Resolving Training Biases via Influence-based Data Relabeling

This repository is the official implementation of our ICLR2022 paper: "Resolving Training Biases via Influence-based Data Relabeling"

Paper Citation

@inproceedings{
kong2022resolving,
title={Resolving Training Biases via Influence-based Data Relabeling},
author={Shuming Kong and Yanyan Shen and Linpeng Huang},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=EskfH0bwNVn}
}

Brief Introduction

RDIA aims to reuse harmful training samples toward better model performance. To achieve this, we use influence functions to estimate how relabeling a training sample would affect model’s test performance and further develop a novel relabeling function R to update the labels of identified harmful samples. Our framework can be shown as the image below:

Requirements

To install requirements:

• Python 3.7

pip install -r requirement.txt

Training & Evaluation

Experiments of RDIA

To train the model, run this command:

python RDIA.py --dataset realsim  --noise 0.8 --alpha 0.0002
python RDIA.py --dataset mnist    --noise 0.8 --alpha 0.0002
python RDIA.py --dataset cancer   --noise 0.8 --alpha 0.0002

The hyperparameter \alpha should be tuned with dataset and noise rate. For example, the experiments over MNIST are shown below:

Due to the limitation of Supplementary material, we only uploaded three datasets. All the datasets could be found in here.

Experiments of RDIA-LS

To train the model, run this command:

python RDIA-LS.py --result_dir <path to your dir>  --dataset cifar10 --noise_rate 0.8 --weight_decay 0      --gamma 0.05 --gpu 1

python RDIA-LS.py --result_dir <path to your dir>  --dataset mnist   --noise_rate 0.8 --weight_decay 0.00001 --gamma 0.2  --gpu 1

The hyperparameter \gamma should be tuned with dataset and noise rate. Average test accuracy on MNIST and CIFAR10 with 80% noise rata is :

We also provide the test accuracy vs. number of epochs below:

Results

All the experimental results could be reproduced by the aforementioned code. We have already shown some of the experimental results in the 'Training & Evaluation' section.

Acknowledgement

We would like to thank Zifeng Wang et al. for their fascinating work and public code.