Full report: https://git.io/JGdDM
In this project, we reproduced the main results presented in [1]. Thus, we built from scratch (using PyTorch) the Mahalanobis-based score with Input pre-processing and feature ensembling as suggested in [1]. Also, we re-trained from scratch a ResNet34 neural network [3] on CIFAR-10 [5], SVNH [6] and CIFAR-100 [5] datasets and evaluated the performance of the Mahalanobis-based score on different In-Distribution (ID) /Out-of-Distribution (OOD) setups. We achieve very comparable results to [1]. We have also conducted an ablation study to understand the correlation between validation accuracy of the NN and the performance ofthe mahalanobis score OOD detector. We found that the more well-trained a NN, the best is the performance of the score. We studied the contribution of each layer of the ResNet to the overall performance and concluded that, counter-intuitively, the first layers are not necessary bad predictors but are actually good calibrators to the overall performance. Finally, we extended this score beyond Images-classification tasks by applying it to text data using a pretrained BERT BASE[7] model. In these experiments, we achieved good baseline performances. These experiments could be the ground for a more text-adapted confidence score. Future work could also explore how gradient-based input preprocessing could be implemented in a language model, such as BERT. The table below resumes our results on the setup ID=CIFAR10 and OOD=SVNH with Resnet34 model:
| Model | ROCAUC | AUPRin | AUPROut | Acc | TNR AT 95% TPR |
|---|---|---|---|---|---|
| Mahalanobis Single Layer with input-preprocessing at 0.005 | 98.5 | 96 | 98.8 | 93.8 | 93.1 |
| Mahalanobis Single Layer without input-preprocessing | 94.3 | 92.2 | 93.9 | 87.9 | 56.7 |
| Mahalanobis with feature ensembling and input-preprocessing at 0.005 | 98.9 | 98.3 | 99.3 | 97.6 | 97 |
Our code relies on the PyTorch and HuggingFace libraries.
ResNet_train.ipynb is a notebook with all necessary steps to train a ResNet34 model on either CIFAR-10, CIFAR-100 or SVHN. This notebook uses the class definition in resnet.py
Single_layer_experiments.ipynb is a notebook gathering all single-layer experiments, including appropriate hyperparameter fine-tuning and performance studies.
Feature_ensemble.ipynb is a notebook implementing the feature-ensemble approach to OOD detection. This notebook first pre-computes Mahalanobis scores and saves them to disk, then measures performance.
Ablation_study.ipynb is the implementation and visualization of our main ablation study, which is a sensitivity analysis of OOD detection performance to the base model's validation accuracy.
BERT_experiment.ipynb contains all our text data experiments, using a BERT model. This notebook only relies on scripts from bert_utils.py.
BERT_finetune.ipynb allows to fine-tune a pre-trained BERT model for the downstream IMDB review classification task.
mahalanobis.py gathers all scripts related to Mahalanobis score computing for images with ResNet, including input pre-processing, along with other useful functions.
bert_utils.py gathers all scripts related to Mahalanobis score computing for texts with BERT, including input pre-processing, and other useful code.
resnet.py contains the class definition of the ResNet model we use in the rest of our code.
References:
[1] Kimin Lee et al.A Simple Unified Framework for Detecting Out-of-Distribution Samples and AdversarialAttacks. 2018. arXiv:1807.03888 [stat.ML].
[2] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich.Deep Anomaly Detection with Outlier Exposure.1952019. arXiv:1812.04606 [cs.LG].
[3] Kaiming He et al.Deep Residual Learning for Image Recognition. 2015. arXiv:1512.03385 [cs.CV].
[4] Apoorv Vyas et al. “Out-of-Distribution Detection Using an Ensemble of Self Supervised Leave-outClassifiers”. In: (Sept. 2018).
[5] Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. “CIFAR (Canadian Institute for Advanced Re-200search)”. In: http://www.cs.toronto.edu/~kriz/cifar.html.
[6] Yuval Netzer et al. “Reading Digits in Natural Images with Unsupervised Feature Learning”. In: NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011. 2011. url:http://ufldl.stanford.edu/housenumbers/nips2011_housenumbers.pdf.
[7] Jacob Devlin et al.BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding.2052019. arXiv:1810.04805 [cs.CL].
[8] Andrew L. Maas et al. “Learning Word Vectors for Sentiment Analysis”. In:Proceedings of the 49thAnnual Meeting of the Association for Computational Linguistics: Human Language Technologies. Portland,Oregon, USA: Association for Computational Linguistics, 2011, pp. 142–150.url:http://www.aclweb.org/anthology/P11-1015.210
[9] Xiang Zhang, Junbo Zhao, and Yann LeCun. “Character-Level Convolutional Networks for Text Classifi-cation”. In:arXiv:1509.01626 [cs](Sept. 2015). arXiv:1509.01626 [cs].