18
Took me 3 hours to figure out why my AI model kept mislabeling cats as dogs
Turns out I had 200 blurry training images mixed in with the sharp ones. The model was learning from the noise instead of the actual features. Anyone else run into bad data ruining hours of work?
2 comments
Log in to join the discussion
Log In2 Comments
wren_jackson1mo ago
Did you check if the blurry ones were all from the same source or category? Could be there's a pattern to the bad data that makes the model latch onto it harder.
3
daniel_martinez921mo ago
Totally agree with @wren_jackson on that. I had a similar thing happen with a project where all the bad photos came from one old camera, and the model just learned to spot that camera's quirks instead of the actual subject. It's wild how quickly a model will latch onto a stupid pattern if you don't catch it early. Blurry data is a nightmare because you don't always see it until the results are garbage.
2