Matt Crosby (Imperial Collge London): Kinds of Intelligence: Testing AI on Animal Cognition Tasks

Thursday morning lectures

ABSTRACT: I will give a brief overview of the ongoing work of the Kinds of Intelligence team at the Leverhulme Centre for the Future of Intelligence and then focus on the Animal-AI project. Our aim is to bring insights from comparative cognition to the AI community and facilitate work towards more general, and cognitively-inspired, kinds of intelligence. To this end we recently ran the Animal-AI Olympics competition. Participants were given a small simulated arena with simple physics and a few basic objects to interact with. Their goal was to submit a robust food retrieval agent. The only information they had about the tests were that they were constructed in the arena out of known items, required the retrieval of designated food objects, and were based on tests passed by animals. We also provided a syllabus of capabilities required to solve our testbed including navigation, spatial memory, object permanence, and simple causal reasoning. Now that the competition has finished, we can reveal our testbed and present it as an open challenge for AI. I will also present our plans to extend the environment and eventually move the paradigm from simulation to the real world.

BIO: Matthew Crosby is a postdoctoral researcher on the Kinds of Intelligence project at the Leverhulme Centre for the Future of Intelligence. He is based at Imperial College London where he works with Murray Shanahan on the broad area of artificial cognition. He has an interdisciplinary background spanning AI, robotics, cognitive science, and philosophy of mind and is interested in modern AI approaches that span this range of disciplines. His long-term goal is to better understand what it means to ‘have a mind’ by knowing how to build an artificial system that has one.