Sean Everett explains how AI image models like DALL·E and Stable Diffusion organize training data using "mutually exclusive and collectively exhaustive" categories. He uses Halloween candy as an analogy to show how sorting Skittles, M&Ms, and Snickers into distinct groups helps systems retrieve and process information efficiently.
Mutually exclusive and collectively exhaustive. What does that mean? High syllable counts for sure, and essentially it's organize and simplify. So let's use an example: Skittles, M&M's, and Snickers bars. Like any child in at least America with their Halloween bag of candy, right? So they get home from that Halloween, they start organizing. So the Skittles, you've got a bunch of different colors. M&M's, you have different colors. Snickers bars, different. Skittles and M&M's are round. One is chocolate, one is more just pure sugar. Snickers is also brown. And so you start to organize by color, by shape, by flavor. And then what you find is you have a bunch of different piles. And maybe on the left is the Skittles, maybe in the middle are the M&M's, maybe on the right is the Snickers bar. So it goes from just like sugar to chocolate, sugar to pure chocolate. And essentially what you've done with that is you've just created your first AI model, right? So you've categorized shapes, colors, and flavor into different buckets. And then when a new flavor, a new piece of candy comes in, you compare it and contrast it to each one of those buckets. So each one of those buckets are mutually exclusive and collectively exhaustive. It has the entire sample from your Halloween candy bag. So when you go to run an AI model or an image model like DALL-E or Stable Diffusion, all it's doing is it's going out to the internet and saying, Skittles, give me a picture of Skittles. M&M's, give me a picture of that. And that's basically it. We'll delve deeper soon.
