Loading…
20
My buddy who trains models in Seattle told me he labels 10% of his own data wrong on purpose
We were grabbing coffee near Pike Place last Thursday and he said his team adds noisy labels on purpose during fine tuning. I asked why and he said it keeps the model from getting too confident on weird inputs. I tried it on my own small image sorter and my error rate on blurry pics went from 18% down to 9%. Has anyone else messed with adding bad labels on purpose? Curious how much noise you can add before it just makes stuff worse.
1 comments
Log in to join the discussion
Log In1 Comment
clairer7924d ago
Maybe the sweet spot depends on how noisy your blurry pics already are, like matching the label mess to real world mess.
3