Noisy brain: Statistical properties of the neural activity in the brain

Relatively reliable behaviour obscures the fact that brain is an inherently stochastic system. From the thermal fluctuations of the molecules to the sparse sampling of external inputs, at every spatial and temporal scale noise is an ubiquitous part of information processing in the brain. How does such a system represent information in a reliable manner?

Recent technological advances have made it possible to record activity of 100s of neurons simultaneously. Analysis of such activity has revealed two surprisingly features of the stimulus evoked activity in different brain regions. First, as the animal engages in a task, the activity of neurons increases but the variability of the activity decreases. This is surprising because in a stochastic system increase the activity should also increase the variability. Interestingly, the amount of variability reduction is positively correlated with behaviour performance. Second, activity of a network is confined to a low-dimensional manifold. This is surprising because neurons are variable and weakly correlated so naively one would expect that the activity should have a relatively high dimensionality — comparable to the number of neurons in the network. Curiously, the low-dimensional nature of the activity constrains the behavioural repertoire of the animals.

In my talk I will discuss various mechanisms that may underly the quenching of variability of stimulus induced activity. Next, I will present a synthetic approach to design neuronal network whose activity is confined to a low-dimensional manifold and discuss why the dynamical structure of the neuronal activity may constrain the behavioural repertoire of an animal.