Two-Way Is Better Than One: Bidirectional Alignment with Cycle Consistency

Bidirectional projector alignment for exemplar-free class-incremental learning

BiCyc overview

BiCyc studies prototype drift in exemplar-free class-incremental learning. Instead of learning only a one-way post-hoc projection from old features to new features, this project jointly learns two maps, old→new and new→old, with a cycle-consistency objective so that prototype transport and representation learning co-evolve during incremental training.

The key idea is to reduce accumulated transport error between feature spaces. The old→new map transports stored Gaussian prototypes into the current representation space, while the new→old map provides backward alignment. Cycle consistency regularizes the two maps toward near-inverse behavior, helping preserve geometry and reduce forgetting.

Highlights

  • Accepted to ICLR 2026.
  • Introduces bidirectional projector alignment with cycle consistency for EFCIL.
  • Reduces prototype drift and improves old-class retention across CIFAR-100, TinyImageNet, ImageNet-100, and CUB-200.
  • Code: HXuSz11/BiCyc_ICLR2026