The Rise of Computer-Aided Explanation | Hacker News Reader

The Rise of Computer-Aided Explanation

quantamagazine.org

36 pointsernesto958 comments

Better than either approach is to take both the objections and the computer-assisted explanations seriously. Then we might ask the following: What qualities do traditional explanations have that aren’t currently shared by computer-assisted explanations? And how can we improve computer-assisted explanations so that they have those qualities?

I am interested in this line of reasoning. Can anyone point me to relevant discussion (preferably scholarly)?

We haven't had the rise of computer-aided explanation yet. The article is more about the problem of not having it. It's needed; "Why did the classifier do that" is starting to become a big problem. Google is having PR problems because their image classifier labeled black people as gorillas.

They can probably use their classifier in feedback mode to generate a canonical image of what the gorilla recognizer is looking for. Publishing that image would create worse PR problems. But at least there's some way to get insight into what's happening. That's been a big problem with ANNs - you get a matrix of values out, but there's no "meaning" associated with it.

>> Chomsky compares the approach to a statistical model of insect behavior. Given enough video of swarming bees, for example, researchers might devise a statistical model that allows them to predict what the bees might do next. But in Chomsky’s opinion it doesn’t impart any true understanding of why the bees dance in the way that they do.

Chomsky is at liberty to pursue an understanding of bee behavior. The engineers and scientists who created the system were interested in translating, not understanding. It seems to me that any system, whether based on a statistical model or some other approach, that achieves a valid translation is clearly a success. For all we know the brains of human translators work similarly.

Indeed. It's easy to intuitively claim that "just" being able to accurately predict the behavior of a system isn't the same thing as having a "true understanding" of it, but are there any observable, tangible differences? I wonder why Chomsky drew the distinction.

One is reminded of Dijkstra's line about how asking whether computers can think is like asking whether submarines can swim.

Claiming there's no difference between prediction and understanding reminds me to what philosophers of science call "instrumentalism": theories are only to be judged by how well they work in predicting facts. But imagine a world in which Newton, instead of formulating the law of universal gravitation and the laws of motion, built a complex machine that, given the relative positions of a set of planets, could work out their subsequent trajectories, with astonishing precision. Nobody would be able to explain why it works in any "deeper" way. Would that world be undistinguishable by ours? If anything, would have Einstein had the same chances of devising general relativity? And Schwarzschild of deriving from it that black holes should exist? Doesn't the fact that we care whether black holes exist, despite how little they admittedly influence the set of mundane phenomena we observe everyday, imply that this world is distinguishably different from that with the accurate, hypothetical Newton black box?

> But imagine a world in which Newton ... built a complex machine that, given the relative positions of a set of planets, could work out their subsequent trajectories..

Thanks for the reply, I see what you're saying. Just to get metaphysical though, can't one argue that that's essentially what Newton did? Granted the machine he made was mathematically elegant, and the study of its components has been fruitful, but nonetheless it's the nature of physics that one never knows whether a formalism is really describing what's going on, or whether it might someday be superseded. Can we not imagine, say, a future where the art of warping spacetime is so well-understood that Newton's theories about a gravitational "force" are seen as just a bit of predictive machinery - useful for making approximations, but bearing no connection to the "truth" underneath?

I guess that that's what happened, and what's happening everyday... Newton action at a distance wasn't convincing enough, so others sought other, "deeper" explanations, among which most notable is of course general relativity. (And it's probably noteworthy that, from an "instrumental" point of view, its benefits were initially marginal...) The difference to me is between saying "the explanatory power of this current theory can (and will eventually) be outdone" and "being able to explain is a dimension of no importance in a scientific theory". Maybe that is Chomsky's concern, even though this instrumentalist attitude is far from new or unknown (think of Copenhagen (lack of) interpretation of quantum mechanics, and its (in)famous motto: "Shut up and calculate"!) If you are interested in this line of reasoning, I found David Deutsch ("The beginning of infinity") and Karl Popper ("Conjectures and Refutations") two passionate (and opinionated) voices!

Thanks for the recommendations. Admittedly I'm just being perverse in applying this to fundamental physical laws but the distinction feels very important in areas like linguistics (where researchers' ideas of "truth" tend not to be expressible in calculus).

That is, I picture an exchange where a deep-learning researcher says "our clustering algorithms found that languages A & B share characteristic C", and a linguistics professor replies "Aha, but the underlying truth is that A & B belong to the X class of languages, and X-like languages always have C". Intuitively the latter version somehow feels more robust, but in practice they're basically equivalent, and the first version arguably carries less baggage.

Reply on news.ycombinator.com