A recently published paper co-authored by Senior Psychology lecturer Leonidas Doumas ‘A Theory of Relation Learning and Cross-domain Generalisation’ puts forth new arguments on the key differences between human and machine learning systems. Leonidas explains this below.
In the past few years there has been a lot of press about machine learning systems doing some pretty remarkable things. For instance, groups from Google’s DeepMind have published several high profile papers showing machine systems can out-perform humans on a number of video games, and board games like chess and Go. One of the “dirty little secrets” in the machine learning community, though, is that these systems, while spectacular at single tasks, are terrible at generalisation. So, while a machine system might learn to play one game better than any human, that system will perform no better than a completely untrained system when trying to play a related game. For example, the model that learned to play chess was no better than a completely untrained network when trying to play go, and a machine trained to play a video game like Breakout was completely unable to transfer that learning to a related game like Pong.
Humans, by contrast, are exceptional generalisers. We routinely use what we know about one domain to reason about another. When we learn a concept like “more” we can apply it to anything like money, bales of hay, or patience. And when we learn to play one video game, we get better at playing related video games for free.
My colleagues have developed a machine learning system inspired by how humans learn about the world. This system learns how to represent the world from examples, and then uses its representations to make analogies between situations it has experience with and new situations that it comes across. We argue that a key difference between humans (and our system) and other machine learning systems is that humans learn structured (or symbolic) representations of the world and use those representations to characterise and reason about new domains. We show that just like a human learner, our system can play a video game like Pong immediately after learning to play a game like Breakout (it is the first machine system to show zero-shot (or immediate) generalisation between video games), or perform a psychology task like solving analogy problems after experience with unrelated domains.
Leonidas Doumas has lectured at the University of Edinburgh since 2013. Prior to that he worked at Indiana University and the University of Hawaii. The paper ‘A Theory of Relation Learning and Cross-domain Generalization’ was co-written with colleagues from the Max Planck Institute for Psycholinguistics, Radbound University and the University of Illinois. The full paper is available here https://arxiv.org/pdf/1910.05065.pdf