of ELIZA, the theoretical mathematician Doron Zeilberger now names his computer as a co-author of his papers. He calls it Shalosh B. Ekhad, a play on the Hebrew name of his IBM 3B1.) 258
The AI systems designer Roger Grosse has named two paths to this sort of wired sensibility: “Predictive Learning” and “Representational Learning”.25° That first approach is what Maes’s movie machine pusued. The computer is simply checking what it encounters against a database. It teaches itself to predict based on what has been seen before. This sort of knowledge begins with massive amounts of data and then hunts for patterns, tests their reliability, and improves by mapping quirks and similarities. Google engineers have a device that can gaze into a human eye and spot signs of impending optical failure. Is the machine smarter than your ophthalmologist? Hard to know, but let’s just say this: It has seen, studied and compared millions of eyes to find patterns that nearly perfectly predict a diagnosis. It can review in seconds more cases than your doctor will see in a lifetime - let alone recall and compare at sub-millimeter accuracy. Fast, thorough predictive algorithms make what might once have been regarded as AI disappear. The machine isn’t all that wise; it just knows a lot.
On the other path, the one of “representational learning” the machine uses a self- sketched image of the world, a “representation.” Computers using predictive methods to recognize 10,000 numbers pulled from a database of scrawled hand writing now identify 90 percent of the images. Self-trained machines, however, line up each scanned pixel against a representation of the very idea of writing. They screen millions of pictures with nary a mistake. Faces, disease markers, obscure sounds - all these become scrutable not because the machines have been told what to look for, but because they’ve sort of figured it out themselves. The Al is actually starting to think, much as you or | might, first by building up a picture of the world and then applying it, much as a child might build comprehension of traffic rules just by watching Mom driving every day.
With this representation finished, these nearly alive “thinking” meshes navigate by themselves. You can see already the competition lingering here - who can build the most sensitive model of the world? You? A machine? Even today basic versions of representational Als can study a map and name the most important roads. They can predict cracks in computer networks days before a fault. These programs take longer to train. They are harder to program - and they demand almost unimaginable amounts of computing power - but what emerges is a subtle, lively kind of insight. A machine with a representational understanding of Mozart's 41 symphonies can write you an extremely convincing 424 - or, if you wish, an even earlier First Symphony based on what it knows about his evolution as a composer. It can do it again and again. In seconds. The basic attitude of these researchers behind this technology runs, they confess, like this: Mozart was a fantastic composer. If he wrote
258 He calls it: See Nielsen 259 The AI systems designer: Roger Grosse, “Predictive Learning vs. Representational Learning”, Building Intelligent Probabilistic Systems: 2013
190
HOUSE_OVERSIGHT_018422