Beyond Static Costs: Learning-Dynamics Aware Loss Functions for Long-Tailed Classification

Researchers introduced Learning-Dynamics Aware Loss (LDAL), a new loss for long-tailed image classification.

Unlike static reweighting, LDAL adjusts class weights using feature strength, prediction entropy, and changes across epochs.

The paper reports stronger benchmark performance than existing reweighting losses, with little added training cost.

The approach could help improve minority-class accuracy in real-world vision datasets where class imbalance is common.