PhD Dissertation Proposal by HILLIER Dylan Alexander Slavin | Situated Social Learning:How Can Agents Learn Alongside Other Agents?

ABOUT THE TALK

Much of what humans know is learnt from others, by copying them, being taught and learning together. Most AI systems instead learn from individual experience or from existing human data, which reveals little about how social abilities develop. This thesis asks how social learning can develop in situated agents, which, like children, learn through interacting in a shared environment. Such agents face a chicken-and-egg problem: learning from others requires understanding them, yet agents have no reason to develop that understanding until it helps them learn. Existing work largely sidesteps this by building the social part in, or studies only the simplest form, imitating an expert.

For imitation (completed), we show across environments from gridworlds to simulated robot locomotion that learning from others is closely linked to modelling them, and that general-purpose learning signals can help drive the emergence of both. For teaching (in progress), we introduce environments based on how animals teach, such as ants leading a nestmate to food. Teachers learn to teach in these environments and come to represent both what a pupil knows and how capable it is. We close with plans to study how teaching adapts to different pupils.

More broadly, the thesis seeks to understand how social learning can develop without being designed in. This understanding could inform theories of childhood development, and how we build embodied agents that better learn from and teach the people around them.

ABOUT THE SPEAKER

Dylan HILLIER is a PhD candidate at the school of Computing and Information Systems, Singapore Management University, under the supervision of Prof. Pradeep Varakantham. Additionally, he is a SINGA scholar at A*STAR under the supervision of Dr. Cheston Tan. His research focuses on how learning in situated social settings relates to the ability of agents to model each other.