"Some say they see poetry in my paintings; I see only science." -Georges Seurat



Tuesday, March 5, 2013

Time.com Article on Computer Learning Widely Misses the Mark


A recent article published in the science and technology section of Time.com makes some truly spectacular errors and misunderstandings about artificial intelligence and computer learning. And I thought it would be instructive for people who are interested in these topics to consider why this article is inaccurate, and where and how it misses the mark so badly. The author starts out by claiming that computers are extra stupid:

Just how stupid is your computer? The short answer is that it's really, really stupid. The longer answer is that it's stupider than a slime mold.






The author describes how spectacular it is that a slime mold can make its way through a microscopic maze to get at a food source at the end of the maze. Which I agree, this sounds neat and very cool, but not exactly an incredible example of mental powers. Then he claims this maze-finding behavior is an example of "learning," but really this is an inaccurate description. It's probably an example of homing behaviors, in which an organism navigates or orients based on a simple signal from the environment, the way a moth will fly towards a flame (or the way bugs will fly toward lights at night). This is homing, and it is not cognitively-sophisticated behavior. And even if it were, computers can easily do it, and have been doing so for decades. For instance, the military has been using laser-guided bombs and heat-seeking missiles for years and years, which rely on the same principle (the program says "go towards the laser dot" or "go towards a heat-source"). But according to the author, this is an example that "the slime mold learns -- something your computer will never, ever do."

This is a truly curious claim, first, that slime molds are doing some amazing feats of learning (which his example did not demonstrate), and second, that our computers cannot learn (either at the level of slime molds or beyond). He goes on to suggest that the main technological hurdle holding back computer intelligence is that computers use digital signals to communicate, while neurons' firing intensity or firing rate can vary in a non-digital (analog) fashion. Um, no, sorry buddy.

Many computer systems were once analog, in many ways that's how they started. But they SUCKED so badly as analog systems (programming them to do something useful was a huge hassle) that it was a spectacular breakthrough when they went digital. Analog computers are really fast and efficient at doing a few things, like solving differential equations, just not with the repeatability and accuracy offered by digital computers. "Analog computers of the 1960s were physically large, tedious to program and required significant user expertise. However, once programmed, they rapidly solved a variety of mathematical problems, notably differential equations, without time-domain discretization artifacts, albeit with only moderate accuracy. Analog computers were superseded by digital ones long ago..." (Cowen, 2005)

Going back to analog wouldn't magically solve hosts of artificial intelligence problems, it would just overly complicate things. So it's not clear from the Time article what exactly computer learning has to do with the analog/digital issue.

Plus, neurons aren't strictly analog. Their output certainly is analog, but their inputs are not quite. They have a firing threshold that says "if my inputs don't sum to over some given amount X, then I will not fire." Any given neuron will either fire or not fire within some fixed amount of time based on its inputs and whether its threshold is exceeded. This is a binary decision (fire/don't fire) for each individual neuron. It's probably most accurate to consider a neuronal system as a weird hybrid combination of analog and digital computing units all wired together in complex ways.

In any case, it's my understanding that analog systems can be modeled to any desired degree of fidelity using floating-point operations within digital computers. Both of these considerations help explain why we can accurately model neuronal behaviors inside a digital computer to a high degree of fidelity, without resorting to the use of strictly analog systems. The author's points just make no sense.

How neurons compute (what substrate they are using, or whether its analog or digital or a mix) isn't what makes brains smart. It's the patterns, the conglomerations of neurons and how they are interconnected. No single neuron is smart. Even a small group is pretty dumb, comparable to the slime mold that the author uses as an example of "learning." The real magic comes from hooking up massive amounts of neuronal groups and feeding them into sets of one another, hooking them up in appropriate ways (figuring out how to do this is the real tricky part, but we'll get there). But again, it's all in the patterns. If we can copy or capture how these patterns work or what they are doing and implement them in another substrate (i.e., a digital computer), there is no reason why we could not get these stupid computers to be as smart or smarter than people. In some real sense, computer could become people-like.

Still, modern day computers are not stupid. They aren't human-level geniuses, exactly, but they certainly aren't as dumb as the author portrays them to be. Computers as sophisticated (or less) than your current iPhone or laptop are helping land airplanes, translate human languages, planning flights, answering phones, detecting faces and emotional expressions in photographs, and literally a billion other smart activities, far beyond the capabilities of a measly slime mold. Just a few decades ago, such feats would have seemed near-impossible for computers, but this "AI-creep" has been slow enough that it feels invisible to most observers. And of course, every time AI and computer learning programmers conquer another human-like feat of intelligence (checkers, chess, Go, Poker, Jeopardy, face recognition, optical character recognition, detecting credit card fraud, translating languages, flying airplanes, etc., etc.), the goalposts move and detractors claim that "well that wasn't really an instance of 'true' intelligence anyway"!

I think there really is a potentially-amazing story underlying this article, regarding the transistors that can change based on their inputs ("memristors" or memory-resistors). That technology, or the idea behind it, could be a major leap forward in allowing massively-parallel computers to be wired together (the way a brain is a massively parallel collection of simple processors called neurons), and to adjust the strength of their inter-connections more easily and on-the-fly. Simulating parallel computing on serial processors, the way complex cognitive modeling is often done now, is an inefficient way to go about doing parallel computing. It's much more elegant and simpler to tackle the problem actually using parallel processing, which the memristors mentioned in the article could maybe possibly allow. I think this is the story that the author meant to hit, but missed. Badly.

No comments:

Post a Comment