Cyborg (Illustration: Runran/flickr)
Overview

The field of AI has made dramatic progress in the last two decades. The list of triumphs runs long, from the first computer program to beat a human champion at the game of Go to a system that surpasses human scientists in predicting the structure of proteins. However, even the best of today’s state-of-the-art AI systems remain far below the level of humans on many of the core abilities associated with intelligence. While they excel in narrow domains, they largely lack the commonsense and intuitive knowledge that underlies intelligent behavior, as well as the ability to form abstract concepts and transfer knowledge fluidly between situations.  While there has been some progress on “meta-learning” — learning how to learn or what learning strategies to use in different situations, today’s AI systems are not yet capable of open-ended “lifelong” learning, unlike their naturally intelligent counterparts.

Our project aims to address foundational weaknesses in how intelligence itself is understood, which we believe is inhibiting progress at the frontiers of AI. Building on the Santa Fe Institute’s interdisciplinary culture and contributions to natural, artificial, and collective intelligence, our project will inspire new breakthroughs in AI research through concentrated, collaborative effort. To that end, we will bring together AI specialists and other researchers who think deeply about the nature of intelligence — irrespective of discipline.

Some of the most successful AI approaches have been inspired by natural intelligence. Such approaches include neural networks and evolutionary computation, as well as methods inspired by the immune system, ant colonies, and economic markets. However, the implementations of AI systems inspired by nature are often based on an outdated and shallow understanding of how these natural systems work. A broader and deeper understanding of the sources of intelligence in natural systems may dramatically improve AI, and may also provide us with more appropriate metrics of intelligence — not just the speed and accuracy of answers, but robustness to noise, flexibility in different domains, adaptability to new problems and environments, and scalability to problems involving larger sets of interacting components. Our project will advance AI research by facilitating closer engagement with multiple fields interested in natural intelligence, including psychology, neuroscience, philosophy, ethology, information theory, and many others.