Prediction Machines book cover

Book interview: Joshua Gans, ‘Prediction Machines’

Image credit: Harvard Business Review Press

As artificial intelligence becomes cheaper we can use it to make more and better predictions. Economist Joshua Gans argues that this is the application AI has been waiting for.

“Artificial Intelligence is this phrase that we use when we are discussing whether we can replicate the human mind – or something much, much better – with machines.” So says Joshua Gans, one-third of the authorship trio responsible for the new book ‘Prediction Machines’, which sets out to explain the ‘simple economics of AI’. But, continues Gans, “the recent 10-year explosion in AI methods is actually not doing that as a whole thing.”

Today’s machine-learning and deep-learning methods, says the University of Toronto economist, “only do one thing, which is prediction. They take a set of data and generate information you need to make a decision.” They can’t reason and they’re not very good at extrapolation and don’t always have perspective: “and so there are a lot of other parts to ‘intelligence’ missing. Sometimes if they are good at one specific job – such as telling you what is in a photograph – it can look like they’ve got those things. But, ultimately all that’s going on under the hood from a statistical perspective is they are getting better at prediction.”

While this might sound as though it is a deliberate narrowing of scope for the technology, Gans thinks that it’s a “useful way of looking at AI because its real job is to lower the cost of prediction, and once you start to lower the cost you can start to ask yourself what you can do with that. If we now have a prediction in a place where we didn’t have one before, what would that open up for us?”

It’s not about job-killing robots, self-driving cars and self-managing organisations – the concepts that inspire simultaneous wonder and dread in the readers of tabloid newspapers – but the increased ability to make better, cheaper decisions in environments where we are surrounded with what looks like uncertainty.

 

We read it for you

Prediction Machines

One of the biggest engineering challenges is how to make better predictions in an environment of greater uncertainty. One of the ways we can address this is by leveraging Artificial Intelligence to make the future more certain. Based on the fundamental economic principle that as commodities become cheaper they become more abundant, the authors contest that as the price of AI comes down, so does the price of making decisions, which means that we can make more of them more accurately. But decision-making has more to it than data analysis – it requires judgement, which is where humans come in.

 

Although it’s tempting to separate AI (or the ‘prediction machines’ of the book’s title) from human thinking along the fault line that divides objectivity and subjectivity, Gans’s interpretation of the relationship between the two is far more subtle. He admits that there is a case to be made for machines crunching data while humans work out how to convert the outcomes into decisions, but “there’s a lot more going on. They can predict subtle things that we’ve never noticed before.”

Gans cites a start-up in Toronto that employs deep-learning methods to understand the demand for certain supermarket products. “Now, you would think that if you were in charge of restocking a product such as yoghurt, you’d pretty much know what drove demand instinctively. So how is a machine going to help you with that?” Gans goes on to explain that once the numbers had all been crunched it was found that a big influence on yoghurt demand was outside temperature… “even in Canada. So if it dropped from -5°C to -10°C that could have a huge effect on demand. This is important because until that point people hadn’t taken into account temperature in yoghurt demand planning. They could see that it was seasonal, but not on a day-to-day basis.”

This is different from what we’ve learned by experience. We’ve always intuitively known that when there’s a big televised football match on, supermarkets will sell more beer and demand for electricity will go up at half-time when those not drinking beer will boil kettles to make tea. “But what’s really getting people excited is the stuff we’ve never seen before. So we’re running these prediction machines and we’re really starting to learn something about the magnitude of demand, which means that money can now be saved on over-stocking or avoiding stock shortages. For want of a better term, these are insights that we can react to quickly rather than have a committee debate about. And these examples crop up all over the shop.” Gans says that while it is ‘great’ to have this ability to predict, “sometimes it’s not enough.” True to his economics background he is not just interested in rewards, but also costs associated with error so we can use these predictions to make effective judgements.

Gans admits that while the prediction machines of the book title might conjure up images of sci-fi futurology, the reality is that we are “talking about computers. It’s an algorithm that has arisen out of a method of observing a lot of data and outcomes. You can represent that as a lot of ‘if-then’ statements and it can then be distributed and used. One of the interesting things about this – and this is the learning aspect – is that they can improve the algorithm as they get more and more information.”

‘AI’s real job is to lower the cost of prediction.’

Joshua Gans

If you are a product manager you have a choice: either deploy an algorithm that works well or see if there are benefits to updating it on-the-fly, which contains the risk “that the algorithm might do something stupid. There is a trade-off here, and the big concern is making errors.”

Gans then tells a familiar story. Because the economist dislikes shopping for clothes he only does so infrequently. “I know that by my second purchase on that day I’ll get a call from my credit card company wanting me to verify that the card isn’t being used fraudulently. Now, there’s a cost to that for both parties. I’m now inconvenienced and the company is wasting resources checking I’m not a teenager who’s stolen my card. That’s generated by the machine. When there are different judgements in place – and the credit companies are starting to get better at this – say, related to the level of service, then that judgement comes from product management policy.”

Gans reiterates that the main application at the moment for AI is prediction. But the problem is, he says, “you can’t just let this loose and see where we go from there. We’ve got to get better at targeting, especially in dynamic advertising. But we’re writing a book about this in a dynamic time and I think we’ll get better at solving these sorts of problems, partly because we’re aware of them and can articulate them, and we may soon be reflecting on a time when it didn’t work so well. And part of our book says: ‘Sure, people are predicting the wrong stuff, but we’ve got to assume that they’ll work that out.’”

‘Prediction Machines’ by Ajay Agrawal, Joshua Gans and Avi Goldfarb is from Harvard Business Review Press £22.00

Extract

‘Cheap changes everything’

Cheaper prediction will mean more predictions. This is simple economics: when the cost of something falls, we do more of it. For example, as the computer industry began to take off in the 1960s and the cost of arithmetic began to fall, we used more arithmetic in applications where it was already an input, such as the US Census Bureau, the US Department of Defense and Nasa. We later began to use the newly cheap arithmetic on problems that weren’t usually arithmetic problems, such as photography. Whereas we once solved photography with chemistry, when arithmetic became cheap enough, we moved to an arithmetic-based solution: digital cameras. A digital image is just a string of zeroes and ones that can be reassembled into a viewable image using arithmetic.

The same goes for prediction. Prediction is being used for traditional tasks, like inventory management and demand forecasting. More significantly, because it is becoming cheaper it is being used for problems that were not traditionally prediction problems. To see a problem and reframe it as a prediction problem is called ‘AI insight’ and engineers all over the world are acquiring it. We are transforming transportation into a prediction problem. Autonomous vehicles have existed for over two decades, though limited to places with detailed floor plans such as factories and warehouses. These floorplans meant that engineers could design their robots to manoeuvre with basic ‘if-then’ logical intelligence: if a person walks in front of the vehicle, then stop. However, those vehicles could not function outside a highly predictable, controlled environment – until engineers reframed navigation as a prediction problem.

Edited extract from ‘Prediction Machines’ by Ajay Agrawal, Joshua Gans and Avi Goldfarb, reproduced with permission

Recent articles

Info Message

Our sites use cookies to support some functionality, and to collect anonymous user data.

Learn more about IET cookies and how to control them

Close