in fact in my school (ESIEE), multilevel neuronal network were fashion (Yann Lecun was a reference as ancient from the school). what was limiting was compute power (we were thinking about specialized hardware mimicking life)... Experts systems were more applicable, like natural language processing by deterministic methods (semantic graphs, my colleague worked on that in the 90s until 2k bubble )... with low computation power, testing was hard too, and small networks don't work. hard to get popular this way.
finally I get to think solution of statistical methods like google translate was the future. it came back as a surprise for me, like AI fashion, and strangely I rediscover Yann's name. don't tell me it is new... it is renewed. It remind me Jed booklet on the future, telling LENR will make robotic evolve because much of the engineering will be simplified, helping to focus on AI. 2017-04-17 16:42 GMT+02:00 Jed Rothwell <jedrothw...@gmail.com>: > John Berry <berry.joh...@gmail.com> wrote: > > >> It might have limited application, but mostly, I don't see it, too often >> success and failure is just an inch apart. >> > > Yes! That is an important point. Unfortunately, failure is a more likely > outcome. There are countless way to make an experiment fail, but only a > narrow range of ways to make it work. > > A minor change to an experiment makes it go off the rails and no one > notices. The example I often point to was Shockley's initial refusal to > look at zone refining purification. If he had continued to refuse, I doubt > Bell Labs could have made practical transistors when they did. > > Another recent example is the use of computer neural networks in > artificial intelligence. Going back to the 1950s people had an intuitive > feeling this should work. It resembles actual biological brains, which we > know are capable of intelligence. But little progress was made, and the > approach was ignored or even denigrated during the "AI winter" eras. > Finally, about 10 years ago, the method was revived and greatly improved by > using multi-level networks, where one network feed results into another. > This, finally, produced outstanding results, unlike anything previously > seen. This is the basis for the program that beat one of the world's best > go players, and it is the basis for remarkable recent improvements in > Google translate. See: > > https://blog.google/products/translate/found-translation- > more-accurate-fluent-sentences-google-translate/ > > This progress also came about because computer hardware is so much faster > and cheaper. There are many examples of experiments that failed because > they done before their time. They worked later on after better instruments > were devised. > > - Jed > >
