📋 Official implementation of Explainable Robust Learning MLNThis repository is official im the following paper:
Elucidating Robust Learning with Uncertainty-Aware Corruption Pattern Estimation
Our contributions are as follows
✔️ We propose a simple yet effective robust learning method leveraging a mixture of experts model on various noise settings.
✔️ The proposed method can not only robustly train from noisy data, but can also provide the explainability by discovering the underlying instance wise noise pattern within the dataset as well the two types of predictive uncertainties(aleatoric and epistemic)
✔️ We present a novel evaluation scheme for validating the set-dependent corruption pattern estimation performance.
Objective
Architecture
torch==1.7.1
torchvision==0.8.2
matplotlib==3.4.1
scikit-learn==0.24.1
gensim==4.0.1
scipy==1.6.2
seaborn==0.11.1
Pillow==8.2.0
Please download mannually TREC dataset
TREC TREC
e.g., mnist on class conditional noise setting
mkdir ckpt
mkdir res
cd scripts
./ccn_mnist.sh
CIFAR10
| Flipping Rate | F-correction | Co-teaching | Co-teaching+ | JoCoR | MLN(ours) |
|---|---|---|---|---|---|
| Symmetry-20% | 68.74±0.20 | 78.23±0.27 | 78.71±0.34 | 85.73±0.19 | 84.20±0.05 |
| Symmetry-50% | 42.71±0.42 | 71.30±0.13 | 57.05±0.54 | 79.41±0.25 | 77.88±0.07 |
| Symmetry-80% | 15.88±0.42 | 26.58±2.22 | 24.19±2.74 | 27.78±3.06 | 41.83±0.10 |
| Asymmetry-40% | 70.60±0.40 | 73.78±0.22 | 68.84±0.20 | 76.36±0.49 | 76.62±0.07 |
Noise Transition Matrix on CIFAR10
aleatoric uncertainty for the ambiguous set is higher than the clean set and larger for more label noise rate.
estimated noise transition matrix for partioned sets are:
If you find this work useful please consider citing it:
@article{papername,
title={title},
author={authors},
journal={arXiv preprint arXiv:xxxx.xxxxx},
year={2021}
}