Combating Catastrophic Forgetting

Developed a novel Weighted Prototype Update for a CIFAR-10 classifier.

  • Developed a novel Weighted Prototype Update for a CIFAR-10 classifier (on BeiT features) that generalizes to same-distribution datasets without labels
  • Prevented catastrophic forgetting, limiting worst-case accuracy to 97.4%
  • Designed a Clustering-based Mean Shift prototype update algorithm for Unsupervised Domain Adaptation
  • Achieved ~5% accuracy improvement on out-of-distribution data while maintaining performance on prior data