Книга: The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake Our World

Sex, lies, and machine learning

Your digital future begins with a realization: every time you interact with a computer-whether it’s your smart phone or a server thousands of miles away-you do so on two levels. The first one is getting what you want there and then: an answer to a question, a product you want to buy, a new credit card. The second level, and in the long run the most important one, is teaching the computer about you. The more you teach it, the better it can serve you-or manipulate you. Life is a game between you and the learners that surround you. You can refuse to play, but then you’ll have to live a twentieth-century life in the twenty-first. Or you can play to win. What model of you do you want the computer to have? And what data can you give it that will produce that model? Those two questions should always be in the back of your mind whenever you interact with a learning algorithm-as they are when you interact with other people. Alice knows that Bob has a mental model of her and seeks to shape it through her behavior. If Bob is her boss, she tries to come across as competent, loyal, and hardworking. If instead Bob is someone she’s trying to seduce, she’ll be at her most seductive. We could hardly function in society without this ability to intuit and respond to what’s on other people’s minds. The novelty in the world today is that computers, not just people, are starting to have theories of mind. Their theories are still primitive, but they’re evolving quickly, and they’re what we have to work with to get what we want-no less than with other people. And so you need a theory of the computer’s mind, and that’s what the Master Algorithm provides, after plugging in the score function (what you think the learner’s goals are, or more precisely its owner’s) and the data (what you think it knows).

Take online dating. When you use Match.com, eHarmony, or OkCupid (suspend your disbelief, if necessary), your goal is simple: to find the best possible date you can. But chances are it will take a lot of work and several disappointing dates before you meet someone you really like. One hardy geek extracted twenty thousand profiles from OkCupid, did his own data mining, found the woman of his dreams on the eighty-eighth date, and told his odyssey to Wired magazine. To succeed with fewer dates and less work, your two main tools are your profile and your responses to suggested matches. One popular option is to lie (about your age, for example). This may seem unethical, not to mention liable to blow up in your face when your date discovers the truth, but there’s a twist. Savvy online daters already know that people lie about their age on their profiles and adjust accordingly, so if you state your true age, you’re effectively telling them you’re older than you really are! In turn, the learner doing the matching thinks people prefer younger dates than they really do. The logical next step is for people to lie about their age by even more, ultimately rendering this attribute meaningless.

A better way for all concerned is to focus on your specific, unusual attributes that are highly predictive of a match, in the sense that they pick out people you like that not everyone else does, and therefore have less competition for. Your job (and your prospective date’s) is to provide these attributes. The matcher’s job is to learn from them, in the same way that an old-fashioned matchmaker would. Compared to a village matchmaker, Match.com’s algorithm has the advantage that it knows vastly more people, but the disadvantage is that it knows them much more superficially. A na?ve learner, such as a perceptron, will be content with broad generalizations like “gentlemen prefer blondes.” A more sophisticated one will find patterns like “people with the same unusual musical tastes are often good matches.” If Alice and Bob both like Beyonc?, that alone hardly singles them out for each other. But if they both like Bishop Allen, that makes them at least a little bit more likely to be potential soul mates. If they’re both fans of a band the learner does not know about, that’s even better, but only a relational algorithm like Alchemy can pick it up. The better the learner, the more it’s worth your time to teach it about you. But as a rule of thumb, you want to differentiate yourself enough so that it won’t confuse you with the “average person” (remember Bob Burns from Chapter 8), but not be so unusual that it can’t fathom you.

Online dating is in fact a tough example because chemistry is hard to predict. Two people who hit it off on a date may wind up falling in love and believing passionately that they were made for each other, but if their initial conversation takes a different turn, they might instead find each other annoying and never want to meet again. What a really sophisticated learner would do is run a thousand Monte Carlo simulations of a date between each pair of plausible matches and rank the matches by the fraction of dates that turned out well. Short of that, dating sites can organize parties and invite people who are each a likely match for many of the others, letting them accomplish in a few hours what would otherwise take weeks.

For those of us who are not keen on online dating, a more immediately useful notion is to choose which interactions to record and where. If you don’t want your Christmas shopping to leave Amazon confused about your tastes, do it on other sites. (Sorry, Amazon.) If you watch different kinds of videos at home and for work, keep two accounts on YouTube, one for each, and YouTube will learn to make the corresponding recommendations. And if you’re about to watch some videos of a kind that you ordinarily have no interest in, log out first. Use Chrome’s incognito mode not for guilty browsing (which you’d never do, of course) but for when you don’t want the current session to influence future personalization. On Netflix, adding profiles for the different people using your account will spare you R-rated recommendations on family movie night. If you don’t like a company, click on their ads: this will not only waste their money now, but teach Google to waste it again in the future by showing the ads to people who are unlikely to buy the products. And if you have very specific queries that you want Google to answer correctly in the future, take a moment to trawl through the later results pages for the relevant links and click on them. More generally, if a system keeps recommending the wrong things to you, try teaching it by finding and clicking on a bunch of the right ones and come back later to see if it did.

That could be a lot of work, though. What all of these illustrate, unfortunately, is how narrow the communication channel between you and the learner is today. You 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. The learner can still decide to ignore you, if it thinks you’re lying or are low on self-knowledge, but at least it would be able to take your input into account. For this, the model needs to be in a form that humans can understand, such as a set of rules rather than a neural network, and it needs to accept general statements as input in addition to raw data, as Alchemy does. All of which brings us to the question of how good a model of you a learner can have and what you’d want to do with that model.

Оглавление книги