Two-Way Is Better Than One (ICLR 2026)

June 24, 2026 · View on GitHub

This repository contains the official implementation for the ICLR 2026 paper:

Two-Way Is Better Than One: Bidirectional Alignment with Cycle Consistency for Exemplar-Free Class-Incremental Learning
Hongye Xu, Bartosz Krawczyk

Paper: OpenReview | arXiv

Overview

In exemplar-free class-incremental learning (EFCIL), we cannot store past data, so representation drift makes cached class statistics (e.g., prototypes / Gaussians) stale and causes severe forgetting.
We propose bidirectional alignment with cycle consistency during training, jointly learning two lightweight maps:

  • A: old → new (adapter; transports stored old-class statistics into the current feature space),
  • D: new → old (distiller; regularizes the current representation toward the previous backbone),

together with stop-gradient gating and a cycle-consistency loss so that transport and representation co-evolve.

Codebase

Our implementation is based on the FACIL benchmark:

Setup

1) Create conda env + install dependencies

conda create -n yourenv python=3.10 -y
conda activate yourenv

# PyTorch (CUDA 12.6 wheels)
pip install torch==2.7.0 torchvision==0.22.0 torchaudio==2.7.0 \
  --index-url https://download.pytorch.org/whl/cu126

# Core libs
pip install timm==1.0.15 einops==0.8.1 \
  numpy==2.2.5 scipy==1.15.2 pandas==2.2.3 scikit-learn==1.6.1 \
  matplotlib==3.10.0 pillow==11.2.1 tqdm==4.67.1 pyyaml==6.0.2

Notes:

  • For GPU training, ensure your NVIDIA driver supports CUDA 12.6.
  • If you want a minimal dependency set, keep only the “Core libs” and remove the “Optional” block unless required by your run.

Datasets

1) Download datasets

Please download datasets following your preferred convention (cluster/shared storage, etc.).

2) Set dataset root path

Set the dataset root by editing:

  • src/datasets/dataset_config.py
    • modify _BASE_DATA_PATH to your local dataset root directory.

Reproducing Experiments

We provide scripts under scripts/.

TinyImageNet (10 tasks × 20 classes)

bash scripts/tiny-10x20.sh
``$

### \text{CIFAR}-100 (10 \text{tasks}  \times  10 \text{classes})
$``bash
bash scripts/cifar-10x10.sh

Citation

If you find this repository useful, please cite our paper:

@inproceedings{xu2026twoway,
  title     = {Two-Way Is Better Than One: Bidirectional Alignment with Cycle Consistency for Exemplar-Free Class-Incremental Learning},
  author    = {Xu, Hongye and Krawczyk, Bartosz},
  booktitle = {The Fourteenth International Conference on Learning Representations},
  year      = {2026},
  url       = {https://openreview.net/forum?id=7UfZAxKo5K},
  eprint    = {2606.05675},
  archivePrefix = {arXiv}
}

Acknowledgements

This repository is built upon the AdaGauss, FACIL benchmark and related continual-learning tooling. We thank the respective authors for releasing their code.