Combining Models

[00:00:00]
>> Will Sentance: We can actually also combine up models to use our smile detector to maybe even generate new smiles. We create a new model, so we're gonna take our existing, I don't know if I can draw this. We're out of space, unfortunately. I feel bad to mess with them.

[00:00:28]
I will squeeze in here. This is our smile detector. Takes in an image, and if it's random data, there you go, it comes out and says, no smile. Hopefully that's no smile, looks about no smile. Says, no smile. We can build a new model that generates random squiggles like this, squiggle, didn't think you see that, creator.

[00:01:08]
Its first effort went past this model says, no smile. So it goes, let me, and this is a high level explanation, let me try new squiggle. And so it's getting better, this time it was getting up to, the first time 7% next time it got to 13%, okay.

[00:01:30]
After many tries, it gets to generating squiggles at a 90% conversion. Well, now it's a smile generator within the category of images labeled smile and against a high-level explanation. But at its core, this is what a generative adversarial network is doing. It's got one model that's trying to catch out bad squiggles as not a smile and one that's trying to generate squiggles that will convince this at a 90% confidence that what it's produced is a smile.

[00:02:12]
And once it's done so, it's a smile generator. And while we've moved beyond, very, very recently, the last two or three years, generative adversarial networks is the central way in which generation is done towards what's called diffusion models. At their core, it's still combining models up that allows us to achieve these remarkable generative outcomes.

[00:02:36]
Same thing with text-conditional image generation. We have models that associate text with images and pixels and enable us to actually define the content of image we want. And then to combine that model up, which is capturing the meaning of that text, with image generation. Which again, was sometimes done by adversarial networks, now typically, diffusion models.