Gigaom AI Minute – July 3

Transcript

Humans are able to do something that machines are nowhere near able to emulate, something that is very critical to being able to advance the science of artificial intelligence. If you show a person a picture of something, a photograph of something, they can actually recognize that object, if shown other photographs, when it's partially obscured, a photograph from a different angle, when the lighting is changed, or when the object is underwater or any number of permutations. Machines can't do anything like this.

Machines actually need an enormous number of samples of things to be able to train on. So how is it that humans are able to, with just a sample size of one, identify all kinds of variants of an object? Well, it's because we have a lifetime of being able to look at other objects in different amounts of light or underwater or under different conditions, and we are able to use those techniques effortlessly in this new environment. This is something called transfer learning, which machines are not able to do very well, but it's super important for getting more advances in areas that we don't have nicely labeled data.