Scale-invariant neural dynamics for cognitive behavior
This event is part of the Departmental Seminars.
The brain is operating in a world with rich dynamics across a wide range of timescales, therefore the collective activity of the large of number of neurons in the brain should reflect this multi-scale dynamics. Limited by experimental techniques and the nature of laboratory behavior, most established results in systems neuroscience concern the short-term responses of single neurons to features in the world. Recently, new techniques for large-scale and chronic measurements of neural activity open up the opportunity to investigate neural dynamics across different timescales. In this talk I will present modeling, theoretical and data analysis work on a particular type of temporal dynamics - scale-invariant dynamics - which has been implicated by both behavioral experiments and neural data. I will start with a neural circuit model that combines slow intracellular calcium dynamics with the Post approximation of the inverse Laplace transform to produce scale-invariant sequential neural activity and point out evidence for the elements of the model in neural data. I will then present a theoretical analysis on the ability for a linear recurrent neural network to generate scale-invariant neural activity. I will show that the network connectivity matrix should have a geometric series of eigenvalues and translated eigenvectors if the eigenvalues are real and distinct. Similar but less compact results hold for matrices with complex or degenerate eigenvalues. Finally I will show the existence of repeatable neural dynamics on the timescale of minutes in multiple neural recordings of rodents performing various cognitive tasks. Taken together, these results show that neural activity has important dynamics over a much wider range of timescales then previously thought, and that the requirement for scale-invariance provides interesting constraints for neural circuit models.
Zoom Meeting https://bostonu.zoom.us/j/365061062
Meeting ID: 365 061 062