SAMN eliminates hyperparameter sensitivity in long-tailed classification by using a simple monotonicity constraint on class weights, making it easier to deploy and combine with other techniques.
This paper tackles long-tailed recognition (where some classes have far fewer training examples) by proposing SAMN, a method that adjusts classifier weights without requiring hyperparameter tuning. Instead of using regularization tricks, it directly enforces that class weights follow a monotonic pattern, making it simpler and more robust to use in practice.