About Zeynep: Zeynep Akata is a professor of Computer Science within the Cluster of Excellence Machine Learning at the University of Tübingen. After completing her PhD at the INRIA Rhone Alpes with Prof Cordelia Schmid (2014), she worked as a post-doctoral researcher at the Max Planck Institute for Informatics with Prof Bernt Schiele (2014-17) and at University of California Berkeley with Prof Trevor Darrell (2016-17). Before moving to Tübingen in October 2019, she was an assistant professor at the University of Amsterdam with Prof Max Welling (2017-19). She received a Lise-Meitner Award for Excellent Women in Computer Science from Max Planck Society in 2014, a young scientist honour from the Werner-von-Siemens-Ring foundation in 2019 and an ERC-2019 Starting Grant from the European Commission. Her research interests include multimodal learning and explainable AI.

Title: Explainability and Compositionality for Visual Recognition

Abstract: Clearly explaining a rationale for a classification decision to an end-user can be as important as the decision itself. Many existing approaches for deep visual recognition are generally opaque and do not output any justification for the decision; contemporary vision-language models can describe image content but fail to take into account class-discriminative image properties which justify visual predictions. In this talk, I will present my past and current work on Explainable Machine Learning combining vision and language where we show (1) how to learn compositional representations of images that go beyond recognition and towards understanding, (2) how to to learn iterative binary sub-decisions, inducing sparsity and transparency in the decision making process and (3) how to to transfer knowledge across heterogeneous modalities, even though these data modalities may not be semantically correlated.

About Olga: Dr. Olga Russakovsky is an Assistant Professor in the Computer Science Department at Princeton University. Her research is in computer vision, closely integrated with the fields of machine learning, human-computer interaction and fairness, accountability and transparency. She has been awarded the AnitaB.org's Emerging Leader Abie Award in honor of Denice Denton in 2020, the CRA-WP Anita Borg Early Career Award in 2020, the MIT Technology Review's 35-under-35 Innovator award in 2017, the PAMI Everingham Prize in 2016 and Foreign Policy Magazine's 100 Leading Global Thinkers award in 2015. In addition to her research, she co-founded and continues to serve on the Board of Directors of the AI4ALL foundation dedicated to increasing diversity and inclusion in Artificial Intelligence (AI). She completed her PhD at Stanford University in 2015 and her postdoctoral fellowship at Carnegie Mellon University in 2017.

Title: Perception, interaction and fairness: key components of visual recognition

Abstract: There is growing attention to the critically important questions of fairness, privacy, transparency, ethics, and responsible use of computer vision. In this talk, I'll highlight some of our recent research contributions in this space. I'll also discuss the importance of diversity and inclusion to making progress. There is much work to be done, and everyone's expertise is much-needed -- one of the goals of the talk would be to showcase paths forward for anyone who wants to get involved in this line of research.

Student Speakers: Nobline Yoo, Sunnie S.Y. Kim and Angelina Wang

About Natalia: Natalia Neverova is a Research Scientist at Facebook AI Research (FAIR), interested in statistical machine learning and computer vision with emphasis on deep learning and human centered applications. Before coming to FAIR in 2016, she completed her PhD at INSA Lyon (France) and the University of Guelph (Canada) under the guidance of Dr. Christian Wolf and Dr. Graham Taylor, working on deep learning and computer vision models for human motion analysis. She also spent several months as a visiting researcher at Google ATAP.

Title: Towards universal predictors for object correspondences

Abstract: We will talk about the problem of learning inter- and intra-category dense correspondences, which is one of the fundamental problems of visual recognition, also tightly coupled with 3D reconstruction tasks. We will see how training multi-class predictors for establishing correspondences, together with enforced cycle consistency terms, allow for automatic discovery of structural relationships between different object categories (such as people and animals) in un unsupervised way, as well as achieving state-of-the-art alignment of 3D meshes without using specialized techniques.

About Diane: Diane Larlus is a principal research scientist at Naver Labs Europe and leads a Chair on Lifelong representation learning within the MIAI research institute of Grenoble. Her research mainly focuses on learning representations with weak supervision, continual learning, and visual search. As a PhD student, she worked at INRIA Grenoble, France. After a postdoctoral experience at TU Darmstadt, Germany, she joined the European research center of Xerox. She now works at NAVER LABS Europe.

Title: Lifelong visual representation learning

Abstract: Computer vision has found its way towards an increasingly large number of applications. One reason for this success is the development of large and powerful deep learning architectures which produce visual features that are generic enough to be applied directly to - or be the starting point of - a large variety of target tasks. Yet, training such generic architectures requires large-scale data, extensive human annotations, and heavy computational resources. In order to reduce the training cost of transferable descriptors, recent approaches have looked at ways to allow for noisy, fewer, or even no annotations to perform such pretraining. The first part of this presentation will cover our recent contributions in this direction. Second, we consider the deployment of a model which is sequentially exposed to different visual domains and incrementally adapts through model updates. Most standard learning approaches lead to fragile models which are prone to drift when such updates are performed, a problem known as the 'catastrophic forgetting' issue. During this talk, we will discuss a strategy based on meta-learning and domain randomization designed to mitigate catastrophic forgetting.