Abstract: In this talk, I'll give four good reasons to do embodied machine learning. Learning in an agent that can perceive and interact with its environment is fundamentally different from other ML settings. Unlike a disembodied model, an embodied learner must learn to perceive and move in addition to mastering whatever specific behaviour the user may be interested in. This may seem like a disadvantage, because there is in some sense more to learn, but it can also be an advantage when trying to replicate human cognitive behaviours like reasoning and generalization. Having a body and being necessarily located at a specific place at a given time places strong constraints on the learner's experience, which in turn leads to more human-like learning outcomes. These results suggest that embodied learning may be an important part of what is needed to convincingly replicate human linguistic intuitions and behaviours in a machine.

Bio: Felix Hill is a Research Scientist at DeepMind, where he works on replicating in artificial systems how we as humans learn and use natural language. His work is at the intersection of natural language processing & understanding and reinforcement learning, particularly focusing on embodied language learning, taking inspiration from cognitive science and psychology. He completed his PhD in computational linguistic at Cambridge University in 2016 in the Natural Language Information and Processing Group, supervised by Anna Korhonen. During his PhD he also spent time working at the LISA lab, Montreal, with Yoshua Bengio.