional, suggesting new paradigms for their design.
References
1. T. Yang et al. Bifurcation instructed design of multistate machines. Proceedings of the
National Academy of Sciences, 120(34):e2300081120, 2023
2. J. Mao et al. The training process of many deep networks explores the same low-dimensional
manifold. arXiv preprint arXiv:2305.01604, 2023.
3. R. Ramesh, et al. A picture of the space of typical learnable tasks. Proc. of International
Conference of Machine Learning (ICML), 2023.