Learning Confidence Estimates for Neural Networks

This repository contains the code for the paper Learning Confidence for Out-of-Distribution Detection in Neural Networks. In this work, we demonstrate how to augment neural networks with a confidence estimation branch, which can be used to identify misclassified and out-of-distribution examples.

To learn confidence estimates during training, we provide the neural network with "hints" towards the correct output whenever it exhibits low confidence in its predictions. Hints are provided by pushing the prediction closer to the target distribution via interpolation, where the amount of interpolation proportional to the network's confidence that its prediction is correct. To discourage the network from always asking for free hints, a small penalty is applied whenever it is not confident. As a result, the network learns to only produce low confidence estimates when it is likely to make an incorrect prediction.

Bibtex:

@article{devries2018learning,
  title={Learning Confidence for Out-of-Distribution Detection in Neural Networks},
  author={DeVries, Terrance and Taylor, Graham W},
  journal={arXiv preprint arXiv:1802.04865},
  year={2018}
}

Results and Usage

We evalute our method on the task of out-of-distribution detection using three different neural network architectures: DenseNet, WideResNet, and VGG. CIFAR-10 and SVHN are used as the in-distribution datasets, while TinyImageNet, LSUN, iSUN, uniform noise, and Gaussian noise are used as the out-of-distribution datasets. Definitions of evaluation metrics can be found in the paper.

Dependencies

PyTorch v0.3.0
tqdm
visdom
seaborn
Pillow
scikit-learn

Training

Train a model with a confidence estimator with train.py. During training you can use visdom to see a histogram of confidence estimates from the test set. Training logs will be stored in the logs/ folder, while checkpoints are stored in the checkpoints/ folder.

Use the following settings to replicate the experiments from the paper:

VGG13 on CIFAR-10

python train.py --dataset cifar10 --model vgg13 --budget 0.3 --data_augmentation --cutout 16

WideResNet on CIFAR-10

python train.py --dataset cifar10 --model wideresnet --budget 0.3 --data_augmentation --cutout 16

DenseNet on CIFAR-10

python train.py --dataset cifar10 --model densenet --budget 0.3 --epochs 300 --batch_size 64 --data_augmentation --cutout 16

VGG13 on SVHN

python train.py --dataset svhn --model vgg13 --budget 0.3 --learning_rate 0.01 --epochs 160 --data_augmentation --cutout 20

WideResNet on SVHN

python train.py --dataset svhn --model wideresnet --budget 0.3 --learning_rate 0.01 --epochs 160 --data_augmentation --cutout 20

DenseNet on SVHN

python train.py --dataset svhn --model densenet --budget 0.3 --learning_rate 0.01 --epochs 300 --batch_size 64  --data_augmentation --cutout 20

Out-of-distribution detection

Evaluate a trained model with out_of_distribution_detection.py. Before running this you will need to download the out-of-distribution datasets from Shiyu Liang's ODIN github repo and modify the data paths in the file according to where you saved the datasets.

Example commands for running the out-of-distribution detection script:

Baseline

python out_of_distribution_detection.py --ind_dataset svhn --ood_dataset all --model vgg13 --process baseline --checkpoint svhn_vgg13_budget_0.0_seed_0

ODIN

python out_of_distribution_detection.py --ind_dataset cifar10 --ood_dataset tinyImageNet_resize --model densenet --process ODIN --T 1000 --epsilon 0.001 --checkpoint cifar10_densenet_budget_0.0_seed_0

Confidence

python out_of_distribution_detection.py --ind_dataset cifar10 --ood_dataset LSUN_crop --model vgg13 --process confidence --checkpoint cifar10_vgg13_budget_0.3_seed_0

Confidence scaling

python out_of_distribution_detection.py --ind_dataset svhn --ood_dataset iSUN --model wideresnet --process confidence_scaling --epsilon 0.001 --checkpoint svhn_wideresnet_budget_0.3_seed_0