Does a though live in an individual neuron?

Or do all neurons in the human brain participate in all ideas?

And it's been very hard to test in humans because

you can't put a probe on every single neuron in the human brain.

In an artificial neural network, we have this luxury of being able to look at everything that's going on.

[music]

Our project is really asking the question:

"What is a neural network learning inside?"

And we study a specific kind of network called a GAN.

That's a generative adversarial network.

We tell it, "Imagine and image that you haven't seen that looks like these million other images".

[music]

The surprising result of the project is that

neural networks actually show evidence

of composition.

And so the question is, "How the heck is it doing it?"

[music]

If it's just memorizing, then it's approaching things the way we normally program computers to do things, right?

If it's composing, it's sort of a sign that it's thinking in a more human-like way,

that it's understanding the structure of the world.

[music]

But correlation is not the same as causation.

It could be that neuron that correlates with trees

is actually what the neural network is using to think about the color green.

So how do we know the difference?

And just like those individual neurons that correspond to trees or doors,

We found that there are individual neurons that actually correlate with these visible bugs,

with these visible artifacts.

So that was really surprising to us.

Because not only is the network sifting through things, and sorting out things that make sense,

It's also sifting and assigning the things that don't make sense to their own variables as well.

And so it was really surprising to us that we go into a neural network

and do a certain type of brain damage, right?

Basically perform a lobotomy on these twenty neurons,

and instead of doing damage to the network, we actually got the network to perform better.

And so why is it that a network actually has neurons in it that cause problems?

Are mistakes an important part of learning?

It's one of the mysteries that we uncovered. We don't know the answer to that.

But I think that there's more profound reasons to be interested in this

beyond the ancient puzzle of, "How does thinking work?" and, "How do humans work?"

Because we're also using these AI's to build our future world, to build our future societies

and it's important that

we are able to understand, anticipate, and control the world that we create,

and as long as we don't really understand what rules they're applying inside,

we're not going to be able to do that.

And so, I think, I don't know, I think it's the most important thing in the world to study this kind of thing.

[music]