---
license: mit
pipeline_tag: image-classification
tags:
- continual-learning
- meta-learning
- computer-vision
---

# MePo: Meta Post-Refinement for Rehearsal-Free General Continual Learning

This repository contains the model checkpoints and artifacts for **MePo** (**Me**ta **Po**st-Refinement), as presented in the paper [MePo: Meta Post-Refinement for Rehearsal-Free General Continual Learning](https://huggingface.co/papers/2602.07940).

## Introduction

**MePo** is an innovative approach for Pretrained Models (PTMs)-based General Continual Learning (GCL). It addresses the challenges of online datastreams and blurry task boundaries by refining the pretrained backbone before downstream continual learning. Inspired by meta-plasticity and reconstructive memory in neuroscience, MePo adapts representation learning to facilitate rapid adaptation to evolving environments in a rehearsal-free manner.

## Resources

- **Paper:** [MePo: Meta Post-Refinement for Rehearsal-Free General Continual Learning](https://huggingface.co/papers/2602.07940)
- **GitHub Repository:** [SunGL001/MePo](https://github.com/SunGL001/MePo)

## Methodology Overview

MePo improves GCL performance through three core components:
1. **Meta-Learning for Representation Learning**: A bi-level meta-learning paradigm that constructs pseudo task sequences from pretraining data to refine the backbone.
2. **Meta Covariance Initialization**: Initializes a meta covariance matrix as the reference geometry of the pretrained representation space, enabling the exploitation of second-order statistics.
3. **Feature Alignment**: Ensures robust output alignment during downstream tasks.

The method serves as a plug-in strategy that achieves significant gains across various benchmarks such as CIFAR-100, ImageNet-R, and CUB-200.

## Citation

If you find this work useful for your research, please cite:

```bibtex
@misc{sun2026mepometapostrefinementrehearsalfree,
      title={MePo: Meta Post-Refinement for Rehearsal-Free General Continual Learnin}, 
      author={Guanglong Sun and Hongwei Yan and Liyuan Wang and Zhiqi Kang and Shuang Cui and Hang Su and Jun Zhu and Yi Zhong},
      year={2026},
      eprint={2602.07940},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2602.07940}, 
}
```