| Prof. Tinne Tuytelaars |
| Wed 13 Sep 2017, 16:30 - 17:30 |
| IF 4.31/4.33 |
If you have a question about this talk, please contact: Allison Kruk (v1atayl6)
Speaker: Prof Tinne Tuytelaars
Title: Keep on learning
Abstract: We can envision an ideal version of an AI system for visual recognition, that can cope with a wide range of classes or tasks and keeps learning new tasks on a continuous basis, accumulating knowledge over time resulting in both forward as well as backward transfer, while scaling properly both in terms of memory and storage resources. In this work, I'll present a couple of steps we've undertaken moving in this direction, in the context of image classification. First, I'll briefly talk about a system consisting of a growing architecture with multiple experts (one per task), where we developed an expert gate to select the right expert at run time (work presented at CVPR17). Next, I'll discuss our work on a shared model, where catastrophic forgetting is reduced by using task-specific encoders (work to be presented at ICCV17) as well as some of our latest work exploiting Hebbian synapses towards the same goal.