Gautam Singh on Representation Learning for Systematic Generalization | Toronto AIR Seminar скачать в хорошем качестве

Gautam Singh on Representation Learning for Systematic Generalization | Toronto AIR Seminar

Abstract: While the underlying fundamental structure of the physical world is compositional and modular, in general, this structure is hidden in modalities such as images and it is quite elusive how one may discover it. Uncovering this hidden structure through the observation of unlabeled data is crucial for fully harnessing the vast amount of available data today. In this talk, I will discuss my recent architectural contributions in the line of unsupervised object-centric learning. I will highlight the benefits of incorporating powerful decoders, such as transformers, which not only simplify the conventional architectures but also enable them to scale to more complex scenes and potentially expand to modalities beyond images. Additionally, I will introduce a novel binding mechanism capable of not only discovering objects but also decomposing factors of within-object variation, which are often entangled in traditional slot-based methods. Paper: Singh, Gautam, Fei Deng, and Sungjin Ahn. "Illiterate DALL-E Learns to Compose." ICLR. 2022. Singh, Gautam, Yi-Fu Wu, and Sungjin Ahn. "Simple Unsupervised Object-Centric Learning for Complex and Naturalistic Videos." NeurIPS. 2022. Singh, Gautam, Yeongbin Kim, and Sungjin Ahn. "Neural Systematic Binder." ICLR. 2023. Bio: Gautam is a PhD student at Rutgers University, supervised by Sungjin Ahn. His research focuses on unsupervised representation learning, specifically aiming to learn compositional representations such as object-centric representations, which can support systematic generalization in downstream tasks. He has contributed to developing architectures that enable unsupervised emergence of object-centric representation, not just in simple scenes (as was the state of the art previously) but also in visually complex scenes with the ultimate goal of scaling it to real scenes. Before starting his PhD, he worked as a research scientist at IBM Research India after obtaining his bachelor's degree from IIT Guwahati. Toronto AIR Seminar: The Toronto AI Robotics Seminar Series is a set of events featuring young robotics and AI experts. The talks are given by local as well as global speakers and organized by the Faculty and Students at University of Toronto’s Department of Computer Science. We welcome students, researchers and robotics enthusiasts from around the world to join us and interact with the Toronto Robotics Community. Find out more at: https://robotics.cs.toronto.edu/toron...