Spent 4 hours training an AI model on 500 images and it still confused a cat for a dog

I was building a custom image classifier for a hobby project and thought feeding it 500 labeled photos would be enough. Turns out the lighting conditions were way too varied and the model kept misidentifying objects. Has anyone else run into this where their training dataset needs way more preprocessing than expected?

2 comments

2 Comments

river3208d ago

Honestly, "confused a cat for a dog" is basically my entire life in tech right there. I once spent a whole weekend cleaning up a dataset for a plant identifier and my model ended up labeling every fern as a "small angry bird." I think the lighting issue is a killer, I learned the hard way that 500 pictures of my cat in the dark just trains it to recognize shadows, not actual cat features.

mark_price8d ago

And then you think you've fixed it, so you toss in some data augmentation, flip the images, adjust the brightness, and suddenly your model thinks a blurry plastic bag is a Pomeranian. I swear, neural networks see the world like a drunk person squinting through a dirty window. The whole thing makes me wonder if we're really making it smarter or just lucky when it gets something right. It's like training a goldfish to do your taxes, you're amazed it works at all, then it eats the paper.