Proposed a loss function for unbalanced data. Proposed an entropy correction term that combines the inverse ratio of the samples with a coefficient term that increases the confidence of the correct class relative to the confidence of the incorrect answer for each sample, which in longtail CIFAR-10 outperforms the COT(the previous study).
TL;DR
Proposed a loss function for unbalanced data. Proposed an entropy correction term that combines the inverse ratio of the samples with a coefficient term that increases the confidence of the correct class relative to the confidence of the incorrect answer for each sample, which in longtail CIFAR-10 outperforms the COT(the previous study).
Why it matters:
Paper URL
https://arxiv.org/abs/2009.02189
Submission Dates(yyyy/mm/dd)
Authors and institutions
Yechan Kim, Younkwan Lee, Moongu Jeon
Methods
Results
Comments