By Timothy Kennett.
Pedro Domingos, The Master Algorithm: How the Quest for
the Ultimate Learning Machine Will Remake Our World
(Basic Books, 2015; Penguin, 2015)
It is important to consider, sometimes, how the machines we use every day think of us. Or, more precisely, it is important to consider, sometimes, into what categories (pre-defined or machine-learned or some combination of the two) the machines we use every day place us – because machines do not really think: mostly they sort.
For example, Facebook, like pretty much every other business on the internet, spends a lot of its time categorising its users so that it can target them with ads. Facebook bases this categorisation on the potentially very rich data it has available about every user: network maps of them and all their friends, their conversations, what they like, what they ignore, where they live and where they log into Facebook from, and so on. This data is then sorted through for all kinds of correlations by Facebook’s software so that it can display ads that are, hopefully, relevant to its users’ interests. And, although how exactly Facebook’s software works is secret, it is, you can bet your bottom dollar, largely powered by some kind of machine learning, which is in some senses the cool new kid on the block in the world of artificial intelligence and the related world of making computers do useful stuff. It is also the subject of a new book, The Master Algorithm, by Pedro Domingos, a professor of computer science at the University of Washington (Penguin, 2015). Very basically, machine learning is a method of creating computer algorithms and programs that process data in a way that is probabilistic and self-reinforcing, i.e., the more times they run through the data, the more accurate the outcomes they produce. This is not a process that can be repeated indefinitely to produce outcomes that are ever-closer to perfect; usually it will lead to an outcome that is good enough to make predictions. Or, as Domingos writes, “machine learning is the art of making false assumptions and getting away with it.”
So, back to Facebook, which is a very clear example of just how good enough some of the outcomes of machine learning can be. Here are some things about which Facebook has recently shown me ads: becoming a US military chaplain; going away on a Swedenborgian retreat in Montana (n.b. Emanuel Swedenborg was an eighteenth-century mystic who had a vision of heaven and hell; William Blake liked him); buying new-build luxury condominiums in Berlin (this advert was in English); donating to UN refugees (this ad was in German); renting space for my start-up in London; various small businesses in Cornwall, including a yoga practice, a yacht salesman, an art gallery, and several restaurants; buying flights to Canada; attending a heavy metal festival in Germany; a vaping shop in Berkshire, a steakhouse in St Louis, a caf é in Dubai, and a department store in Copenhagen; a magazine called Primitive Archer; websites selling merchandise for Doctor Who and Curb Your Enthusiasm; learning how to love myself with a website called heartsinharmony.
These demonstrate fairly well a number of machine learning’s potential failings. I get adverts related to areas I am interested in like Berlin or start-ups, but that are wildly inappropriate because they assume I have far more money than I actually do (freelance writers, alas, cannot, in general, afford to buy luxury condos). I get ads related to things like Swedenborg in which I am interested in a vague sense, but on which I do not intend to spend my money; things in which I have an abstract, intellectual interest in rather than a concrete, practical one. I get ads related to things I like (Curb) but which offer things I find repugnant (merchandise). I get various hyper-local ads which aren’t relevant to my interests but are presumably served to me because Facebook, for some inscrutable reason, has decided I live in these places. And, lastly, I get seemingly random ads like the military chaplain thing (I am neither religious nor interested in the military, and, as a UK citizen, very unlikely to join the US Army – although now Facebook mentions it…), which must be based on some not very obvious correlation that has been picked up by Facebook’s machine learners.
Domingos summarises these two problems as “overfitting” and the “curse of dimensionality.” Overfitting, which Domingos calls “the central problem in machine learning,” refers to situations when the learning algorithm detects patterns in the data that are not meaningful: when it mistakes noise for signal. This is something like mistaking correlation for causation, and thus serving someone who is interested in start-ups ads about renting office space, and it is a difficult problem to overcome in any large datasets because they potentially have many, many correlations. The curse of dimensionality is a similar problem: the more dimensions there are in a dataset (i.e., the more different variables you are trying to relate), the more correlations you are going to find. As Domingos explains: “as the number of dimensions goes up, the number of training examples you need to locate the concept’s frontiers goes up exponentially. With twenty Boolean attributes, there are roughly a million different possible examples. With twenty-one, there are two million, and a corresponding number of ways the frontier could wind between them.” This is a problem because it makes the learner’s output more noisy and less accurate and because it makes the computation more and more difficult, and therefore expensive and/or slow, and generally inefficient. Inefficiency should be avoided: “In computer science, a problem isn’t really solved until it’s solved efficiently.”
The field of machine learning, as Domingos documents well throughout his book, has found a number of marvelous ways of fudging around this inefficiency to make machine learning a viable tool for very large or very complex datasets. Very recently, for example, a team at Google announced that they had used machine learning – “a state-of-the-art tree search with two deep neural networks” – to create a system that can beat an elite human at a boardgame called Go, a problem that was thought to be sufficiently hard to solve that five or more years of advancement in machine learning would be needed before it could be done properly. Go is such a hard problem to solve because it has a ridiculous number of moves, as the Google team outline: “The search space in Go is vast – more than a googol times larger than chess (a number greater than there are atoms in the universe!). As a result, traditional ‘brute force’ AI methods – which construct a search tree over all possible sequences of moves – don’t have a chance in Go.” Given that brute force AI methods are basically what computers are good at, the solution to the Go problem is to find clever ways to cut down the domain of possible moves that the system should brute force until it is manageable. In this case, the Google team used two different deep neural networks to cut down the search space as much as possible, first by predicting next moves that are deemed likely to be valuable, and second by giving a value to each position independently, rather than by working out the whole game of moves that follow from it. So, instead of considering all possible moves until the end of the game from each situation, the two neural networks learned from previous games (in a dataset of “30 million moves from games played by human experts”) which moves might be good, and then considered only these. While this cleverness transformed the problem of Go from being one that was unsolvable to one that could be solved very well, it did not make it efficient: “Of course, all of this requires a huge amount of compute power” – Wired reports “170 GPU [Graphical Processing Unit] cards and 1200 standard processors, or CPUs.” The assumption is always that the efficiency will come with time, either with improvements in the function of the machine learning systems, or increases in processing speed, or both.
Google’s research into using machine learning to win 2500-year-old board games is part of a larger project that aspires to create a general learning system, one which can be applied to any number of problems rather than to the one problem for which it was designed. Such generality is, Domingos argues, the holy grail of machine learning. The “master algorithm” of his book’s title refers to such a general learner, which he believes will enable the next great spurt of human progress. (Until we come up against the next great bottleneck: Domingos cautions, wisely, that “[o]vercoming a bottleneck does not mean the sky is the limit; it means the next bottleneck is the limit, even if we don’t see it yet.”) Domingos believes that such a master algorithm is most likely to be a result of a system that combines a number of different kinds of machine learning, preferably all five of the major “tribes” of machine learners he identifies: analogizers, Bayesians, symbolists, connectionists, and evolutionaries. The result would be a system that had general intelligence in at least some of the senses that humans have general intelligence: it would be able to approach a wide range of problems and make some headway toward solving them. Like humans, it would likely be limited by whatever heuristics it was using to cut corners and make problem solving easier, but it would be quite easy to modify or enhance the master algorithm to make it more effective at solving any given problem (unfortunately this is still quite hard to do with human brains). The ability to generalise is particularly important and particularly difficult: at the moment, machine learners of all kinds can become very good at any one of a whole range of different tasks, from creating build orders in the strategy video game Starcraft 2 and translating text from Esperanto into Yaruba to generating music and driving cars, but remain incapable of doing any other tasks. If you tried to drive your Google car using the Google Translate learner, for example, things might not go so well.
Domingos is careful to avoid hyping this general learner too much. People who have been working in artificial intelligence for a long time, as Domingos has, remember the difficulties that hype can cause: Yann LeCun, Facebook’s AI Director, says “I’m very careful not to use words that could lead to hype. Because there is a huge amount of hype in this area. Which is very dangerous.” Both are concerned that hype can lead to unreasonable fears, like the popular concern about a “Singularity” after which independently-minded machines might decide to eradicate human life, more reasonable but exaggerated fears, like that there will be widespread unemployment after robots steal everyone’s jobs, and, most pertinently, disappointment. The Master Algorithm is, among other things, a sort of general history of artificial intelligence research, some of the nadirs of which were the “AI winters” of the 1970s and late 1980s, during which, following extravagant boasting and hype from technologists, researchers, and journalists about what AI was just on the verge of accomplishing – and following AI’s failure really to accomplish anything – funding dried up, companies went bust, and most people in the tech industry started hyping a bunch of other stuff instead. Technology writers who have memories that encompass even the dot-com bubble of 2001, let alone things that happened in the twentieth century, can be hard to find, so, in the current climate of febrile excitement, desperate media boasting, and remarkable growth rates (in user base and market capitalization, if not in revenue), Domingos’s level-headed but optimistic stance is refreshing.
If anything, Domingos undersells the significance of machine learning. The Master Algorithm is mostly a general introduction to what machine learning is, how it works, and how it developed, but in the last chapter Domingos speculates a little about what it might become in the immediate future. His suggestions are surprisingly modest, and almost banal. They are also mostly consumer-facing. He imagines a kind of personal assistant that will automate and optimize a lot of the more tedious internet-admin tasks that consume so much of our time:
Everyone will have a detailed model of him- or herself, and these models will talk to each other all the time. If you’re looking for a job and company X is looking to hire, its model will interview your model. […] your model will go on millions of dates so you don’t have to, and come Saturday, you’ll meet your top prospects at an OkCupid-organized party, knowing that you’re also one of their top prospects.
It will recommend movies, and help you with shopping, banking, insurance, buying gifts, etc., etc., and it looks a lot like the kind of intelligent personal assistant Facebook and others are trying to create. Domingos’s idea is a little more ambitious, though: it’s more like a persistent online representation of you, a simulacrum of your character, preferences, and actions in this limited digital environment constructed out of data points generated by your behaviour and, crucially, owned and controlled by you. Indeed, Domingos argues that “[y]ou should be able to tell it as much as you want about yourself, not just have it learn indirectly from what you do. More than that, you should be able to inspect the learner’s model of you and correct it as desired.”
I am pessimistic about the practicality of this kind of learner, although I think that it’s probably technologically feasible and not all that far from becoming a kind of reality, at least in the more limited Facebook-Messenger-assistant sense. Philosophically, I have problems with the idea that quiddity can emerge from data points, even for something as relatively banal and unimportant as my taste in movies, although this may just be me being fanciful and a little irrational. More importantly, I don’t think this kind of learner could function politically in the current environment, although I hope that I am wrong about this. Domingos touches on the reason why, but with far more optimism than I can muster. He asks: “who should you share data with? That’s perhaps the most important question of the twenty-first century.” And he’s not wrong, but even this phrasing is too optimistic: who should I share my data with, as if I have control of it already, and as if all the other actors on- and offline will respect my wishes. Domingos proposes “a new type of company that is to your data what you bank is to your money” as a necessary institution for a new data-driven world, and the idea is attractive and sounds like it could work, although such an institution is perhaps even more of a pipe-dream than the hypothetical general-learner.
If we are moving towards what Shoshana Zuboff has termed “surveillance capitalism” – an economic model predicated on mining and exploiting vast reams of data about human behaviour in order to modify it – then machine learning is to surveillance capitalists what steam engines were to the factory owners of the Industrial Revolution: a necessity for anyone hoping to make any money out of the new system. Surveillance capitalism, or Big Data capitalism, or Silicon Valley, or whatever other name we want to give the business model represented by Google and Facebook et al., depends on the ability to collect vast quantities of data and, just as crucially, the ability to sort and process this data to do something useful with it. It is not much use to the NSA, for example, to collect everyone’s mobile phone records if some human analysts then have to read them all looking for patterns. The security researcher Bruce Schneier argues that “When we talk about surveillance, we tend to concentrate on the problems of data collection: CCTV cameras, tagged photos, purchasing habits, our writings on sites like Facebook and Twitter. We think much less about data analysis. But effective and pervasive surveillance is just as much about analysis.” And, right now and for the foreseeable future, machine learning in some form will be the backbone of such analysis: without machine learning, Big Data wouldn’t be doing very much of anything other than filling server space.
It is because of this industrial use of machine learning that I am more pessimistic than Domingos about its future use. Right now, neither data nor machine learning algorithms are in the hands of consumers, and there is little incentive to transfer ownership, especially while issues like privacy and data-management remain niche interests rather than political priorities. Indeed, Jaron Lanier and others have emphasized how AI systems can obscure the human labour that goes into them, even though it is essential to their good outcomes: Google Translate could not function without the vast corpus of human-translated text from which it learnt, and Google’s Go AI could not have beaten any human players without the large database of human Go games it had to study. Or, to put it another way, Domingos may have got the whole labour relation the wrong way around: just as much as machine learning will be working for us, we will be working for it.
My biggest cause for optimism is somewhat perverse: perhaps things will get so bad in the current environment that people will start to care, and controlling their data will become the kind of issue about which they might vote. For now, despite its generally upbeat tone, Domingos’s book is a good start. Understanding how machine learning works – its strengths and weaknesses when processing data; its reliance on large, well-manicured datasets; data’s reliance on it to be more than just a large collection of numbers; the kinds of errors it is likely to produce – is a necessary step toward understanding how an internet-enabled data economy functions, how it affects you, and what, if anything, you want to do about it. Domingos ends his book with a plea to his readers: “My fondest wish is that your reaction to this book will be like my reaction to the first AI book I read, over twenty years ago: there’s so much to do here, I don’t know where to start.” He has written The Master Algorithm, he says, because he wants to encourage more people to get into the field of AI research. He adds a final request: “If one day you invent the Master Algorithm, please don’t run to the patent office with it. Open-source it.” This is his tacit acknowledgement of the political problems of the data economy: a master algorithm exclusively owned by a narrow range of already powerful interests is going to bring out all the wrong connotations of the word “master,” and won’t do all that much to benefit those of us whose venture capitalist funding is, how shall I say it, somewhat limited.
ABOUT THE AUTHOR
Timothy Kennett lives and writes in London.
