Project
Continuous learning
The world around us is continuously evolving (think e.g. of fashion, news items or social media trends). A static artificial agent, trained in the lab and then deployed in the real world, will quickly get outdated. To keep up the pace, an artificial agent should be evolving as well, gradually increasing its knowledge about the world and expanding its horizons. In this project, we study how to make an artificial agent learn multiple tasks, while keeping its memory footprint constant. This involves retaining learned knowledge, adapting to particular working conditions, and improving its internal representation (embedding) along the way. In particular, we will focus on visual image understanding, including complex tasks such as multiple object interactions (triples of ) .