GoodAI has partnered with the Leverhulme Centre for the Future of Intelligence, Imperial College London, and the University of Cambridge to bring you the competition which will benchmark AI agents against multiple animal species using a range of established animal cognition tasks.
The agents will be tested on 100 tasks split testing visual processing, navigation, spatial memory, tool use, and basic causal reasoning amongst many other cognitive skills. While some of these tasks may seem trivial to a human — and to some animals — they can be challenging for AI agents. In particular, while AI agents may be good at solving a specific task, their intelligence is rarely general enough to tackle a number of previously-unseen and varied tasks.
The prize pool has recently tripled to $30,000 thanks to the support of AWS, AIJ, and a standalone prize of $4,000 from the Whole Brain Architecture Initiative for the most biologically plausible entry.
The competition will officially launch in June 2019; however, potential participants can already start preparing using the “playground”, a virtual environment which includes many different objects for the AI agent to interact with. Participants can customize their “playground” while training, by changing the objects and the layouts, but the exact details of the testing environments will remain secret.
Marek Rosa, CEO and CTO of GoodAI, said: “On the journey to reaching human-level artificial intelligence it is always a challenge to measure the exact progress. This competition is a great starting point and introducing tasks from the animal cognition world will help compare cognitive capabilities between AI agents and animals. I am looking forward to seeing how participants go about the challenge.”