Back to Home
CIFAR-100 Long-Tail Classification Project

CIFAR-100 Long-Tail Classification Project

H. Choi
Deep Learning Lab Course Project

This project addresses the long-tail classification problem on CIFAR-100 dataset. We built a hierarchical scheme clustering 100 classes into 20 superclasses and applied a two-stage retraining approach with class-balanced sampling and LDAM (Large Margin Cosine Loss) to improve performance on tail classes.

Project Page

Methodology

  • Hierarchical Clustering: 100 classes → 20 superclasses
  • Loss Function: L_class + α·L_superclass
  • Two-stage Training: Class-balanced sampling + LDAM
  • Dataset: CIFAR-100 with long-tail distribution

Key Results

  • Overall Top-1 Accuracy: 48.8% → 75.7% (+26.9%)
  • Tail Classes Performance: 36.6% → 70.3% (+33.7%)
  • Significant improvement in tail class recognition
  • Effective hierarchical learning approach