Daniel Whiteson

How do you train machine learning if you don't know what you're looking for?

Yeah, this is a big question in machine learning and more broadly in artificial intelligence.

It's a whole field called anomaly detection.

How do you find something that's out of the ordinary if you don't know what you're looking for?

Because that's what I want, right?

I want to find the big surprise, the thing that makes us go, what is that even?

And so we have techniques there.

Anomaly detection says, well, let's learn to describe what's expected.

And then we'll think about anything that's different from that.

And so you train machine learning.

You give it a bunch of examples.

You say, here's the kind of thing we're expecting.

Figure out how to think about that so that if we give you something you haven't seen before, you can flag it.

And so what machine learning does is, for example, it takes all the things that you aren't interested in and it learns to transform that into some internal mathematical space and then transform it back.

And it becomes really good at doing that for the kind of things you've been training it for.

And then when something new and weird comes, then that transformation fails.

It's like, well, I don't know how to transform this there and back.

And so it's just an example, but there's lots of ways that you can train machine learning to flag something unusual.

But it's hard, and you never really know if there's something there that you've missed.

It's got to bother you a little bit, right?