PhD Talk: Few-shot behavior learning

We humans can remember a face after only seeing it once, we can also grasp and identify an object we have never interacted with before. We perform those motor-perceptual tasks with a limited amount of information and experience. In an opposite manner to our capabilities, the current state-of-the-art algorithms for computer-vision and high-level manipulation use deep learning methods which consume a great number of training samples and obtain a limited generalization. This is specifically problematic for robotic systems that are supposed to operate in a wide range of environments while performing multiple different tasks. The urging question thus is: Which prior elements and algorithms would allow a robot to learn those motor-perceptual tasks with the same flexibility and limited resources that a human has and uses? In this talk, I present some of the limitations of deep learning methods and I propose to start a PhD thesis in which I research how to learn robot behaviors from less information.