paper
, code
TL;DR#
- I read this because.. : trick I used in #57
- task : long-tail image classification
- problem : In the real-world, classes are often unbalanced
- idea : logit adjustment based on label frequency
- architecture : ResNet-32, ResNet-50
- objective : Add the values that go into the exponential of softmax plus the frequency per class multiplied by $\tau$.
- baseline : ERM, weight normalisation, Adaptive, Equalized
- data : CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, iNaturalist2018
- evaluation : balanced error(average by class)
- result : outperform baselines
- limitation / things I cannot understand : Reading, but not understanding formulas and logic
Details#
Post-hoc logit adjustment#
Logit adjusted softmax cross-entropy#
Result#
