Fiete Lab; McGovern Institute, MIT
Cognitive maps for navigation in the brain
A good understanding of the world depends on factorizing it into its parts and learning the regularities within it. The brain solves this problem both generally and in specific problem domains. The hippocampal complex -- famous from human studies for its role in constructing memories involving the who, what, when, and where of biographical events – has been a rich playground for understanding the neural circuits underlying these functions. This is because neural activity recorded from the hippocampus and associated areas provides a strikingly explicit representation of abstract low-dimensional variables, especially those that are key to solving cognitive problems but are latent, or not directly specified by sensory input. Such neural representations in the hippocampus were originally discovered in the form of spatial maps recorded from freely behaving rats (work which was awarded the 2014 Nobel prize), and recently several groups have begun to explore such cognitive maps in non-spatial domains. I will 1) highlight the important pieces of the phenomenology of these representations, 2) describe mechanistic circuit models, and 3) briefly summarize recent progress in characterizing the nature of these representations through unsupervised discovery of latent low-dimensional structure from population data. I will conclude with an overview, based on recent experimental work, of how understanding these circuits in the realm of navigation gives insights into their use in non-spatial cognitive representations as well.
Center for Brain Science, Harvard University
Manifold discovery of cognitive maps
Historically, a fundamental paradigm in systems neuroscience for elucidating brain function has been to identify "single-neuron receptive fields." From this point of view, a functional description of a neural system is equated with a list of single-cell receptive fields, i.e., a description of the trial-averaged activity of each cell, conditioned on one or more sensory or behavioral variables. However, in light of the on-going proliferation of exciting multi-neuronal recording techniques, there has been a push from the theoretical and computational communities to develop techniques and ideas which can leverage population recordings to provide insightful new descriptions of brain function. In this seminar, I will describe two synergistic research projects (one by myself and Ryan Low, the other by Rishi Chaudhuri and Ila Fiete), organized around the idea of studying neural circuits in terms of the "intrinsic geometry" of their population activity. Specifically, I will give a detailed description of the manifold learning paradigm from machine learning and data science, and how we have applied novel techniques of this kind to infer properties of "cognitive maps" represented in the neural activity recorded from the hippocampus and thalamus of rats navigating in real and abstract spaces (as described in the primer by Ila Fiete). These techniques provide a language for characterizing neural population activity in a concise and unsupervised manner, and more importantly, we will demonstrate their potential for discovering novel properties of neural systems.