Tweeted By @rasbt

Active learning usually focuses on labeling examples with high prediction uncertainty to improve performance. Plot twist: removing the most difficult examples (based on confidence & variability) helped reaching good performance with *fewer* examples https://t.co/4XPqh9DsOf 1/4 pic.twitter.com/Grm9QmP2mz

— Sebastian Raschka (@rasbt) January 6, 2022