Adaptive Distillation via Online Drift Projection

Class-adaptive preservation for exemplar-free class-incremental learning

ODP overview

This project studies heterogeneous representation drift in prototype-based exemplar-free class-incremental learning. Instead of applying uniform preservation strength to all old classes, Online Drift Projection estimates class-wise drift by projecting stored old-class prototypes into the current feature space.

The projected prototype displacement gives an exemplar-free drift signal for each old class. The method converts this signal into drift-weighted distillation, strengthening preservation for high-drift classes while relaxing constraints on stable classes. This class-adaptive view improves the stability-plasticity trade-off under strict exemplar-free constraints.

Highlights

  • Introduces online drift projection for estimating class-wise representation drift.
  • Uses projected prototype displacement as an exemplar-free drift proxy.
  • Applies drift-weighted distillation to focus preservation on high-drift classes.
  • Evaluated on CIFAR-100, TinyImageNet, ImageNet-100, and CUB-200.