C-BRIC JUMP e-Workshop

Abstract: Being able to learn continuously over a lifetime is of critical importance in biology, but remains a challenge for machines. A fundamental problem, known as Catastrophic Forgetting, is that deep learning as an optimization process seeks to optimize network parameters for the current data, which may yield suboptimal settings for older data. I will present several new methods recently developed to avoid such forgetting. In one method, we learn both a model for the task at hand (e.g., image classification), and a generator for the learned data (using a generative adversarial network, GAN). When the classifier is trained with new, previously unseen image classes, we avoid forgetting about previously learned classes by generating instances of those, and inter-mixing them with the new classes. In another approach, we train out-of-network bias units for each task (e.g., each set of image classes to be learned in succession). This allows the network to learn new tasks with minimal interference towards older tasks, as the bias units specific to each task place the network in a different operating regime for each task. Another method deploys out-of-network biasing units for each task, allowing the network to operate in better separated realms for different tasks. On standard benchmark datasets, we show that these methods surpass the state of the art in minimizing interference across tasks, and thus in enabling lifelong learning.

 

Bio: Laurent Itti received his M.S. degree in Image Processing from the Ecole Nationale Superieure des Telecommunications (Paris, France) in 1994, and his Ph.D. in Computation and Neural Systems from Caltech (Pasadena, California) in 2000. He has since then been an Assistant, Associate, and now Full Professor of Computer Science, Psychology, and Neuroscience at the University of Southern California. Dr. Itti's research interests are in biologically-inspired computational vision, in particular in the domains of visual attention, scene understanding, control of eye movements, and surprise. This basic research has technological applications to, among others, video compression, target detection, and robotics. Dr. Itti has co-authored over 150 publications in peer-reviewed journals, books and conferences, three patents, and several open-source neuromorphic vision software toolkits.