Towards neural networks that provably know when they don't know

This repository contains the code that was used to obtain the results reported in https://arxiv.org/abs/1909.12180. In it we propose a Certified Certain Uncertainty (CCU) model with which one can train deep neural networks that provably make low-confidence predictions far away from the training data.

Training the models

Before training a CCU model, one has to first initialize a Gaussian mixture model on the datasets from the in- and out-distribution 80 Million Tiny Images.

python gen_gmm.py --dataset MNIST --PCA 1 --augm_flag 1

The PCA option refers to the fact that we use a modified distance metric.

Most models in the paper are trained via the script in run_training.py. Hyper parameters can be passed as options, but defaults are stored in model_params.py. For example the following lines train a plain model, ACET model and a CCU model on augmented data on MNIST.

python run_training.py --dataset MNIST --train_type plain --augm_flag 1
python run_training.py --dataset MNIST --train_type ACET --augm_flag 1
python run_training.py --dataset MNIST --train_type CEDA_GMM --augm_flag 1 --use_gmm 1 --PCA 1 --fit_out 1

For all commands (except gen_gmm.py) one can specify, which GPU to train on via the --gpu option. All model paths for models that one wishes to use as base models or for testing should be stored in model_paths.py.

The GAN model is trained using the code in https://github.com/alinlab/Confident_classifier and then only imported using the ImportModels notebook. The same goes for OE which we train with https://github.com/hendrycks/outlier-exposure except that we substitute in our architecture, as well as DE.

An ODIN model and single layer Maha model can be generated from the pretrained base model by running in this notebook as well.

MCD and EDL can be trained from scratch in the same notebook.

Testing the models

The out-of-distribution detection statistics (Table 2) are generated by specifying the models one wishes to test in model_paths.py and then running

python gen_eval.py --dataset MNIST --drop_mmc 1

The result of our adversarial noise attack (Table 1) comes from

python gen_attack_stats.py --datasets MNIST --wide_format 1 --fit_out 1

where mutliple datasets could be specified. The gen_attack_stats.py script dumps full information, like all confidences and perturbed samples in results/backup. This path has to be specified when reproducing Figure 2 and 3 with their respective notebooks.

Pre-trained Models

Our pre-trained CCU model weights are available here. Since the gen_eval.py script expects to load models but we are providing model weights, you need to load and save the downloaded weights by running for example

python load_pretrained.py --dataset CIFAR10 --pretrained CIFAR10.pt

This saves a file CIFAR10.pth that you can register in model_paths.py.

Cite us

@article{meinke2020towards,
  title={Towards neural networks that provably know when they don't know},
  author={Meinke, Alexander and Hein, Matthias},
  conference={International Conference on Learning Representations},
  year={2020}
}