Gigaom AI Minute – January 29

Transcript

In our desire to build robots with artificial intelligence, we face five big challenges. The first was seeing things and recognizing them. The second was contextualizing them. The third was contextualizing action such as video or how things change. But let's say you did that. Let's say the AI could do all of those things--we're nowhere near that, but let's say just for a moment say we can. We actually aren't that much better off, because we don't know how to teach it to do transfer learning.

If I show you a statue of a lion about a foot tall, and then I show you a dozen other photos and ask you to spot the lion, in them, you wouldn't have any trouble. Even if the lion is half obscured by a tree, or underwater, or upside down, or laying on its side, or has peanut butter smeared all over it. I would wager you've never seen a lion statue with peanut butter smeared all over it, but you would recognize it. This is because human beings can take a lifetime experience of seeing things like objects obscured by water, or covered with substances like peanut butter, and apply that knowledge to a new task. That's transfer learning, and we don't understand how we do it, let alone how to teach a computer to do it.