Learning without neurons
Learning and neural computation are usually associated with neural networks; but many of these properties come from them being highly interconnected many-body systems with collective dynamics. To what extent can inevitable processes in disordered non-equilibrium physical systems show similar neural network-like behavior? We study these questions in the abstract as well as in disordered molecular and mechanical systems. Learn more about our physical learning research.
Recent representative work:
Neural network-like behavior in the collective dynamics of molecules (Nature)
Supervised physical learning in mechanics (PNAS) A review: Learning without neurons (Ann. Rev. Cond. Mat.)
Recent collaborators:
Nagel, Winfree
Molecular information and evolution
The success of life derives from its ability to exploit molecular information. We study how the physical nature of information in life introduces constraints but also opportunities for error correction and compact information processesing that are not possible with abstrated notions of information. This work combines non-equilibrium statistical mechanics, information theory and molecular biology to understand the minimal ingredients needed for robust self-replication and evolution. See our molecular information publications for more details.
Recent representative work:
A minimal scenario for the origin of non-equilibrium order (arXiv)
Closed ecosystems extract energy through self-organized nutrient cycles (PNAS)
Physical soft modes constrain epistasis (MBE)
Recent collaborators:
Elowitz, Hwa, Pincus, Szostak
