Finding the gorilla
"A really smart machine will think like an animal," predicted Temple Grandin at last weekend's Singularity Summit. To an animal, she argued, a human on a horse often looks like a very different category of object than a human walking. That seems true; and yet animals also live in a sensory-driven world entirely unlike that of machines.
A day later, Melanie Mitchell, a professor of computer science at Portland State University, argued that analogies are key, she said, to human intelligence, producing landmark insights like comparing a brain to a computer (von Neumann) or evolutionary competition to economic competition (Darwin). This is true, although that initial analogy is often insufficient and may even be entirely wrong. A really significant change in our understanding of the human brain came with research by psychologists like Elizabeth Loftus showing that where computers retain data exactly as it was (barring mechanical corruption), humans improve, embellish, forget, modify, and partially lose stored memories; our memories are malleable and unreliable in the extreme. (For a worked example, see The Good Wife, season 1, episode 6.)
Yet Mitchell is obviously right when she says that much of our humor is based on analogies. It's a staple of modern comedy, for example, for a character to respond on a subject *as if* it were another subject (chocolate as if it were sex, a pencil dropping on Earth as if it were sex, and so on). Especially incongruous analogies: when Watson asks - in the video clip she showed - for the category "Chicks dig me" it's funny because we know that as a machine a) Watson doesn't really understand what it's saying, and b) Watson is pretty much the polar opposite of the kind of thing that "chicks" are generally imagined to "dig".
"You are going to need my kind of mind on some of these Singularity projects," said Grandin, meaning visual thinkers, rather than the mathematical and verbal thinkers who "have taken over". She went on to contend that visual thinkers are better able to see details and relate them to each other. Her example: the emergency generators at Fukushima located below the level of a plaque 30 feet up on the seawall warning that flood water could rise that high. When she talks - passionately - about installing mechanical overrides in the artificial general intelligences Singularitarians hope will be built one day soonish, she seems to be channelling Peter G. Neumann, who talks often about the computer industry's penchant for repeating the security mistakes of decades past.
An interesting sideline about the date of the Singularity: Oxford's Stuart Armstrong has studied these date predictions and concluded pretty much that, in the famed words of William Goldman, no one knows anything. Based on his study of 257 predictions collected by the Singularity Institute and published on its Web site, he concluded that most theories about these predictions are wrong. The dates chosen typically do not correlate with the age or expertise of the predicter or the date of the prediction. I find this fascinating: there's something like an 80 percent consensus that the Singularity will happen in five to 100 years.
Grandin's discussion of visual thinkers made me wonder whether they would be better or worse at spotting the famed invisible gorilla than most people. Spoiler alert: if you're not familiar with this psychologist test, go now and watch the clip before proceeding. You want to say better - after all, spotting visual detail is what visual thinkers excel at - but what if the demands of counting passes is more all-consuming for them than for other types of thinkers? The psychologist Daniel Kahneman, participating by video link, talked about other kinds of bias but not this one. Would visual thinkers be more or less likely to engage in the common human pastime of believing we know something based on too little data and then ignoring new data?
This is, of course, the opposite of today's Bayesian systems, which make a guess and then refine it as more data arrives: almost the exact opposite of the humans Kahneman describes. So many of the developments we're seeing now rely on crunching masses of data (often characterized as "big" but often not *really* all that big) to find subtle patterns that humans never spot. Linda Avey, founder of the personal genome profiling service 23andMe and John Wilbanks are both trying to provide services that will allow individuals to take control of and understand their personal medical data. Avey in particular seems poised to link in somehow to the data generated by seekers in the several-year-old self-quantified movement.
This approach is so far yielding some impressive results. Peter Norvig, the director of research at Google, recounted both the company's work on recognizing cats and its work on building Google Translate. The latter's patchy quality seems more understandable when you learn that it was built by matching documents issued in multiple languages against each other and building up statistical probabilities. The former seems more like magic, although Slate points out that the computers did not necessarily pick out the same patterns humans would.
Well, why should they? Do I pick out the patterns they're interested in? The story continues...
Wendy M. Grossman's Web site has an extensive archive of her books, articles, and music, and an archive of all the earlier columns in this series.