‘The Alignment Problem’ Review: When Machines Miss the Point
Sophisticated algorithms can do everything they are supposed to do, performing wonders, and still make bad recommendations and dodgy claims.
Image from article: By Brian Christian
Norton, 476 pages, $27.95
By David A. Shaywitz
Oct. 25, 2020 4:52 pm
In the mid-1990s, a group of software developers applied the latest computer learning to tackle a problem that emergency-room doctors were routinely facing: which of the patients who showed up with pneumonia should be admitted and which could be sent home to recover there? An algorithm analyzed more than 15,000 patients and came up with a series of predictions intended to optimize patient survival. There was, however, an oddity—the computer concluded that asthmatics with pneumonia were low-risk and could be treated as outpatients. The programmers were skeptical.
Their doubts proved correct. As clinicians later explained, when asthmatics show up to an emergency room with pneumonia, they are considered so high-risk that they tend to be triaged immediately to more intensive care. It was this policy that accounted for their lower-than-expected mortality, the outcome that the computer was trying to optimize. The algorithm, in other words, provided the wrong recommendation, but it was doing exactly what it had been programmed to do.
The disconnect between intention and results—between what mathematician Norbert Wiener described as “the purpose put into the machine” and “the purpose we really desire”—defines the essence of “the alignment problem.” Brian Christian, an accomplished technology writer, offers a nuanced and captivating exploration of this white-hot topic, giving us along the way a survey of the state of machine learning and of the challenges it faces.
The alignment problem, Mr. Christian notes, is as old as the earliest attempts to persuade machines to reason, but recent advances in data-capture and computational power have given it a new prominence. To show the limits of even the most sophisticated algorithms, he describes what happened when a vast database of human language was harvested from published books and the internet. It enabled the mathematical analysis of language—facilitating dramatically improved word translations and creating opportunities to express linguistic relationships as simple arithmetical expressions. Type in “King-Man+Woman” and you got “Queen.” But if you tried “Doctor-Man+Woman,” out popped “Nurse.” “Shopkeeper-Man+Woman” produced “Housewife.” Here the math reflected, and risked perpetuating, historical sexism in language use. Another misalignment example: When an algorithm was trained on a data set of millions of labeled images, it was able to sort photos into categories as fine-grained as “Graduation”—yet classified people of color as “Gorillas.” This problem was rooted in deficiencies in the data set on which the model was trained. In both cases, the programmers had failed to recognize, much less seriously consider, the shortcomings of their models.
We are attracted, Mr. Christian observes, to the idea “that society can be made more consistent, more accurate, and more fair by replacing idiosyncratic human judgment with numerical models.” But we may be expecting too much of our software. A computer program intended to guide parole decisions, for example, delivered guidance that distilled and arguably propagated underlying racial inequalities. Is this the algorithm’s fault, or ours?
To answer this question and others, Mr. Christian devotes much of “The Alignment Problem” to the challenges of teaching computers to do what we want them to do. A computer seeking to maximize its score through trial and error, for example, can quickly figure out shoot-’em-up videogames like “Space Invaders” but struggles with Indiana Jones-style adventure games like “Montezuma’s Revenge,” where rewards are sparse and you need to swing across a pit and climb a ladder before you start to score. Human gamers are instinctively driven to explore and figure out what’s behind the next door, but the computer wasn’t—until a “curiosity” incentive was provided.
Imitation is another learning device: A computer may teach itself by observing, say, the video of a highly skilled car driver. Yet even this method can prove tricky, since an expert may never make the mistakes of a beginner. What’s more, an expert may have an intrinsic ability that the learner is unlikely ever to acquire. A skilled driver might easily navigate a cliff-side road, Mr. Christian explains, but a prudent computer-driver might be better off selecting an inland alternative—humility unlikely to be learned by imitation alone.
Mr. Christian notes that computers may one day be able not only to learn our behaviors but also intuit our values—figure out from our actions what it is we’re trying to optimize. This possibility offers the hope of robust cooperative human-machine learning—an area of especially promising research at the moment—but it raises a number of thorny concerns: What if an algorithm intuits the “wrong” values, based on its best read of who we currently are but perhaps not who we aspire to be? Do we, Mr. Christian asks, really want our computers inferring our values from our browser histories?
For all the progress of technology, computers won’t—can’t—solve our most vexing problems. “Machine learning,” Mr. Christian observes, “is an ostensibly technical field crashing increasingly on human questions.” Rather than offering a magic solution, computers provide us with “an unflinching, revelatory mirror.” The image it offers can be discomfiting, but it can also help us by making biases “real” and “measurable” rather than “gossamer, ethereal, ineffable.” At the same time, Mr. Christian reminds us, we must attend to “the things that are not easily quantified or do not easily admit themselves into our models.” He adds that the “ineffable need not cede entirely to the explicit”—a timely reminder that even in our age of big data and deep learning, there will always be more things in heaven and earth than are dreamt of in our algorithms.
Dr. Shaywitz, a physician-scientist, is the founder of Astounding HealthTech, a lecturer in the Department of Biomedical Informatics at Harvard Medical School, and an adjunct scholar at the American Enterprise Institute.
Appeared in the October 26, 2020, print edition as 'When Machines Miss the Point.'
Comments
Post a Comment