Gigaom AI Minute – July 2

Transcript

The Holy Grail of artificial intelligence is being able to point a machine at a big old pile of data and have it figure out everything that's going on and all the relationships between all the different elements with no human involvement at all. We are clearly a long way away from that. Instead we use pretty orderly data sets in order to train our AI.

Computers are pretty good at recognizing handwriting because there's a large corpus of handwritten documents that have been transcribed. Similarly machines are good at translating from one language to another because there are a lot of documents that have been translated already. The recent advances we've made in being able to identify images is also because recently there's been a data set of labeled images on which artificial intelligence can be trained.

Video, however, has presented new challenges, mainly because, although there's an armload of video publicly available on YouTube and Facebook and places like that, it is poorly tagged--especially the action that is going on inside of the video.

That's something that Google DeepMind is starting to tackle. They took a bunch of video, and they put it in Amazon Turk and got people to label 400 different activities that are occurring in those videos, like "playing tennis," and they're using that to train an artificial intelligence. In a way, this is a really interesting problem because every frame of a video is in essence a prediction, a place for a computer to make a prediction about the next frame. And so I think we can expect computers to make big advances in this even though it is seemingly an incredibly difficult area in which to train artificial intelligence.