Monday April 29, 2013 at 16h30 in Celestijnenlaan 200A (auditorium 00.225)
Deep Transfer Learning: Generalizing Knowledge Across Domains
by Jan Van Haaren (PhD student DTAI)
Machine learning algorithms are said to learn if their performance on a task improves while gaining more experience. The current paradigm in machine learning defines experience as the amount of available data, and traditional machine learning algorithms assume lots of high-quality data are available. Unfortunately, this assumption often does not hold in real-world scenarios. In this seminar, I will present transfer learning as a solution to overcome machine learning's dependence on large amounts of high-quality data. Transfer learning algorithms consider data from another task, called the source task, in addition to data from the target domain.
After providing the necessary preliminaries, I will introduce a novel deep transfer learning algorithm called TODTLER, alongside two state-of-the-art approaches. I will discuss TODTLER's advantages from a theoretical point of view and present some early experimental results. Aimed at a broad audience, this seminar will provide an intuitive discussion of the current state of the art in deep transfer learning.