AI has made rapid progress in recent years, overtaking human performance on complex tasks. However, there remain many areas where AI cannot compete with even simple biological organisms. Such areas include understanding object permanence and intuitive physics, one-shot learning, intelligent exploration and generalising to unexpected situations. This project explores these areas looking towards understanding and then closing the gap between biological and artificial intelligence.
The Animal-AI Olympics is an AI competition using experiments translated from the animal cognition literature. Animal cognition has a long history of creating ingenious experiments for empirically investigating intelligent behaviour across biological species. A standard experimental paradigm begins by familiarising an animal with a new apparatus and training it until it can successfully complete a simple task. Once the animal achieves a certain level of performance, a new dimension is introduced to the task, constructed to test whether the animal has learnt generalisable properties to solve the problem or was just relying on blindly repeating previously successful actions. This approach is in stark contrast to most modern AI testbeds, where the training dataset is often drawn from exactly the same source as the test set. It is only recently that progress in deep learning has made it possible to even consider such a competition. We hope that it will illuminate the recent successes in AI, and also identify areas in which there is still a long way to go to meet animal-level general intelligence.