February 25, 2015


The Robots Are Coming (John Lanchester, 3/05/15, London Review of Books)

Watson's achievement is a sign of how much progress has been made in machine learning, the process by which computer algorithms self-improve at tasks involving analysis and prediction. The techniques involved are primarily statistical: through trial and error the machine learns which answer has the highest probability of being correct. That sounds rough and ready, but because, as per Moore's law, computers have become so astonishingly powerful, the loops of trial and error can take place at great speed, and the machine can quickly improve out of all recognition. The process can be seen at work in Google's translation software. Translate was a page on Google into which you could type text and see it rendered into a short list of other languages. When the software first launched, in 2006, it was an impressive joke: impressive because it existed at all, but a joke because the translations were wildly inaccurate and syntactically garbled. If you gave up on Google Translate at that point, you have missed many changes. The latest version of Translate comes in the form of a smartphone app, into which you can not only type but also speak text, and not just read the answer but also have it spoken aloud. The app can scan text using the phone's camera, and translate that too. For a language you know, and especially with text of any length, Translate is still somewhere between poor and embarrassing - though handy nonetheless, if you momentarily can't remember what the German is for 'collateralised debt obligation' or 'haemorrhoid'. For a language you don't know, it can be invaluable; and it's worth taking a moment to reflect on the marvel that you can install on your phone a device which will translate Malay into Igbo, or Hungarian into Japanese, or indeed anything into anything, for free.

Google Translate hasn't got better because roomfuls of impecunious polymaths have been spending man-years copying out and cross-referencing vocabulary lists. Its improvement is a triumph of machine learning. The software matches texts in parallel languages, so that its learning is a process of finding which text is statistically most likely to match the text in another language. Translate has hoovered up gigantic quantities of parallel texts into its database. A particularly fertile source of these useful things, apparently, is the European Union's set of official publications, which are translated into all Community languages. There was a point a few years ago when the software, after improving for a bit, stopped doing so, as the harvesting of parallel texts began to gather in texts which had already been translated by Translate. I don't know how, but they must have fixed that problem, because it's been getting better again. You could argue that this isn't really 'learning' at all, and indeed it probably isn't in any human sense. The process is analogous, though, in terms of the outcome, if that outcome is defined as getting better at a specific task.

Put all this together, and we can start to see why many people think a big shift is about to come in the impact of computing and technology on our daily lives. Computers have got dramatically more powerful and become so cheap that they are effectively ubiquitous. So have the sensors they use to monitor the physical world. The software they run has improved dramatically too. We are, Brynjolfsson and McAfee argue, on the verge of a new industrial revolution, one which will have as much impact on the world as the first one. Whole categories of work will be transformed by the power of computing, and in particular by the impact of robots.

For many years the problem with robots has been that computers are very good at things we find difficult but very bad at things we find easy. They are brilliant at chess but terrible at the cognitive skills we take for granted, one of the most important being something scientists call SLAM, for 'simultaneous localisation and mapping': the ability to look at a space and see it and know how to move through it, all simultaneously, and with good recall. That, and other skills essential to advanced robotics, is something computers are useless at. A robot chess player can thrash the best chess player in the world, but can't (or couldn't) match the motor and perceptual skills of a one-year-old baby. A famous demonstration of the principle came in 2006, when scientists at Honda staged a public unveiling of their amazing new healthcare robot, the Asimo. Asimo is short (4'3") and white with a black facemask and a metal backpack. It resembles an unusually small astronaut. In the video Asimo advances towards a staircase and starts climbing while turning his face towards the audience as if to say, à la Bender from Futurama, 'check out my shiny metal ass'. He goes up two steps and then falls over. Tittering ensues. It is evident that a new day in robotics has not yet dawned.

That, though, was nine years ago, and Moore's law and machine learning have been at work. The new generation of robots are not ridiculous. Take a look online at the latest generation of Kiva robots employed by Amazon in the 'fulfilment centres' where it makes up and dispatches its parcels. (Though pause first to enjoy the full resonance of 'fulfilment centres'.) The robots are low, slow, accessorised in a friendly orange. They can lift three thousand pounds at a time and carry an entire stack of shelves in one go. Directed wirelessly along preprogrammed paths, they swivel and dance around each other with surprising elegance, then pick up their packages according to the instructions printed on automatically scanned barcodes. They are not alarming, but they are inexorable, and they aren't going away: the labour being done by these robots is work that will never again be done by people. It looks like the future predicted by Wassily Leontief, a Nobel laureate in economics, who said in 1983 that 'the role of humans as the most important factor of production is bound to diminish in the same way that the role of horses in agricultural production was first diminished and then eliminated by the introduction of tractors.'

Large categories of work, especially work that is mechanically precise and repetitive, have already been automated; technologists are working on the other categories, too. Brynjolfsson and McAfee:

Rodney Brooks, who co-founded iRobot, noticed something else about modern, highly automated factory floors: people are scarce, but they're not absent. And a lot of the work they do is repetitive and mindless. On a line that fills up jelly jars, for example, machines squirt a precise amount of jelly into each jar, screw on the top, and stick on the label, but a person places the empty jars on the conveyor belt to start the process. Why hasn't this step been automated? Because in this case the jars are delivered to the line 12 at a time in cardboard boxes that don't hold them firmly in place. This imprecision presents no problem to a person (who simply sees the jars in the box, grabs them, and puts them on the conveyor belt), but traditional industrial automation has great difficulty with jelly jars that don't show up in exactly the same place every time.

It's that problem, and others like it, that many observers think robots are beginning to solve. This isn't just a First World issue. The Taiwanese company Foxconn is the world's largest manufacturer of consumer electronics. If you're reading this on an electronic gadget, there is a good chance that it was made in one of Foxconn's factories, since the firm makes iPhones, iPads, iPods, Kindles, Dell parts, and phones for Nokia and Motorola and Microsoft. It employs about 1.2 million people around the world, many of them in China. At least that's how many it currently employs, but the company's founder, Terry Gou, has spoken of an ambition to buy and deploy a million robots in the company's factories. This is nowhere near happening at the moment, but the very fact that the plan has been outlined makes the point: it isn't only jobs in the rich part of the world that are at risk from robots. The kind of work done in most factories, and anywhere else that requires repetitive manual labour, is going, going, and about to be gone.

And it's not just manual labour. 

Posted by at February 25, 2015 12:54 PM

blog comments powered by Disqus