I have been skeptical of artificial intelligence for a long time, and especially neural networks since they are incomprehensible even to the experts who develop them. Other algorithms have proven to be much better. But then three years ago there was some sort of major breakthrough for which Steve's explanation only confirmed to me that neural networks are still incomprehensible, but researchers seem to have stumbled onto the right recipe. The main example he gave was the ability to recognize images in pictures, which Facebook highlights with their ability to recognize faces in uploaded pictures. The problem is that the machine has to be "trained", or perhaps that should be "brain washed". A set of 1.5 million pictures was used with 22,000 items that had been identified in the pictures by hand. This entire set was fed into the machine learning system, so the machine could correctly associate images with items. Afterwards, it was possible for the machine to recognize images with just over a 95% accuracy, which is about the same as for a human. On still images. For audio, the recognition was about 90% for clear voice and 80% if there is background noise. Not bad, but we are still talking about some fairly heavy duty hardware on two special aspects of human intelligence. The real challenge is machine learning without all the tiger parenting, which Steve hinted at, but for which the results weren't quite as impressive. It is still going to be a while before we can design the Terminator.
Thursday, April 16, 2015
Part of the fun of working on SkyNet is that we get to hear presentations on the latest technology from experts gathered in from around Silicon Valley and even from other parts of the world. Today's lecture was from Steve Oberlin of NVidia on "Deep Learning" using Neural networks. Their graphics hardware is able to process the neural network algorithms some 10X to 100X faster than regular computers, making them a key player as Google, Apple, Facebook, and Baidu all compete to produce the most efficient and accurate algorithms for recognizing images and audio.
Posted by Looney at 8:46 PM