Just hit 10,000 AI-generated cat pictures in my training set and my model still can't tell a tabby from a tiger

I've been working on this image classifier project for 6 months, and I finally hit 10,000 custom-generated cat images using Stable Diffusion. I thought that number would be a magic bullet. Nope. The other day, it confidently labeled a picture of my orange tabby as 'juvenile Bengal tiger, 92% confidence.' My friend saw the output and just said, 'Bro, your AI is on something.' Has anyone else hit a big dataset milestone only to have their model do something completely ridiculous? What was your 'wait, what?' moment?

2 comments

2 Comments

oliver2421mo ago

Ten thousand pictures and it still thinks your cat is a tiger?

wyattbennett1mo ago

Honestly, that's kind of fair. A cat is basically a tiny, fluffy tiger when you think about it. Same body shape, same predator eyes. The tech is just seeing the family resemblance. Maybe we're the ones who are wrong for calling them different things.

-1