On cars, old, new and future; science & technology; vintage airplanes, computer flight simulation of them; Sherlockiana; our English language; travel; and other stuff
LONG BEACH, California, about 20 miles up the 405 from where I live, had an international conference this past December on NIPS, Neural Information Processing Systems. The London Review of Books, published 5500 miles from here, had news of NIPS in its January 25, 2018, issue.
David Runciman, who teaches at Cambridge, contributed an LRB “Diary” feature on the NIPS conference, a gathering of some 8000 researchers, many of them specializing in machine learning. Here are some tidbits on things I’ve learned about this technology from his LRB article and a bit more internet searching.
Machine Learning versus Human Learning. For a long time, Artificial Intelligence’s holy grail was emulation of the human brain. One aspect of AI models the brain’s neurons, its specialized cells that transmit nerve impulses. Machine learning, though, proceeds by different means.
Runciman observes, “Trying to get machines to mimic how the human mind works turns out to be a distraction. Machines that are capable of learning do not have to be capable of thinking the way we do.”
“What they require,” Runciman continues, “are vast amounts of data, which they can filter at incredible speed searching for patterns on the basis of which they make inferences.” The machines then seek optimized patterns by comparing one with another.
“In other words,” writes Runciman, “they learn from their mistakes, without stopping to wonder what it really means to learn or to make a mistake. It is amazing how much can be accomplished by machines operating like this. It’s as though they were just waiting to be told that they should try doing it their way, not ours.”
AlphaGo and AlphaZero. Dennis Hassabis was one of the speakers at NIPS. He’s co-founder and CEO of Google’s DeepMind, “best known,” Runciman notes, “for having taught an algorithm (AlphaGo) to play the fearsomely demanding game of Go better than any human being has ever managed.”
AlphaZero, the latest DeepMind project, teaches itself games from scratch; its only human input is providing basic game rules. It accomplishes this by playing games against itself, over and over, learning from its mistakes, and deriving an optimal strategy.
“When AlphaZero was let loose on chess,” Runciman says, “it took just four hours before it could play well enough to defeat the best rival program, Stockfish, which could already see off any human grand master…. When applied to Go—a far more difficult game—the AlphaZero algorithm, as Hassabis put it, acquired in the space of 72 hours the level of knowledge that human civilization had painstakingly accumulated over three thousand years.”
“Then,” Runciman notes, “they left it running, Within weeks it had taken Go to a level that was beyond anything that had been previously imagined. It did the same for chess, effectively turning it into a different game.” The algorithm’s moves made no sense until later they emerged as part of a winning strategy that only AlphaZero could have seen.
After a month, researchers switched off AlphaZero. What else could be gained by letting it gobble up more electricity? Whom else could it play?
Is Machine Learning the Ultimate? Not necessarily. Runciman observes, “… outside the world of perfect information games [like chess or Go] even the smartest machines regularly misidentify objects, misinterpret language, and misunderstand nuance…. As another NIPS speaker pointed out, deep learning machines are still capable of mistaking turtles for rifles (don’t ask me how…).”
Efficiency versus Plasticity. Runciman says, “To the machine, everything is about getting the job done. The danger of conceiving intelligence in these terms is that it makes everything about getting the job done. What we gain in efficiency, we lose in plasticity.”
And maybe we start mistaking turtles for rifles? ds
© Dennis Simanaitis, SimanaitisSays.com, 2018