Zoran
Tiganj
Boston
University
Memory Across Scales: Integrating Computational Models and Electrophysiological Data
It is well known that, all things being equal, the accuracy of mammalian memory is better for events that took place at more recent past than at more distant past. I will present a biologically plausible computational framework that can account for this gradual decay of memory over multiple seconds. The framework relies on sequentially activated time cells that constitute an internal timeline. Information about what happened when is dynamically updated and always available in the brain. Having an internal timeline makes various useful computations easily achievable. For instance, if an animal knows its momentary speed, it can compute spatial distance from landmarks and construct place cells. Also, from the internal timeline it is straightforward to construct prediction of the future that is based on the spatiotemporal structure of the input signals. I will discuss the utility of these computations in the context of brain-inspired machine learning and artificial intelligence. Finally, I will present single unit data from electrophysiological recordings in rodents that support the existence of a neural timeline.