Learning to adapt: paradigms and challenges in adapting deep models to domain and semantic shift

Elisa Ricci

University of Trento, IT

Abstract

Deep networks have revolutionized the computer vision field. Unfortunately, the impressive performance gains in several tasks have come at the price of the use of massive amounts of labeled data. As the cost of collecting and annotating data is often prohibitive, given a target task where no labeled samples are available, it would be desirable to build models that can leverage information from labeled data of a different but related domain. However, a major obstacle in adapting models to the target task is the shift in data distributions across different domains. This problem, referred to as domain shift, has motivated several research efforts in domain adaptation. Another major issue with deep networks is their inherent difficulty to learn sequentially over several tasks without forgetting knowledge obtained from the previous tasks. This last problem is addressed by continual learning. In this talk I will provide an overview of existing methods and most recent trends in the area of domain adaptation and continual learning for visual recognition, with special emphasis on recent approaches which benefit from self-supervised learning techniques. I will also outline important achievements in several applications and discuss the current challenges on building models which can learn continuously and adaptively in real-world scenarios.