Tel Aviv University International Prize in Biophysics Award Ceremony and Mini-Symposium: Embodied learning and computation in physical systems

Abstract:

Our current physical models of high-dimensional computation and learning typically seek to mimic neural network architectures element by element. Forcing physical systems to mimic computational abstractions such as neural networks carries significant overhead. I'll describe an alternative: exploiting the collective dynamics already present in physical systems to implement both learning and computation more compactly. For computation, we show that dense reversible interaction networks of molecules can serve as compact yet expressive substrates. Competition, cooperativity, and irreversibility emerge as key physical principles that set the expressive power of such systems. For learning, we show that high-pass filtering in plastic processes naturally realizes a Boltzmann-like learning rule. This single mechanism can train a physical system for diverse tasks: Pavlovian conditioning, supervised classification, and generative tuning of bet-hedging ratios. Crucially, this rule is model-free and can compensate for unknown or unmodeled interactions present in living cells. Together, these results suggest design principles for trainable cellular circuits and reveal how molecular systems might learn statistical features of their environments through experience alone.

Additional Talks:

• Michael Faran, TAU, Coarse-Graining by the Stochastic Landscape Method

• Amir Ohad, TAU, Plant Whisking: Mechanical Sensing in Climbing Plants

• Chaviva Sirote-Katz, TAU, Breaking Mechanical Holography

• Nir Sherf, TAU, Fast Relaxation and Inference through Stochastic Resetting

• Yamin Ben-Shimon, TAU, Machine Learning the Entropy to Estimate Free Energy Differences without Sampling Transitions