[DL輪読会]Deep Anomaly Detection Using Geometric Transformations

DEEP LEARNING JP [DL Papers] Hirono Okamoto, Matsuo Lab http://deeplearning.jp/ 1

• http://deeplearning.jp/

: Deep Anomaly Detection Using Geometric Transformations n NIPS 2018 accepted n : Izhak Golan, Ran El-Yaniv : n n ( flip ) n n AUROC !" ($) OC-SVM, DAGMM, DSEBM, ADGAN !& ($) !' ($) SOTA !( ($) ( ( ) )

: n 1: ( ) n 2: ( ) 1 (2 )

: n n n One Class SVM

: n (PCA, Robust-PCA, deep autoencoders, ADGAN…) n n One Class SVM L2

: n n n One Class SVM (KDE, Robust-KDE, DSEBM…)

: n n n One Class SVM (SVDD, Deep SVDD...)

: n n n n or n λ

: step1 n k n identity transformation n n x cross-entropy deep k-class 72(=2x3x3x4) !"

: step2 Dirichlet Normality Score n softmax y(x) n α x (Dirichlet ) α

: step1: step2: Dirichlet α

: n n n α k y !" ($) f !" ($)

: n One-Class SVM (OC-SVM) n RAW-OC-SVM n CAE-OC-SVM n Deep One-Class Classification (E2E-OC-SVM) n ICML2018 n Deep structured energy-based models (DSEBM) n ICML 2016 n n Deep Autoencoding Gaussian Mixture Model (DAGMM) n ICLR 2018 n n Anomaly Detection with a Generative Adversarial Network (ADGAN) n AnoGAN(IPMI 2017) n GAN

: n CIFAR-10 n 10 n CIFAR-100 n 100 n 6000 32x32 600 32x32 20 n Fashion-MNIST n 10 7000 28x28 n CatsVsDogs n ASIRRA n 2 12500 360x400 64x64

: n AUROC AUROC(area under an ROC curve) AUROC http://www.randpy.tokyo/entry/roc_auc

• http://www.randpy.tokyo/entry/roc_auc

: CIFAR-10 SOTA

: CIFAR-100 SOTA

: Fashion-MNIST CatsVsDogs SOTA

n n n GAN SOTA n n n ( ?)