PLEASE NOTE:
- This seminar is a hybrid in-person and online event. Please click here to fill out the virtual registration form.
- In-person registration will close 24 hours before the event, and online registration will close at 15:30 on Wednesday, May 8th. Those who miss the in-person registration deadline are encouraged to register to attend virtually instead.
- Refreshments will be served immediately after this seminar.
Title:
Unsupervised manifold learning using low-distortion Riemannian alignment of tangent spaces
Abstract:
With the advent of modern experimental techniques that enable the recording of hundreds of neurons simultaneously, the focus in neuroscience has shifted towards studying the coordinated activity of neural populations, which can often be captured by low-dimensional latent representations. The organization of behaviorally-relevant neural activity in several recent works reveals striking topological organization that can be approximated as neural manifolds. Extracting physiological and behavioral information from these topologies is an active field of study that falls under the remit of manifold learning techniques. While manifold learning techniques have enjoyed much success in elucidating organizational principles in large datasets in various fields like genetics and oncology, many of these techniques introduce distortions while embedding higher-dimensional geometrical data into lower-dimensional representations. Such distortions can compromise the accuracy of insights obtained through these methods.
In this talk, I will present a novel measure of distortion to facilitate comparison across manifold learning techniques. Having established this metric, I will present a new unsupervised manifold learning technique, Riemannian Alignment of Tangent Spaces (RATS) that aims to recover low-distortion embeddings of data manifolds. I will demonstrate that compared to alternative methods, low-distortion embeddings generated by RATS excel in the visualization and deciphering of underlying function across a diverse range of neurophysiological datasets, making it a promising tool to understand neural manifolds.
Biography:
Dr. Devika Narain is an Associate Professor of Neuroscience at the Erasmus University Medical Center, Rotterdam, The Netherlands and the Principal Investigator of the Circuits for Temporal Dynamics lab. Her group combines mathematical, engineering, and systems neuroscience methods to investigate how neural circuits generate precisely-timed motor behaviors to facilitate clinical use of brain-machine interface technology.
Before returning to the Netherlands, she was a postdoctoral fellow at MIT’s McGovern Institute for Brain Research in Prof. Mehrdad Jazayeri’s lab, working on timing in frontal and thalamocortical circuits in nonhuman primates. Afterwards, she became a rodent cerebellar physiologist studying motor timing with Prof. Chris De Zeeuw in Rotterdam. Previously, she was trained at the NASA Ames Research Center in Mountain View, California and the Max Planck Institute for Intelligent Systems in Tuebingen, Germany. She holds engineering, cognitive science, and neuroscience degrees and studied in Bangalore, Palo Alto, Munich, and Amsterdam.
