CS Events

PhD Defense Neural Unsupervised Structure-Aware Representation Learning
Export Event Save iCal Send to Google Calendar Send to Yahoo Calendar Send to Outlook Live Send to Microsoft Outlook Save iCal Preserve formatting in description (only supported in some calendar applications) Close Tuesday, November 26, 2024, 09:00am - 10:15am

Speaker: Gautam Singh

Location : CoRE 305

Committee: 

Professor Sungjin Ahn (Chair)

Professor Hao Wang

Professor Yongfeng Zhang

Prof. Francesco Locatello (External)

Event Type: PhD Defense

Abstract: Although deep learning models have shown an impressive performance, they still lack in several important aspects such as robustness, systematic generalization, interpretability, reasoning, and the ability to create new knowledge from limited experience. To address these limitations, learning representations of data that capture its underlying hidden structure is thought to be crucial. This dissertation develops architectures and algorithms for learning structure-aware representations without any human supervision or labels, aiming to capture (1) the 4D spatiotemporal structure and (2) the compositional object-centric structure of visual scenes. In Part One, we learn the 4D spatiotemporal structure by integrating observations over time and employing predictive coding. This improves novel view synthesis over baselines, pointing to a superior representation of the underlying scene geometry and dynamics. In Part Two, we introduce a novel object-centric representation learning method by inverting a flexible decoder, demonstrating for the first time the ability to decompose complex scene images and render systematically novel images. We further extend this method to handle videos through two routes: a sequential route via recurrence and a parallelizable route via causal attention. In Part Three, we shift focus towards systematic generalization and concept reuse. We develop a novel method that not only disentangles objects but also learns intra-object factor concepts that are optimized for reusability across scenes. Lastly, we develop a novel image-to-image translation benchmark to measure the ability of deep learning models to generalize to systematically novel visual scenes.

: 

Contact  Prof. Sungjin Ahn

Zoom Link
https://rutgers.zoom.us/j/91834809053?pwd=Ej92OzVV0RtGlZVeC3Lh2VJyaaRYcB.1