If you’ve had a cursory eye on the news over the past few years, chances are you’ll have noticed that machine learning and artificial intelligence are getting lots of air time. Whether it’s spooky algorithms employed by your favourite social networking site or fears of a robot rebellion, machine learning has never been more present in the public consciousness. To find out more, Chalkdust attended a lecture by Mike Jordan—undoubtedly the most influential voice on this hottest of hot topics.
Jordan, an award-winning academic who has also advised several machine-learning related companies and developed toolkits that are now the industry standard, is the perfect person to provide a level-headed response to the hype. He has been involved in machine learning for over 30 years and his broad experience gives him authority in both the academic and commercial spheres.
Given the recent explosion in news coverage, you might be forgiven for thinking that machine learning is undergoing something of a revolution. From an academic perspective, this is not quite correct. “Really, the techniques have not dramatically evolved over the last twenty years. The introduction to machine learning class I taught in 1999 is still pretty much the same today.” Instead, Jordan suggests that we are being more creative in our applications, with “every big business in the world employing machine learning on some level”. Previously, these methods had been used in the back-end of businesses, for example by Amazon to reduce their levels of credit card fraud and by shipping companies to predict how many of a product would be needed in a certain part of the world. Now, however, machine learning algorithms are in our homes and our pockets in the form of targeted advertising, personalised recommendations and (soon, according to Jordan) domestic robots.
Although this is an exciting development, as your fridge, television and car all become smarter and more personalised, it isn’t without risks. “If Google messes up and shows you a rubbish search result, it’s frustrating. If machine learning is applied to healthcare and it goes wrong, then lots of people could die.” Like all of mathematics, development of the theory is slow. “People are using these models, and we haven’t got good ways of quantifying the error. We don’t know how well these things scale, and algorithms are being applied in areas that they were never intended for.” Jordan also suggests that journalists, hungry for clickbait, are partly at fault. “I’d love to see people write articles who know about the stuff, and have spent time getting involved in it. But at the same time, I realise people want views and clicks.” (Rising to the challenge, we have given a quick primer on machine learning on page 69.)
So if the pace of change is slow, does that mean we needn’t fear a robot revolution? “Anybody who tells you that we’ll be making machines smarter than we are, doesn’t know what they’re talking about. Not in our lifetime, not for a long time.” Several times during his lecture, Jordan makes the distinction between learning and intelligence. Machine-learning algorithms are simple and dumb, and we have no idea how to achieve the genuine understanding that constitutes intelligence in a human being. As a concrete example, Jordan talks about natural language processing (“the most challenging problem in the field”). “If I make up a word, and say ‘the gretch walked from London to Cambridge in an hour’, you understand lots about this gretch thing. If I add the word ‘gretch’ to a machine learning algorithm, it shifts its knowledge by a miniscule amount.” Thus, flexibility would seem to be the thing missing from machine learning. The ability to improvise, hypothesise and to infer genuine meaning (rather than statistical likelihood) from a sentence are all natural features of human communication that seem impossible from the current perspective of large data sets and finely tuned parameters. Although progress to true intelligence seems like a slow road, Jordan is adamant that machine-learning enabled devices will fill our homes in the near future. “In ten years, I expect most people to own a household robot. It’ll be like having a cellphone, but it will move around.”
It’s important to be prepared for these changes, to teach statistics and computer science together and to make sure that people can retrain as the workplace becomes more automated. “In previous tech revolutions, you had thirty years to adapt. This time it’ll be faster.” Jordan is insistent that these advancements are distributed equitably, regularly travelling abroad to spread the word about machine learning, and developing his software as open source, so that anybody can make use of its incredible power. The transformative effect of machine learning is already starting to be felt in China, a country that Jordan has visited several times. Here, banks have traditionally been hesitant about lending money to individual citizens, and access to cash can often be difficult for those living in rural communities. Now, machine learning methods are being applied by Ant Financial to lend money based on people’s online shopping history. Loans are thus easier to come by, and less risky for the lenders, than ever before. In this case it seems like everybody wins, and that the hype around the impact of ‘dumb algorithms’ is justified after all. Just don’t expect to hold a conversation with your toaster any time soon.
Mike Jordan was speaking as part of the G-Research lecture series, in which world experts in science and technology give accessible talks about cutting-edge research.
More from Chalkdust
- We announce the winner of this coveted prize
- Have your say, and help to choose the 2020 Readers' Choice
- We review the seventh of this year's nominees for the Book of the Year, and open the vote for the readers' favourite
- We review the sixth of this year's nominees for the Book of the Year
- We review the fifth of this year's nominees for the Book of the Year
- We review the fourth of this year's nominees for the Book of the Year