Yoram Burak, Edmond and Lily Safra Center for Brain Sciences, Racah Institute of Physics, The Hebrew University of Jerusalem
In recent years there has been considerable interest in understanding theoretically how a neural network\\\'s architecture affects its ability to maintain information about the past. I will discuss this question in the context of networks that encode a continuous variable, such as an angle or position, within a persistent memory state, as observed in diverse brain areas. Since neural activity is noisy, the stored variable is expected to gradually drift from its initial state, leading to degradation of the precision in which the memory can be recalled. I will introduce a fundamental statistical relation that relates this process to another aspect of noise in the neural activity, namely: the limited rate at which network spikes transmit information about the stored memory to an external observer. This relation takes the form of a rigorous inequality. I will explain the inequality, and will then consider several interesting consequences.
University of Heidelberg, INF 328, SR 25, 15:00-16:00