Grant: Canada Research Chairs

The proposed research is in artificial intelligence (AI), which spans Computer Vision and Machine Learning. Specifically, Dr. Leonid Sigal proposes to study the use of computers to automatically perceive and understand visual content in typical everyday scenarios, in order to make predictions or help make informed decisions. A key component of such AI systems is learning, allowing acquisition of sophisticated models of the visual world from large image or video corpora. This program focuses on fundamental under-explored scientific challenges, including learning from limited, structured and/or multi-modal data (images + descriptive text + sound), by developing novel neural architectures and objectives. Dr. Sigal and his research team are working on a branch of learning algorithms known as deep learning or deep neural networks: biologically-inspired algorithms and architectures that systematically learn to extract hierarchical representations and abstractions from data to help with visual cognition tasks. Such algorithms are at the core of most recent artificial intelligence advances, from speech recognition to face- and image-tagging. In the context of such architectures, Dr. Sigal is focusing on use of multiple modalities of data. Not only can visual representations and interpretations be enhanced by other modes of perception; but also such representations provide a direct ability for vision systems to naturally interact with the human users and operators. Further, in comparison to humans, most current vision algorithms require very large amounts of fully labeled data (e.g., images with human-annotated categories or object boxes). To address this, he and his team are developing general systems that can perform multiple visual tasks and are able to deal effectively with all forms (e.g., image-, region- or pixel-level), and amounts, of supervision. Finally, the program addresses ability to build content creation tools that produce indistinguishably from real images, and videos, subject to interactive user specifications. While addressing these fundamental scientific challenges, Dr. Sigal’s research program is also intent on applying these technical advances to build solutions for real-word problems, spanning visual data curation, content creation, augmented reality (e.g., virtual try-on) and medical diagnostic applications. Ultimately, this work may support a wide variety of applications and have a significant impact on both academia and industry.