Will man learn from machine?

Drum roll please. Experts have hailed the amazing achievements of AI agents that pulled off a resounding triumph over…Ms Pac-Man.

No kidding. A team from Microsoft-owned Canadian firm Maluuba used a ground-breaking method called ‘reinforcement learning’ to help AI achieve a perfect score of 999,990 in the popular video game.

So how did it work? While most AI is powered by people-generated data, in order to conquer the beribboned Ms Pac-Man, some 150 independent AI agents set to work either finding pellets or avoiding ghosts. A ‘senior AI manager’ took all suggestions on board, but ultimately decided where Ms Pac-Man should move in order to maximise pellet-collecting and the chance of survival.

Senior research manager at Maluuba, Harm Van Seijen, told the BBC: “There’s this nice interplay between how they [AI] have to, on the one hand, co-operate based on the preferences of all the agents, but at the same time each agent cares only about one particular problem.”

So far the rise of AI has been rather one-sided, with machines gaining intelligence using data gleaned from human behaviour. Now it seems that reinforcement learning could change this. As the technology develops, could we potentially learn from AI just as it has learned from us? Should we sit up and take notice of the success of ‘reinforcement learning’ and try to make the human decision-making process more like AI’s?

Insights posed this question to world-renowned AI expert, Professor Nello Cristianini of the University of Bristol, who has written extensively about AI for publications such as the New Scientist. It turns out he is rather sceptical about whether AI-inspired techniques could be successfully applied to human processes in the workplace, academia or elsewhere.

Nello says, “Modern AI is statistical and data-driven, and I do not see how we could train, say, a student to perform the same operations. While current AI systems are very effective, they do not work in the same way we work.”

So what are the key differences?

“For example, image recognition, product recommendation and spell checkers, all make use of statistical correlations discovered in very large sets of data, to compute probabilities of certain events,” says Nello. “A pattern in the pixels of an image might increase the probability that this image contains a lion, but this pattern might be meaningless to us; a certain statistical relation in the log file of Amazon might mean that a certain book might be of your interest, but this would not translate into anything that you can recognise in terms of genre of the book, or personality of the reader.

“Correlations are very useful to make predictions, but are not causations: modern AI bypasses the need to understand the mechanisms behind things, the need for causes, or explanations, and is satisfied with making useful predictions.”

To summarise, by their very design AI processes are not something humans can emulate. However, perhaps conscious of a fast-changing and constantly disrupted world, Nello doesn’t rule out the prospect of humans learning from machines entirely.

He says, “The way in which you detect a spam email is different than the way in which your spam filter does it, but they can both be useful. We do not know if this will change in the future. AI has surprised us before.”

AI-inspired processes? Watch this space…