Lifelong Learning for Visual Recognition Systems

Recent breakthroughs of machine learning and visual recognition are fundamentally impacting our today's society. Two impressive show-cases are autonomous vehicles in automotive industry or smart diagnosis systems in modern health care. However, current solutions are static by design, i. e., once they are trained, they heavily rely on non-changing environments and tasks. In contrast, humans are incredibly successful in adapting to new environments by asking appropriate questions, in recognizing previously unseen objects, and in interactively expanding their knowledge over time.
In this thesis, we draw inspiration from human learning by transferring individual aspects to continuously learning systems and introduce the lifelong learning cycle as a combination of involved topics.