Michael Burke
Thu 06 Feb 2020, 12:45 - 14:00
IF, G.03

If you have a question about this talk, please contact: Jodie Cameron (jcamero9)

Title: On the need for inductive biases in robot learning


Abstract: Advances in deep learning and deep reinforcement learning have resulted in an explosion of researchers naively throwing learning algorithms at robots. Time and time again, these approaches fail in real world settings, because robotics is not i.i.d, and is subject to numerous constraints around safety, alongside physical and sensory limitations. Using a learning from demonstration setting, this talk will discuss the need for greater structure and prior knowledge in deep learning models for robot learning and control. In particular, I will focus on how representation learning can be improved using knowledge of desired control strategies, process dynamics and symbolic constructs, or even statistical information about the generative process of demonstration sequences as a supervisory signal.