The tech craze du jour is machine learning (ML). Billions of dollars of venture capital are being poured into it. All the big tech companies are deep into it. Every computer science student doing a PhD on it is assured of lucrative employment after graduation at his or her pick of technology companies. One of the most popular courses at Stanford is CS229: Machine Learning. Newspapers and magazines extol the wonders of the technology. ML is the magic sauce that enables Amazon to know what you might want to buy next, and Netflix to guess which films might interest you, given your recent viewing history.
To non-geeks, ML is impenetrable, and therefore intimidating. Exotic terminologyabounds: neural networks, backpropagation, random forests, Bayesian networks, quadratic classifiers – that sort of thing. Accordingly, a kind of high priesthood has assembled around the technology which, like all priesthoods, tends to patronise anyone who wonders whether this arcane technology might not be, well, entirely good for humanity. “Don’t you worry about a thing, dear,” is the general tone. “We know what we’re doing.”
When I mentioned ML to a classicist friend of mine recently, he replied: “What, exactly, is the machine learning?” That turns out to be the key question. Machine learning, you see, is best understood as a giant computer-powered sausage-making machine. Into the machine is fed a giant helping of data (called a training set) and, after a bit of algorithmic whirring, out comes the sausage – in the shape of a correlation or a pattern that the algorithm has “learned” from the training set.
The machine is then fed a new datastream, and on the basis of what it has “learned”, proceeds to emit correlations, recommendations and perhaps even judgments (such as: this person is likely to reoffend if granted parole; or that person should be granted a loan). And because these outputs are computer-generated, they are currently regarded with awe and amazement by bemused citizens who are not privy to the aforesaid algorithmic magic.
It’s time to wean ourselves off this servile cringe. A good place to begin would be to start using everyday metaphors for all this exotic gobbledegook.
Cue Maciej Cegłowski, who describes himself as “a painter and computer guy” who lives in San Francisco and maintains one of the most delightful blogs to be found on the web. Last month, Cegłowski was invited to give a talk at the US Library of Congress in which he proposed a novel metaphor. “Machine learning,” he says, “is like a deep-fat fryer. If you’ve never deep-fried something before, you think to yourself: ‘This is amazing! I bet this would work on anything!’ And it kind of does. In our case, the deep fryer is a toolbox of statistical techniques. The names keep changing – it used to be unsupervised learning, now it’s called big data or deep learning or AI. Next year it will be called something else. But the core ideas don’t change. You train a computer on lots of data, and it learns to recognise structure.”
“But,” continues Cegłowski, “the fact that the same generic approach works across a wide range of domains should make you suspicious about how much insight it’s adding. In any deep-frying situation, a good question to ask is: what is this stuff being fried in?”
The cooking oil, in the case of machine learning, is the data used for training. If the data is contaminated – by error, selectivity or bias – so too will be the patterns learned by the software.
And of course, the ML priesthood knows that, so the more conscientious practitioners go to considerable lengths to try to detect and correct for biased results in applications of the technology. For an increasing number of ML applications, though, the training sets are just huge collections of everyday conversations – culled, for example, from social media. That sounds OK: after all, ordinary speech is just that. But a remarkable piece of research by AI researchers at Princeton and the University of Bath reveals that even everyday speech has embedded biases of which most of us are unaware. “Language itself contains recoverable and accurate imprints of our historic biases,” they write, “whether these are morally neutral as towards insects or flowers, problematic as towards race or gender, or even simply veridical, reflecting the status quo for the distribution of gender with respect to careers or first names.” And of course these hidden biases are inevitably captured by machine learning.
I suspect that Wittgenstein would have loved this research: it confirms his belief that the meaning of a word is not to be found in some abstract definition, but in its use in everyday language. Maybe ML geeks should read his Tractatus.