Cold Start Problem in Vision Active Learning

Active learning promises to improve annotation efficiency by iteratively selecting the most important data to be annotated first. However, we uncover a striking contradiction to this promise: active learning fails to select data as efficiently as random selection at the first few choices. We identify this as the cold start problem in vision active learning. We seek to address the cold start problem by exploiting the three advantages of contrastive learning: (1) no annotation is required; (2) label diversity is ensured by pseudo-labels to mitigate bias; (3) typical data is determined by contrastive features to reduce outliers.

Paper

This repository provides the official implementation of the following paper:

Making Your First Choice: To Address Cold Start Problem in Vision Active Learning
Liangyu Chen¹, Yutong Bai², Siyu Huang³, Yongyi Lu², Bihan Wen¹, Alan L. Yuille², and Zongwei Zhou²
¹Nanyang Technological University, ²Johns Hopkins University, ³Harvard University
Medical Imaging with Deep Learning (MIDL), 2023
NeurIPS Workshop on Human in the Loop Learning, 2022
paper | code | poster

If you find this repo useful, please consider citing our paper:

@article{chen2022making,
  title={Making Your First Choice: To Address Cold Start Problem in Vision Active Learning},
  author={Chen, Liangyu and Bai, Yutong and Huang, Siyu and Lu, Yongyi and Wen, Bihan and Yuille, Alan L and Zhou, Zongwei},
  journal={arXiv preprint arXiv:2210.02442},
  year={2022}
}

Installation

The selection part of code is developed on the basis of open-mmlab/mmselfsup. Please see mmselfsup installation.

Dataset preparation

All datasets can be auto downloaded in this repo.

MedMNIST can also be downloaded at MedMNIST v2.

CIFAR-10-LT is generated in this repo with a fixed seed.

Pretrain

⚠️ Note: the current implementation is based on MMSelfSup which is not maintained. I updated the new version of pretraining and selection code in v2, based on MMPreTrain. However, this repository is reserved as a minimal and compatible implementation : )

Pretrain on all MedMNIST datasets

cd selection
bash tools/medmnist_pretrain.sh

Pretrain on CIFAR-10-LT

bash tools/cifar_pretrain.sh

Select initial query

Select initial queries on all MedMNIST datasets

bash tools/medmnist_postprocess.sh

Select initial queries on CIFAR-10-LT

bash tools/cifar_postprocess.sh

Train with selected initial query

cd training
cd medmnist_active_selection # for selecting with actively selected initial queries
# cd medmnist_random_selection # for selecting with randomly selected initial queries
# cd medmnist_uniform_active_selection # for selecting with class-balanced actively selected initial queries
# cd medmnist_uniform_random_selection # for selecting with class-balanced randomly selected initial queries
for mnist in { bloodmnist breastmnist dermamnist octmnist organamnist organcmnist organsmnist pathmnist pneumoniamnist retinamnist tissuemnist }; do bash run.sh $mnist; done

cliangyu / CSVAL