Speaker: Dr. Bogdan Raducanu
Lifelong Learning with Generative Memory Replay
This talk intends to introduce the concept of lifelong learning (or continuous learning) and its application to the field of computer vision. The most common way to train a neural network is considering that we have all the necessary data from the beginning. But in case of real problems, this is not the proper scenario. For example, if we consider the case of a robot that has to continuously explore new sites or video surveillance, where the representation of a person must be updated to make its identification more robust, the neural network has to be retrained to accommodate this new information. In other words, another training strategy is required which is designated as 'sequential learning'. But a recognized problem associated with this paradigm refers to 'catastrophic forgetting', i.e., the neural network tends to forget prior knowledge when re-trained with new data. There are several strategies to limit the effects of catastrophic forgetting, and one of them is 'generative memory replay'.
Dr. Bogdan Raducanu received a degree in computer science from the Polytechnic University of Bucharest (Romania, in 1995) and the title of doctor "Cum Laude" also in computer science from the University of the Basque Country, in 2001. Between 2002-2004 he was a researcher postdoctoral fellow at the Technical University of Eindhoven, in the Netherlands. In 2004 he joined the Computer Vision Center (CVC) with a "Ramón y Cajal" scholarship and since 2009 he is a senior researcher and project manager. His research interests are in the areas of machine vision and deep learning with applications to lifelong learning, generative models, and robotics.