Скачать или смотреть Visual Deep Learning with Limited Labels or Data

Abstract:
It is an exciting time for research in computer vision. Significant breakthroughs have happened by the adoption of deep learning in the field. However, much of this progress is built upon massive data collection and annotation endeavors. The question remains -- how can we reduce this effort and make machine learning algorithms learn more efficiently with less data or labels, much like human do? In this talk, I will present several of our recent works geared towards addressing this challenging problem for visual tasks. I will describe our work on learning decoding transforming encoder-decoder architectures for few-shot and semi-supervised learning; on self-supervised learning of viewpoint using controllable generative adversarial networks, along with an analysis-by-synthesis framework; and on the self-supervised discovery of 3D shapes from image collections and videos. I will conclude the talk with a look to the future and share my thoughts on open questions that remain unaddressed in learning from limited labels or data.

Bio:
Shalini De Mello is a Principal Research Scientist and Research Lead in the Learning and Perception research group at NVIDIA. She is interested in computer vision, machine learning, human-computer interaction (HCI), and learning with limited labels or data. At NVIDIA, she has invented technologies for self-supervised and few-shot learning; gaze, and 2D and 3D head pose estimation, and hand gesture recognition, among others. Her research has pushed the boundaries of HCI in cars and has led to the development of NVIDIA’s innovative DriveIX product for smart AI-based automotive interfaces. She received doctoral and master’s degrees in Electrical and Computer Engineering from the University of Texas at Austin in 2008 and 2004, respectively.