We experimented with MNIST and BDD100K and noticed two recurring issues: about 2\u20134% of samples were noisy or confusing, and there was significant redundancy in the datasets.<\/p>\n
We achieved ~87% accuracy on MNIST with only 10 samples (1 per class), and on BDD, we matched baseline performance with less than ~40% of the dataset after removing obvious redundancies and very low-quality samples.<\/p>\n
This made us wonder why we don\u2019t see more \u201cdataset goldifying\u201d approaches, where datasets are split into something like:<\/p>\n
- \n
- Gold subset (very clean, ~1%)<\/li>\n
- Silver subset (medium, ~5%)<\/li>\n
- Full dataset<\/li>\n<\/ul>\n
Are there any canonical methods or open-source efforts for creating curated gold\/silver subsets of datasets?<\/p>\n<\/div>\n
submitted by \/u\/taranpula39 <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
