³ÉÈËVRÊÓƵ

Jordan DeKraker
Postdoctoral Fellow, The Neuro
Supervisors:ÌýBoris Bernhardt and Alan Evans

Abstract:ÌýMany researchers study the hippocampus as a center for memory, navigation or as a crux in disease. Here, I consider the role of the hippocampus as a driver, or cost function, for deep learning throughout brain evolution and in modern AI. Modern AI takes millions of training samples and is told the corresponding interpretations (outputs) for each. Animals learn fast and through mostly unstructured exploration. I argue that the ability to predict, and especially to remember predictions over long timescales, is a key component of animal learning and an avenue towards more data-efficient and general AI. In part 2, I address the issue that current AI models don't make good long-term predictions and present a way forward through HippoMaps: a tool for quantitative mapping of hippocampal macro-, micro-circuitry, and functional recordings across domains and spatial scales. Spatial comparison of these maps can help us better understand the contrast of the acquisition methods and inter-individual differences, but also the mechanisms through which circuitry gives rise to function. Altogether, this talk leverages computational methods for making bridgeheads between theory-driven and data-driven understanding of learning and intelligence.