[Pr12] deep anomaly detection using geometric transformations
- 1. Deep Anomaly Detection Using Geometric Transformations Kang, Min-Guk Mingukkang1994@gmail.com March, 17, 2019 1/21 PR-148 https://arxiv.org/abs/1805.10917
- 2. Contents 1. Methodology(Anomaly Detection) 2. Deep Anomaly Detection using Geometric Transformation 3. Experiments 2/21
- 4. 4/21 Anomaly Detection One-class classification Multi-class classification
- 5. 5/21 Anomaly Detection One-class classification Multi-class classification We want to find a dog image between cats.
- 6. 6/21 Anomaly Detection One-class classification Multi-class classification Dog Cat Penguin ?????
- 7. 7/21 Anomaly Detection One-class classification Multi-class classification A baseline for Detecting Misclassified and out-of-distribution examples in neural networks(2017, ICLR) Training Confidence-calibrated Classifier for Detecting Out-of-Distribution Samples(2018, ICLR) A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks(2018, NIPS) A loss framework for calibrated anomaly detection(2018, NIPS) Deep Anomaly Detection with Outlier Exposure(2019, ICLR) … Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery(2017, IPMI) Adversarially Learned One-class Classifier for Novelty Detection(2018, CVPR) Deep One-class Classification(2018, ICML) Deep Autoencoding Gaussian Mixture model for Unsupervised Anomaly Detection(2018, ICLR) Generative Probabilistic Novelty Detection with Adversarial Autoencoder(2018, NIPS) …
- 8. 8/21 Anomaly Detection is widely divided three! 1. Statistical Model(Low density rejection Principle) Neural Networks 𝑍1 𝑍2 Scatter
- 9. 9/21 2. Reconstruction based model Neural Networks 𝑍1 𝑍2 Neural Networks L1, L2 distance Anomaly Detection is widely divided three!
- 10. 10/21 2. Reconstruction based model Neural Networks 𝑍1 𝑍2 Neural Networks L1, L2 distance Anomaly Detection is widely divided three!
- 11. 11/21 3. Kernel based model(DSVDD) Anomaly Detection is widely divided three!
- 12. Deep Anomaly Detection using Geometric Transformations 12/21
- 13. 13/21 Neural Networks Create 72 different geometric transfored images (Translation 9 x Flip 2 x Rotation 4 = 72) … Trans(Images) … Pred … gt 0 1 0 0 0 MSE(Pred, gt)
- 14. 14/21 Neural Networks Create 72 different geometric transfored images (Translation 9 x Flip 2 x Rotation 4 = 72) … … Pred 0.612 0.005 0.001 Simple version: Use the sum of maximum softmax probabilities as an anomaly score!
- 15. 15/21 Neural Networks Create 72 different geometric transfored images (Translation 9 x Flip 2 x Rotation 4 = 72) … … Pred 0.002 0.575 0.102 Simple version: Use the sum of maximum softmax probabilities as an anomaly score!
- 16. 16/21 Neural Networks Create 72 different geometric transfored images (Translation 9 x Flip 2 x Rotation 4 = 72) … … Pred Simple version: Use the sum of maximum softmax probabilities as an anomaly score! 0.002 0.005 0.710
- 17. 17/21 Create 72 different geometric transfored images (Translation 9 x Flip 2 x Rotation 4 = 72) … Softmax probability vectors Complicated version: fit a dirichlet probability distribution to softmax probabilities of training image. And use low density rejection principle!… … … … Summation of column elements is 1.
- 19. 19/21https://www.cs.toronto.edu/~kriz/cifar.html Normal samples(training data) Test samples(test data) :1000 *10 images In this paper, pure version of anomaly detection was performed.
- 20. 20/21
- 21. Thank you! 21/21