Representational Continuity
for Unsupervised Continual Learning

This is the Pytorch Implementation for the paper Representational Continuity for Unsupervised Continual Learning

Authors: Divyam Madaan, Jaehong Yoon, Yuanchun Li, Yunxin Liu, Sung Ju Hwang

Abstract

Continual learning (CL) aims to learn a sequence of tasks without forgetting the previously acquired knowledge. However, recent advances in continual learning are restricted to supervised continual learning (SCL) scenarios. Consequently, they are not scalable to real-world applications where the data distribution is often biased and unannotated. In this work, we focus on unsupervised continual learning (UCL), where we learn the feature representations on an unlabelled sequence of tasks and show that the reliance on annotated data is not necessary for continual learning. We conduct a systematic study analyzing the learned feature representations and show that unsupervised visual representations are surprisingly more robust to catastrophic forgetting, consistently achieve better performance, and generalize better to out-of-distribution tasks than SCL. Furthermore, we find that UCL achieves a smoother loss landscape through qualitative analysis of the learned representations and learns meaningful feature representations. Additionally, we propose Lifelong Unsupervised Mixup (Lump), a simple yet effective technique that leverages the interpolation between the current task and previous tasks' instances to alleviate catastrophic forgetting for unsupervised representations.

Contribution of this work

We attempt to bridge the gap between continual learning and representation learning and tackle the two important problems of continual learning with unlabelled data and representation learning on a sequence of tasks.
Systematic quantitative analysis show that UCL achieves better performance over SCL with significantly lower catastrophic forgetting on Sequential CIFAR-10, CIFAR-100 and Tiny-ImageNet. Additionally, we evaluate on out of distribution tasks and few-shot continually learning demonstrating the expressive power of unsupervised representations.
We provide visualization of the representations and loss landscapes that UCL learns discriminative, human perceptual patterns and achieves a flatter and smoother loss landscape. Furthermore, we propose Lifelong Unsupervised Mixup (Lump) for UCL, which effectively alleviates catastrophic forgetting and provides better qualitative interpretations.

Prerequisites

$ pip install -r requirements.txt

Run

Split CIFAR-10 experiment with SimSiam

$ python main.py --data_dir ../Data/ --log_dir ../logs/ -c configs/simsiam_c10.yaml --ckpt_dir ./checkpoints/cifar10_results/ --hide_progress

Split CIFAR-100 experiment with SimSiam

$ python main.py --data_dir ../Data/ --log_dir ../logs/ -c configs/simsiam_c100.yaml --ckpt_dir ./checkpoints/cifar100_results/ --hide_progress

Split Tiny-ImageNet experiment with SimSiam

$ python main.py --data_dir ../Data/ --log_dir ../logs/ -c configs/simsiam_tinyimagenet.yaml --ckpt_dir ./checkpoints/tinyimagenet_results/ --hide_progress

Split CIFAR-10 experiment with BarlowTwins

$ python main.py --data_dir ../Data/ --log_dir ../logs/ -c configs/barlow_c10.yaml --ckpt_dir ./checkpoints/cifar10_results/ --hide_progress

Split CIFAR-100 experiment with BarlowTwins

$ python main.py --data_dir ../Data/ --log_dir ../logs/ -c configs/barlowm_c100.yaml --ckpt_dir ./checkpoints/cifar100_results/ --hide_progress

Split Tiny-ImageNet experiment with BarlowTwins

$ python main.py --data_dir ../Data/ --log_dir ../logs/ -c configs/barlowm_tinyimagenet.yaml --ckpt_dir ./checkpoints/tinyimagenet_results/ --hide_progress

Contributing

We'd love to accept your contributions to this project. Please feel free to open an issue, or submit a pull request as necessary. If you have implementations of this repository in other ML frameworks, please reach out so we may highlight them here.

Acknowledgment

The code is build upon aimagelab/mammoth and PatrickHua/SimSiam

Citation

If you found the provided code useful, please cite our work.

@inproceedings{
  madaan2022rethinking,
  title={Representational Continuity for Unsupervised Continual Learning},
  author={Divyam Madaan and Jaehong Yoon and Yuanchun Li and Yunxin Liu and Sung Ju Hwang},
  booktitle={International Conference on Learning Representations},
  year={2022},
  url={https://openreview.net/forum?id=9Hrka5PA7LW}
}

jaehong31/UCL