Searching for universality in the state space of neural networks

So-called “collective behavior” surrounds us. For example, the solidity of an object is the result of forces that act between neighboring atoms, on a scale nearly one billion times smaller than the object itself. Decades of work in theoretical physics has provided precise mathematical theories that explain how macroscopic phenomena such as solidity emerge from the microscopic interactions among atoms—i.e., from the collective behavior of those atoms. Just as matter is composed of individual atomic units, the brain is composed of individual cellular units, termed neurons. Is it possible that perceptions, thoughts, memories, and actions can be described by the collective behavior of neurons? Could mathematical ideas from physics be applied to the brain’s biology? We are taking advantage of new experimental technology for simultaneously monitoring the activity of many neurons to test these very ideas. Starting with the raw data from neurons, we will borrow strategies from physics—such as the construction of thermodynamic variables from the statistical mechanics of atoms—to build descriptions of neural networks. Preliminary analysis from both the retina and a brain region called the hippocampus suggest that such constructions are feasible. This sets the stage for discovering universal principles governing the collective behavior of neurons, much as work in physics has described universal principles governing inanimate systems.

William Bialek, Princeton University