Machine Justification

14 November 2017

I remember in my teens being told of the wonderful things Artificial Intelligence (AI) would do in the next few years. Now several decades later, some of these seem to be happening. The most recent triumph was of computers teaching each other to play Go by playing against each other, rapidly becoming more proficient than any human, with strategies human experts could barely comprehend. It's natural to wonder what will happen over the next few years, will computers soon have greater intelligence than humanity? (Given some recent election results, that may not be too hard a bar to cross.)

But as I hear of these, I recall Pablo Picasso's comment about computers many decades ago: "Computers are useless. They can only give you answers". The kind of reasoning that techniques such as Machine Learning can result in are truly impressive in their results, and will be useful to us as users and developers of software. But answers, while useful, aren't always the whole picture. I learned this in my early days of school - just providing the answer to a math problem would only get me a couple of marks, to get the full score I had to show how I got it. The reasoning that got to the answer was more valuable than the result itself. That's one of the limitations of the self-taught Go AIs. While they can win, they cannot explain their strategies.

Given this world, one of the big challenges I see for AI is that while we may have figured out Machine Learning in order to teach them to get answers, we haven't got systems that can do Machine Justification for their answers. As AIs make more judgments for us, we'll increasingly run into situations where the answer isn't enough. An AI might be trained in such a way to rule on legal cases, but could we accept a judgment where the AI cannot explain its reasoning?

Given this it seems likely that we will need a new class of "programmer' in the future, one whose job is to figure out why AIs get the answer they do, to deduce the reasoning underlying the AIs skills. We could see many fields where AIs make opaque judgments that we can see are good, but need another approach for us to really learn the theory that underlies their decisions.

This problem is particularly acute since we've discovered that it's awfully easy for these machines to learn undesirable behaviors from their training data, such as discriminating against racial minorities when judging credit ratings.

Like many, I see much of the opportunity of computers is in collaboration with humans. Good use of computers is understanding where the computer is strong (rapidly doing constrained work) and where humans are better, and using a mix. Computers are, at their most intellectual, a tool for the mind. In programming I'm happy to lean on the compiler to help me find errors or suggest alternatives, a practice which I was scolded for as a young programmer. That boundary between where the two are strongest is fluid, and one of the fascinations of the future is how we can best take advantage of its movement.

Further Reading

MIT Technology Review looks at the broad topic of explainability for AI.

Some articles in the dangers of machine learning and undesirable bias from The Atlantic, NPR, and Tech Republic


Brandon Byars, Chris Ford, Christoph Windheuser, Danilo Sato, Dave Elliman, Ian Cartwright, Kent Rahman, Saleem Siddiqui, Sallie Walecka, Tito Sarrionandia, and Vishal Bardoloi discussed drafts of this post on our internal mailing lists.