Wednesday 18th October 2017
Ground Floor Seminar Room
25 Howland Street, London, W1T 4JG
Reconfiguration of states and dynamics in a canonical cognitive circuit revealed by unsupervised decoding
It is often unknown whether neural circuits represent the same variables during sleep as they do during waking; apart from a few examples, the identities of the encoded variables remain unknown in most neural circuits even in the awake animal. I will discuss a technique to identify and decode unknown low-dimensional variables encoded in the time-varying responses of multiple simultaneously recorded single neurons. The method is based on fitting parameterized splines to nonlinear low-dimensional manifold structure in the high-dimensional neural states.
Applied to the cognitive mammalian head direction system, I will show that the method recovers head direction with an accuracy matching supervised methods. I will show that REM sleep states are the same as waking states, but the dynamics during REM are different, and diffusive. By contrast, non-REM states are higher-dimensional. Projected onto waking states, non-REM dynamics are flickering and discontinuous, but actually consist of smooth looping trajectories on the higher-dimensional non-REM manifold. Thus, it is possible to decode an encoded variable without knowledge of its its identity, and to obtain a rich understanding of the structure and dynamics of representations across states, simply from time-series neural data.
The results shed light on the operation of a canonical cognitive circuit during sleep, and can help inform ideas on the role of different stages of sleep.