Cleanlab: AI to Find and Fix Errors in ML Datasets

[@sukanyasaha5458]

This is mind blowing!

[@djethereal99]

Great, honest talk

[@gordonlim2322]

8:00 You said it would be easy if the errors were “totally randomly” but you gave an example of when it is “uniformly random” then there would be only one error per class. I’m not convinced that it would be easy to find errors even if the errors were completely random.

[@TomHutchinson5]

I think that the idea is with enough data, if your errors were random, you could still successfully train a model. The model could still figure out, say, how to classify a dog, even if you have an example where a dog was labelled as an orange. However if a big chunk of your data has dogs labelled as wolves, your model won't properly learn what a dog is. I also hear the "totally random" as spread out across your dataset, instead of all clumped up.

[@googm]

what is the difference between totally random and uniformly random

[@billwang1456]

@@googm Hi, should be not much difference in this talk