Hello redditors,<\/p>\n
The Food-101N<\/a> dataset is a computer vision dataset that is a varient of Food-101 that has extra images and label noise added. I spent some time using an automated data correction platform to really quantify the amount of noise in this dataset. With over 100k examples, manual inspection isn’t an option.<\/p>\n To my surprise, I didn’t just find noise, I also found outliers, ambiguous examples, and duplicates. It was quite an eye-opener seeing thousands of issues that were not included in the “disclaimer” of added label noise by the authors.<\/p>\n Here’s a quick breakdown of what I found:<\/p>\n 27,488 Mislabeled Examples 8,519 Outliers 13,538 Ambiguous Examples 17,510 (Near) Duplicate Examples. <\/p>\n If you’d like to read and see a bit more, you can check out the article<\/a>. There are many visuals that show all of the errors that I wish I could upload here.<\/p>\n * Disclaimer: I am a data scientist for Cleanlab who builds Cleanlab Studio, the automated data correction platform that I used to find these issues.<\/p>\n<\/div>\n submitted by \/u\/cmauck10 <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
