SlideShare a Scribd company logo
1 of 30
Download to read offline
Image Anomaly Detection with
Generative Adversarial Networks
Master-Seminar 1: Data Analytics
Presented by:
Sakshi Singh (305238)
Sakshi Singh (Universität Hildesheim)
Seminar 1: Data Analytics
Image Anomaly Detection with Generative Adversarial Networks
OUTLINE
• Abstract
• Introduction
• Related Work
• Motivation & Hypothesis
• Proposed Algorithm
• Experiments & Results
• Conclusions
• References
Seminar 1: Data Analytics
Sakshi Singh (Universität Hildesheim) 2Image Anomaly Detection with Generative Adversarial Networks
Paper Introduction
• Title: Image Anomaly Detection with Generative Adversarial
Networks
• Authors: Lucas Deecke, Robert Vandermeulen, Lukas Ruff,
Stephan Mandt, Marius Kloft
• Publication date: 2018/9/10
• Published by: Springer, Cham
Seminar 1: Data Analytics
Sakshi Singh (Universität Hildesheim) 3Image Anomaly Detection with Generative Adversarial Networks
ABSTRACT
• Detection of anomalies in high dimensional space is difficult
to achieve.
• The authors proposed a novel approach based on deep
learning: Anomaly Detection using Generative Adversarial
Networks (GAN).
• The method is build on searching a good representation of
that sample in the latent space of generator and if not found,
then the sample is considered anomalous.
• State-of-the-art is achieved on standard benchmark image
datasets.
Seminar 1: Data Analytics
Sakshi Singh (Universität Hildesheim) 4Image Anomaly Detection with Generative Adversarial Networks
INTRODUCTION
• Anomaly detection
 Identifying unusual instances
 Application areas: Astronomy, medicines, fault detection,
intrusion detection
• Traditional Algorithms for Anomaly detection: (drawbacks)
 Focus on low dimensional problem
 Require manual feature engineering
Sakshi Singh (Universität Hildesheim) 5Image Anomaly Detection with Generative Adversarial Networks
Seminar 1: Data Analytics
INTRODUCTION
• Deep Learning
 Omits manual feature engineering
 Works well with High Dimensional data like images
 Application performance areas: Image classification, Natural
language processing, Speech recognition.
• State-of-the-art
 Achieved by GAN in high-dimensional generative modelling
 GAN has two neural networks- Generator and Discriminator
 Generator and Discriminator compete with each other
6
Seminar 1: Data Analytics
Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
INTRODUCTION
• Generator
 Maps random samples from low-dimensional space to high-
dimensional space.
 If generator has learnt a good approximation of training data,
then for a new point there already exists a point in the latent
space which closely resembles this point.
• Discriminator
 Learns to separate actual samples from the samples
generated by the Generator.
7
Seminar 1: Data Analytics
Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
GENERATOR Vs DISCRIMINATOR
8
Seminar 1: Data Analytics
Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
Reference[2]
RELATED WORK
1. Anomaly Detection (Approach 1)
 Generative models estimate anomalies through estimation of
the data distribution p.
 In given data, estimate p̂≈p and declare those samples which
are unlikely under p̂ to be anomalous.
 Methods used:
 Kernel density estimation(KDE)
 Gaussians for active learning of anomalies
 hidden Markov models
 dynamic Bayesian networks
Sakshi Singh (Universität Hildesheim) 9Image Anomaly Detection with Generative Adversarial Networks
Seminar 1: Data Analytics
RELATED WORK
2. Deep Generative Models (Approach 2)
 Variational Autoencoders(VAEs) have been proposed as deep
generative models.
 Two approaches followed:
 Optimizing over variational lower bound, parameters are tuned such
that samples resembling data is generated from a Gaussian prior.
 Train pair of deep convolutional in an autoencoder setup and produce
random samples on compression manifold.
 None of the approach estimates p.
 The approach in paper uses deep generative models in
anomaly detection.
10Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
Seminar 1: Data Analytics
MOTIVATION
• Authors addressed the drawbacks of non-parametric anomaly
detection methods:
 Suffer from curse of dimensionality
 Inadequate for analysis and interpretation of high dimensional data
• Authors overcome these limitations by using Deep Learning
for anomaly detection which is:
 End-to-end deep learning approach
 Aims specifically at task of anomaly detection
11
Seminar 1: Data Analytics
Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
STATE-OF-THE-ART
• Achieved by GANs
• GAN provides framework to generate samples approximately
distributed to p.
• GAN attempts to learn the parameterization of neural
network
 Generator(gθ): maps low-dimensional data from noise prior pz (e.g.
multi-variate Gaussian) to samples in image space (including qθ ≈ p)
 Discriminator(dw): learns to classify the data from p and qθ.
• Discriminator becomes better at separating samples from p
and samples from qθ, while the generator fools the
discriminator by adjusting θ.
12
Seminar 1: Data Analytics
Sakshi Singh (Universität Hildesheim)
GAN and WGAN
• Objective function of GAN
where z: vectors lying in latent space of dimensionality d’<<d5;
minmax optimization: lower bound of an f-divergence.
• Objective function of WGAN(Wasserstein GAN)
NOTE: WGAN uses 1-Wasserstein distance(amount of work to pull one density
onto other). WGAN training is more stable(as GAN training has vanishing
gradients in high dimensions) and is used in the experiments.
13
Seminar 1: Data Analytics
Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
HYPOTHESIS
Given a test sample x, if there exists no z such that gθ(z)≈x, or if
such a z is difficult to find, then it can be inferred that x is not
distributed according to p, i.e. it is anomalous.
where,
gθ: the generator
x : new test sample
z : existing point in latent space
p : data distribution through which generative models
detect anomalies
14
Seminar 1: Data Analytics
Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
PROPOSED ALGORITHM: ADGAN
• GAN (Drawback):
 Discriminator perfectly separates real from fake data, but can’t deal
with samples that are unlike training data
• ADGAN (Advantage):
 The discriminator is never used which is discarded after training, so it
is easy to couple ADGAN with any GAN-based approaches like:
 Generator network VAEs
 Moment matching networks
15
Seminar 1: Data Analytics
Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
PROPOSED ALGORITHM: ADGAN
• ADGAN (Anomaly Detection using GAN) is derived from GAN
algorithm.
• ADGAN comes into play after GAN training has converged.
• Basic workflow of algorithm:
 If generator has captured training data distribution with given new
sample x~p, there exists a point z(in latent space) such that: gθ(z)≈p
 Points that are away from p have no representation or occupy small
probability mass in latent distribution.
 If z does not exist then we say x is not distributed over p and it is
anomalous.
16
Seminar 1: Data Analytics
Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
ALGORITHM: PSEUDOCODE
17
Assumption: GAN training has converged
Goal: Given a new test point x, the algorithm searches for a point z in the
latent space, such that gθ(z)≈p
Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
Seminar 1: Data Analytics
ALGORITHM: EXPLAINATION
1. Initialization:
 z0 ~ pz (pz:noise prior from GAN)
 θ reset back to original for each new testing point
 search is initialized from nseed individual points
2. For t=1,2,…,k steps:
 backpropagate the reconstruction loss l(gθ(zt-1),x)
 update parameterization of the generator with a small amount (result:
series of mapping from latent space that closely resembles x)
 Adjust θ to give generator the extra capabilities
18Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
Seminar 1: Data Analytics
ALGORITHM
• The key idea of ADGAN is that if the generator was trained on the same
distribution x was drawn from, then the average over the final set of
reconstruction losses will assume low values and high values otherwise.
19Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
Seminar 1: Data Analytics
EXPERIMENTS
• Datasets
 MNIST: contains grayscale scans of handwritten digits
 CIFAR-10: contains colored images of real-world objects belonging to
10 classes
 LSUN: image datasets showing different scenes(e.g. bedroom, bridges,
conference rooms)
• Default settings
 Training and test splits remain as default
 All images are rescaled to assume pixel values in [-1,1]
Sakshi Singh (Universität Hildesheim) 20Image Anomaly Detection with Generative Adversarial Networks
Seminar 1: Data Analytics
METHODS & HYPERPARAMETERS
• Tested the ADGAN performance against 4 traditional non-
parametric methods of anomaly detection:
1. KDE with a Gaussian kernel
2. One-class support vector machine(OC-SVM) with a Gaussian kernel
3. Isolation Forest(IF)
4. Gaussian mixture model(GMM)
• For all methods above, the authors reduced the feature
dimensionality before performing anomaly detection. Note:
Done via PCA and Alexnet (refer to table 1, slide 26)
• AnoGAN: Similar to ADGAN but they use an additional intermediate
discriminator layer dw’
21
Seminar 1: Data Analytics
Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
METHODS & HYPERPARAMETERS
• Report the performance of two end-to-end deep learning
approaches:
1. VAEs: scored by evaluating the evidence lower bound (ELBO)
2. DCAE: scored according to reconstructional losses, interpreting a
high loss if new samples differ from training samples.
• VAE’s performance is better than thresholding directly via
prior likelihood in latent space.
• Both VAE and DCAE have same architecture as DCGAN, so
these two are paired with DCGAN in the experiments.
22
Seminar 1: Data Analytics
Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
ASSUMPTIONS & RESULTS
EXPERIMENTAL ASSUMPTIONS
 Set dimensionality of latent space d’=256
 Noise prior pz is same used in GAN training
 nseed=64 (to maintain non-convexity of optimization problem)
 Adam optimizer is used for optimizing the latent vectors and
parameters of generator
 Squared L2 loss for measuring reconstruction quality
EXPERIMENTAL RESULTS
 More iterations helped the performance but this gain saturates quickly
 K=5 is a good value(where k: no. of iterations)
23Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
Seminar 1: Data Analytics
ONE Vs ALL CLASSIFICATION
Task 1: Quantify performance of competing methods
 Follow the original publications on OC-SVMs[6]
 Train models on data from single class from MNIST
 Evaluate each model’s performance on 5000 items randomly selected from
test set
 In each trial, label unseen classes in training data as anomalous
24
Fig. 3. ROC curves for one-
versus-all prediction of
competing methods on
MNIST (left) and CIFAR10
(right), averaged over all
classes. KDE and OC-SVM are
shown in conjunction with
PCA, for detailed
performance statistics see
Table 1.
Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
Seminar 1: Data Analytics
Table 1: reports the AUCs
that resulted from leaving
out each individual class
• In these experiments,
1. Average AUC Score according
to GAN discriminator: MNIST
(0.625) and CIFAR-10 (0.513)
2. Average AUC score according
to prior likelihood pz of final
latent vector: MNIST (0.554)
and CIFAR-10 (0.721) {better}
25
Seminar 1: Data Analytics
Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
UNSUPERVISED ANOMALY DETECTION
Task 2: showcase the use of ADGAN in unsupervised detection
 Build a test set from 300 validation sample images
 Train the generator on LSUN scenes (in these experiments: only
bedroom scene)
 Record whether ADGAN assigns low or high anomaly score
 The images associated with highest and lowest anomaly
scores of three different scene categories: Figure 4,5,6 in next
slides
26Sakshi Singh (Universität Hildesheim)
Seminar 1: Data Analytics
Image Anomaly Detection with Generative Adversarial Networks
27Sakshi Singh (Universität Hildesheim)
Seminar 1: Data Analytics
Image Anomaly Detection with Generative Adversarial Networks
CONCLUSION
• ADGAN is able to outperform
traditional methods of detecting
anomalies in high-dimensional
image samples.
• ADGAN has ability to discern
usual from unusual samples:
anomaly detection
• ADGAN is able to incorporate
many properties of an image
(e.g. colors, canonical
geometrics, foreign
objects(caption)).
28
Seminar 1: Data Analytics
Image Anomaly Detection with Generative Adversarial NetworksSakshi Singh (Universität Hildesheim)
FUTURE IMPROVEMENTS
• Performance can be boosted by additional tuning of the
underlying neural network
• To achieve strong experimental results jointly optimize the
latent vectors and generator parameterization
• Improve by initializing from an approximate inversion of the
generator[3,4] and substituting the reconstruction loss for a
more elaborate variant[5]
29
Seminar 1: Data Analytics
Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
REFERENCES
1. Image Anomaly Detection with Generative Adversarial Networks: by
Lucas Deecke, Robert Vandermeulen, Lukas Ruff, Stephan Mandt, and
Marius Kloft
2. https://developers.google.com/machine-learning/gan/gan_structure,
Overview of GAN Structure, accessed on 2.12.2019
3. Donahue, J., Krähenbühl, P., Darrell, T.: Adversarial feature learning.
In:International Conference on Learning Representations (2017)
4. Dumoulin, V., Belghazi, I., Poole, B., Lamb, A., Arjovsky, M., Mastropietro,
O., Courville, A.: Adversarially learned inference. In: International
Conference on Learning Representations (2017)
5. Ling, H., Okada, K.: Diffusion distance for histogram comparison. In:
Computer, Vision and Pattern Recognition. pp. 246–253. IEEE (2006)
6. Schölkopf, B., Platt, J.C., Shawe-Taylor, J., Smola, A.J., Williamson, R.C.:
Estimating the support of a high-dimensional distribution. Tech. Rep.
MSRTR-99-87, Microsoft Research (1999)
30Image Anomaly Detection with Generative Adversarial NetworksSakshi Singh (Universität Hildesheim)
Seminar 1: Data Analytics

More Related Content

What's hot

Variational Autoencoders For Image Generation
Variational Autoencoders For Image GenerationVariational Autoencoders For Image Generation
Variational Autoencoders For Image GenerationJason Anderson
 
GANs and Applications
GANs and ApplicationsGANs and Applications
GANs and ApplicationsHoang Nguyen
 
Basic Generative Adversarial Networks
Basic Generative Adversarial NetworksBasic Generative Adversarial Networks
Basic Generative Adversarial NetworksDong Heon Cho
 
Generative adversarial networks
Generative adversarial networksGenerative adversarial networks
Generative adversarial networksYunjey Choi
 
Introduction to Generative Adversarial Networks (GANs)
Introduction to Generative Adversarial Networks (GANs)Introduction to Generative Adversarial Networks (GANs)
Introduction to Generative Adversarial Networks (GANs)Appsilon Data Science
 
Generative Adversarial Network (GAN)
Generative Adversarial Network (GAN)Generative Adversarial Network (GAN)
Generative Adversarial Network (GAN)Prakhar Rastogi
 
A Short Introduction to Generative Adversarial Networks
A Short Introduction to Generative Adversarial NetworksA Short Introduction to Generative Adversarial Networks
A Short Introduction to Generative Adversarial NetworksJong Wook Kim
 
Generative Adversarial Network (+Laplacian Pyramid GAN)
Generative Adversarial Network (+Laplacian Pyramid GAN)Generative Adversarial Network (+Laplacian Pyramid GAN)
Generative Adversarial Network (+Laplacian Pyramid GAN)NamHyuk Ahn
 
A Style-Based Generator Architecture for Generative Adversarial Networks
A Style-Based Generator Architecture for Generative Adversarial NetworksA Style-Based Generator Architecture for Generative Adversarial Networks
A Style-Based Generator Architecture for Generative Adversarial Networksivaderivader
 
GAN - Theory and Applications
GAN - Theory and ApplicationsGAN - Theory and Applications
GAN - Theory and ApplicationsEmanuele Ghelfi
 
1시간만에 GAN(Generative Adversarial Network) 완전 정복하기
1시간만에 GAN(Generative Adversarial Network) 완전 정복하기1시간만에 GAN(Generative Adversarial Network) 완전 정복하기
1시간만에 GAN(Generative Adversarial Network) 완전 정복하기NAVER Engineering
 
Generative Adversarial Networks and Their Medical Imaging Applications
Generative Adversarial Networks and Their Medical Imaging ApplicationsGenerative Adversarial Networks and Their Medical Imaging Applications
Generative Adversarial Networks and Their Medical Imaging ApplicationsKyuhwan Jung
 
AnoGAN을 이용한 철강 소재 결함 검출 AI
AnoGAN을 이용한 철강 소재 결함 검출 AIAnoGAN을 이용한 철강 소재 결함 검출 AI
AnoGAN을 이용한 철강 소재 결함 검출 AIHYEJINLIM10
 
Image-to-Image Translation with Conditional Adversarial Nets (UPC Reading Group)
Image-to-Image Translation with Conditional Adversarial Nets (UPC Reading Group)Image-to-Image Translation with Conditional Adversarial Nets (UPC Reading Group)
Image-to-Image Translation with Conditional Adversarial Nets (UPC Reading Group)Universitat Politècnica de Catalunya
 
Introduction to Generative Adversarial Networks
Introduction to Generative Adversarial NetworksIntroduction to Generative Adversarial Networks
Introduction to Generative Adversarial NetworksBennoG1
 
Unsupervised learning represenation with DCGAN
Unsupervised learning represenation with DCGANUnsupervised learning represenation with DCGAN
Unsupervised learning represenation with DCGANShyam Krishna Khadka
 
Graph Neural Network - Introduction
Graph Neural Network - IntroductionGraph Neural Network - Introduction
Graph Neural Network - IntroductionJungwon Kim
 
Introduction to XGBoost
Introduction to XGBoostIntroduction to XGBoost
Introduction to XGBoostJoonyoung Yi
 

What's hot (20)

Variational Autoencoders For Image Generation
Variational Autoencoders For Image GenerationVariational Autoencoders For Image Generation
Variational Autoencoders For Image Generation
 
GANs and Applications
GANs and ApplicationsGANs and Applications
GANs and Applications
 
Basic Generative Adversarial Networks
Basic Generative Adversarial NetworksBasic Generative Adversarial Networks
Basic Generative Adversarial Networks
 
Generative adversarial networks
Generative adversarial networksGenerative adversarial networks
Generative adversarial networks
 
Style gan
Style ganStyle gan
Style gan
 
Introduction to Generative Adversarial Networks (GANs)
Introduction to Generative Adversarial Networks (GANs)Introduction to Generative Adversarial Networks (GANs)
Introduction to Generative Adversarial Networks (GANs)
 
Generative Adversarial Network (GAN)
Generative Adversarial Network (GAN)Generative Adversarial Network (GAN)
Generative Adversarial Network (GAN)
 
Review SRGAN
Review SRGANReview SRGAN
Review SRGAN
 
A Short Introduction to Generative Adversarial Networks
A Short Introduction to Generative Adversarial NetworksA Short Introduction to Generative Adversarial Networks
A Short Introduction to Generative Adversarial Networks
 
Generative Adversarial Network (+Laplacian Pyramid GAN)
Generative Adversarial Network (+Laplacian Pyramid GAN)Generative Adversarial Network (+Laplacian Pyramid GAN)
Generative Adversarial Network (+Laplacian Pyramid GAN)
 
A Style-Based Generator Architecture for Generative Adversarial Networks
A Style-Based Generator Architecture for Generative Adversarial NetworksA Style-Based Generator Architecture for Generative Adversarial Networks
A Style-Based Generator Architecture for Generative Adversarial Networks
 
GAN - Theory and Applications
GAN - Theory and ApplicationsGAN - Theory and Applications
GAN - Theory and Applications
 
1시간만에 GAN(Generative Adversarial Network) 완전 정복하기
1시간만에 GAN(Generative Adversarial Network) 완전 정복하기1시간만에 GAN(Generative Adversarial Network) 완전 정복하기
1시간만에 GAN(Generative Adversarial Network) 완전 정복하기
 
Generative Adversarial Networks and Their Medical Imaging Applications
Generative Adversarial Networks and Their Medical Imaging ApplicationsGenerative Adversarial Networks and Their Medical Imaging Applications
Generative Adversarial Networks and Their Medical Imaging Applications
 
AnoGAN을 이용한 철강 소재 결함 검출 AI
AnoGAN을 이용한 철강 소재 결함 검출 AIAnoGAN을 이용한 철강 소재 결함 검출 AI
AnoGAN을 이용한 철강 소재 결함 검출 AI
 
Image-to-Image Translation with Conditional Adversarial Nets (UPC Reading Group)
Image-to-Image Translation with Conditional Adversarial Nets (UPC Reading Group)Image-to-Image Translation with Conditional Adversarial Nets (UPC Reading Group)
Image-to-Image Translation with Conditional Adversarial Nets (UPC Reading Group)
 
Introduction to Generative Adversarial Networks
Introduction to Generative Adversarial NetworksIntroduction to Generative Adversarial Networks
Introduction to Generative Adversarial Networks
 
Unsupervised learning represenation with DCGAN
Unsupervised learning represenation with DCGANUnsupervised learning represenation with DCGAN
Unsupervised learning represenation with DCGAN
 
Graph Neural Network - Introduction
Graph Neural Network - IntroductionGraph Neural Network - Introduction
Graph Neural Network - Introduction
 
Introduction to XGBoost
Introduction to XGBoostIntroduction to XGBoost
Introduction to XGBoost
 

Similar to Image anomaly detection with generative adversarial networks

Machine Learning ICS 273A
Machine Learning ICS 273AMachine Learning ICS 273A
Machine Learning ICS 273Abutest
 
Deep learning: challenges and applications
Deep learning: challenges and  applicationsDeep learning: challenges and  applications
Deep learning: challenges and applicationsAboul Ella Hassanien
 
Pixel Recurrent Neural Networks
Pixel Recurrent Neural NetworksPixel Recurrent Neural Networks
Pixel Recurrent Neural Networksneouyghur
 
Face-GAN project report.pptx
Face-GAN project report.pptxFace-GAN project report.pptx
Face-GAN project report.pptxAndleebFatima16
 
Generative Adversarial Networks and Their Applications in Medical Imaging
Generative Adversarial Networks  and Their Applications in Medical ImagingGenerative Adversarial Networks  and Their Applications in Medical Imaging
Generative Adversarial Networks and Their Applications in Medical ImagingSanghoon Hong
 
Unsupervised Deep Learning (D2L1 Insight@DCU Machine Learning Workshop 2017)
Unsupervised Deep Learning (D2L1 Insight@DCU Machine Learning Workshop 2017)Unsupervised Deep Learning (D2L1 Insight@DCU Machine Learning Workshop 2017)
Unsupervised Deep Learning (D2L1 Insight@DCU Machine Learning Workshop 2017)Universitat Politècnica de Catalunya
 
GNA 13552928 deep learning for GAN a.ppt
GNA 13552928 deep learning for GAN a.pptGNA 13552928 deep learning for GAN a.ppt
GNA 13552928 deep learning for GAN a.pptManiMaran230751
 
Recent improvements to the RDKit
Recent improvements to the RDKitRecent improvements to the RDKit
Recent improvements to the RDKitNextMove Software
 
GAN Deep Learning Approaches to Image Processing Applications (1).pptx
GAN Deep Learning Approaches to Image Processing Applications (1).pptxGAN Deep Learning Approaches to Image Processing Applications (1).pptx
GAN Deep Learning Approaches to Image Processing Applications (1).pptxRMDAcademicCoordinat
 
one shot15729752 Deep Learning for AI and DS
one shot15729752 Deep Learning for AI and DSone shot15729752 Deep Learning for AI and DS
one shot15729752 Deep Learning for AI and DSManiMaran230751
 
最近の研究情勢についていくために - Deep Learningを中心に -
最近の研究情勢についていくために - Deep Learningを中心に - 最近の研究情勢についていくために - Deep Learningを中心に -
最近の研究情勢についていくために - Deep Learningを中心に - Hiroshi Fukui
 
LRP for hand gesture recogntion.pptx
LRP for hand gesture recogntion.pptxLRP for hand gesture recogntion.pptx
LRP for hand gesture recogntion.pptxshamim806425
 
Few shot learning/ one shot learning/ machine learning
Few shot learning/ one shot learning/ machine learningFew shot learning/ one shot learning/ machine learning
Few shot learning/ one shot learning/ machine learningﺁﺻﻒ ﻋﻠﯽ ﻣﯿﺮ
 
Extracting the Object from the Shadows: Maximum Likelihood Object/Shadow Disc...
Extracting the Object from the Shadows: Maximum Likelihood Object/Shadow Disc...Extracting the Object from the Shadows: Maximum Likelihood Object/Shadow Disc...
Extracting the Object from the Shadows: Maximum Likelihood Object/Shadow Disc...Kan Ouivirach, Ph.D.
 
Introduction to conventional machine learning techniques
Introduction to conventional machine learning techniquesIntroduction to conventional machine learning techniques
Introduction to conventional machine learning techniquesXavier Rafael Palou
 
Learning from Simulated and Unsupervised Images through Adversarial Training....
Learning from Simulated and Unsupervised Images through Adversarial Training....Learning from Simulated and Unsupervised Images through Adversarial Training....
Learning from Simulated and Unsupervised Images through Adversarial Training....eraser Juan José Calderón
 
gan-190318135433 (1).pptx
gan-190318135433 (1).pptxgan-190318135433 (1).pptx
gan-190318135433 (1).pptxkiran814572
 
An Extensive Review on Generative Adversarial Networks GAN’s
An Extensive Review on Generative Adversarial Networks GAN’sAn Extensive Review on Generative Adversarial Networks GAN’s
An Extensive Review on Generative Adversarial Networks GAN’sijtsrd
 

Similar to Image anomaly detection with generative adversarial networks (20)

Machine Learning ICS 273A
Machine Learning ICS 273AMachine Learning ICS 273A
Machine Learning ICS 273A
 
Deep learning: challenges and applications
Deep learning: challenges and  applicationsDeep learning: challenges and  applications
Deep learning: challenges and applications
 
Pixel Recurrent Neural Networks
Pixel Recurrent Neural NetworksPixel Recurrent Neural Networks
Pixel Recurrent Neural Networks
 
Face-GAN project report.pptx
Face-GAN project report.pptxFace-GAN project report.pptx
Face-GAN project report.pptx
 
Face-GAN project report
Face-GAN project reportFace-GAN project report
Face-GAN project report
 
Generative Adversarial Networks and Their Applications in Medical Imaging
Generative Adversarial Networks  and Their Applications in Medical ImagingGenerative Adversarial Networks  and Their Applications in Medical Imaging
Generative Adversarial Networks and Their Applications in Medical Imaging
 
Unsupervised Deep Learning (D2L1 Insight@DCU Machine Learning Workshop 2017)
Unsupervised Deep Learning (D2L1 Insight@DCU Machine Learning Workshop 2017)Unsupervised Deep Learning (D2L1 Insight@DCU Machine Learning Workshop 2017)
Unsupervised Deep Learning (D2L1 Insight@DCU Machine Learning Workshop 2017)
 
GNA 13552928 deep learning for GAN a.ppt
GNA 13552928 deep learning for GAN a.pptGNA 13552928 deep learning for GAN a.ppt
GNA 13552928 deep learning for GAN a.ppt
 
Recent improvements to the RDKit
Recent improvements to the RDKitRecent improvements to the RDKit
Recent improvements to the RDKit
 
GAN Deep Learning Approaches to Image Processing Applications (1).pptx
GAN Deep Learning Approaches to Image Processing Applications (1).pptxGAN Deep Learning Approaches to Image Processing Applications (1).pptx
GAN Deep Learning Approaches to Image Processing Applications (1).pptx
 
one shot15729752 Deep Learning for AI and DS
one shot15729752 Deep Learning for AI and DSone shot15729752 Deep Learning for AI and DS
one shot15729752 Deep Learning for AI and DS
 
最近の研究情勢についていくために - Deep Learningを中心に -
最近の研究情勢についていくために - Deep Learningを中心に - 最近の研究情勢についていくために - Deep Learningを中心に -
最近の研究情勢についていくために - Deep Learningを中心に -
 
LRP for hand gesture recogntion.pptx
LRP for hand gesture recogntion.pptxLRP for hand gesture recogntion.pptx
LRP for hand gesture recogntion.pptx
 
[DL輪読会]ClearGrasp
[DL輪読会]ClearGrasp[DL輪読会]ClearGrasp
[DL輪読会]ClearGrasp
 
Few shot learning/ one shot learning/ machine learning
Few shot learning/ one shot learning/ machine learningFew shot learning/ one shot learning/ machine learning
Few shot learning/ one shot learning/ machine learning
 
Extracting the Object from the Shadows: Maximum Likelihood Object/Shadow Disc...
Extracting the Object from the Shadows: Maximum Likelihood Object/Shadow Disc...Extracting the Object from the Shadows: Maximum Likelihood Object/Shadow Disc...
Extracting the Object from the Shadows: Maximum Likelihood Object/Shadow Disc...
 
Introduction to conventional machine learning techniques
Introduction to conventional machine learning techniquesIntroduction to conventional machine learning techniques
Introduction to conventional machine learning techniques
 
Learning from Simulated and Unsupervised Images through Adversarial Training....
Learning from Simulated and Unsupervised Images through Adversarial Training....Learning from Simulated and Unsupervised Images through Adversarial Training....
Learning from Simulated and Unsupervised Images through Adversarial Training....
 
gan-190318135433 (1).pptx
gan-190318135433 (1).pptxgan-190318135433 (1).pptx
gan-190318135433 (1).pptx
 
An Extensive Review on Generative Adversarial Networks GAN’s
An Extensive Review on Generative Adversarial Networks GAN’sAn Extensive Review on Generative Adversarial Networks GAN’s
An Extensive Review on Generative Adversarial Networks GAN’s
 

Recently uploaded

办理(Vancouver毕业证书)加拿大温哥华岛大学毕业证成绩单原版一比一
办理(Vancouver毕业证书)加拿大温哥华岛大学毕业证成绩单原版一比一办理(Vancouver毕业证书)加拿大温哥华岛大学毕业证成绩单原版一比一
办理(Vancouver毕业证书)加拿大温哥华岛大学毕业证成绩单原版一比一F La
 
dokumen.tips_chapter-4-transient-heat-conduction-mehmet-kanoglu.ppt
dokumen.tips_chapter-4-transient-heat-conduction-mehmet-kanoglu.pptdokumen.tips_chapter-4-transient-heat-conduction-mehmet-kanoglu.ppt
dokumen.tips_chapter-4-transient-heat-conduction-mehmet-kanoglu.pptSonatrach
 
Predictive Analysis - Using Insight-informed Data to Determine Factors Drivin...
Predictive Analysis - Using Insight-informed Data to Determine Factors Drivin...Predictive Analysis - Using Insight-informed Data to Determine Factors Drivin...
Predictive Analysis - Using Insight-informed Data to Determine Factors Drivin...ThinkInnovation
 
Beautiful Sapna Vip Call Girls Hauz Khas 9711199012 Call /Whatsapps
Beautiful Sapna Vip  Call Girls Hauz Khas 9711199012 Call /WhatsappsBeautiful Sapna Vip  Call Girls Hauz Khas 9711199012 Call /Whatsapps
Beautiful Sapna Vip Call Girls Hauz Khas 9711199012 Call /Whatsappssapnasaifi408
 
9654467111 Call Girls In Munirka Hotel And Home Service
9654467111 Call Girls In Munirka Hotel And Home Service9654467111 Call Girls In Munirka Hotel And Home Service
9654467111 Call Girls In Munirka Hotel And Home ServiceSapana Sha
 
9711147426✨Call In girls Gurgaon Sector 31. SCO 25 escort service
9711147426✨Call In girls Gurgaon Sector 31. SCO 25 escort service9711147426✨Call In girls Gurgaon Sector 31. SCO 25 escort service
9711147426✨Call In girls Gurgaon Sector 31. SCO 25 escort servicejennyeacort
 
꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call
꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call
꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Callshivangimorya083
 
VIP Call Girls Service Miyapur Hyderabad Call +91-8250192130
VIP Call Girls Service Miyapur Hyderabad Call +91-8250192130VIP Call Girls Service Miyapur Hyderabad Call +91-8250192130
VIP Call Girls Service Miyapur Hyderabad Call +91-8250192130Suhani Kapoor
 
20240419 - Measurecamp Amsterdam - SAM.pdf
20240419 - Measurecamp Amsterdam - SAM.pdf20240419 - Measurecamp Amsterdam - SAM.pdf
20240419 - Measurecamp Amsterdam - SAM.pdfHuman37
 
Call Girls In Mahipalpur O9654467111 Escorts Service
Call Girls In Mahipalpur O9654467111  Escorts ServiceCall Girls In Mahipalpur O9654467111  Escorts Service
Call Girls In Mahipalpur O9654467111 Escorts ServiceSapana Sha
 
GA4 Without Cookies [Measure Camp AMS]
GA4 Without Cookies [Measure Camp AMS]GA4 Without Cookies [Measure Camp AMS]
GA4 Without Cookies [Measure Camp AMS]📊 Markus Baersch
 
Amazon TQM (2) Amazon TQM (2)Amazon TQM (2).pptx
Amazon TQM (2) Amazon TQM (2)Amazon TQM (2).pptxAmazon TQM (2) Amazon TQM (2)Amazon TQM (2).pptx
Amazon TQM (2) Amazon TQM (2)Amazon TQM (2).pptxAbdelrhman abooda
 
VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...
VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...
VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...Suhani Kapoor
 
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一F sss
 
04242024_CCC TUG_Joins and Relationships
04242024_CCC TUG_Joins and Relationships04242024_CCC TUG_Joins and Relationships
04242024_CCC TUG_Joins and Relationshipsccctableauusergroup
 
{Pooja: 9892124323 } Call Girl in Mumbai | Jas Kaur Rate 4500 Free Hotel Del...
{Pooja:  9892124323 } Call Girl in Mumbai | Jas Kaur Rate 4500 Free Hotel Del...{Pooja:  9892124323 } Call Girl in Mumbai | Jas Kaur Rate 4500 Free Hotel Del...
{Pooja: 9892124323 } Call Girl in Mumbai | Jas Kaur Rate 4500 Free Hotel Del...Pooja Nehwal
 
Call Girls In Dwarka 9654467111 Escorts Service
Call Girls In Dwarka 9654467111 Escorts ServiceCall Girls In Dwarka 9654467111 Escorts Service
Call Girls In Dwarka 9654467111 Escorts ServiceSapana Sha
 
Saket, (-DELHI )+91-9654467111-(=)CHEAP Call Girls in Escorts Service Saket C...
Saket, (-DELHI )+91-9654467111-(=)CHEAP Call Girls in Escorts Service Saket C...Saket, (-DELHI )+91-9654467111-(=)CHEAP Call Girls in Escorts Service Saket C...
Saket, (-DELHI )+91-9654467111-(=)CHEAP Call Girls in Escorts Service Saket C...Sapana Sha
 

Recently uploaded (20)

办理(Vancouver毕业证书)加拿大温哥华岛大学毕业证成绩单原版一比一
办理(Vancouver毕业证书)加拿大温哥华岛大学毕业证成绩单原版一比一办理(Vancouver毕业证书)加拿大温哥华岛大学毕业证成绩单原版一比一
办理(Vancouver毕业证书)加拿大温哥华岛大学毕业证成绩单原版一比一
 
Call Girls in Saket 99530🔝 56974 Escort Service
Call Girls in Saket 99530🔝 56974 Escort ServiceCall Girls in Saket 99530🔝 56974 Escort Service
Call Girls in Saket 99530🔝 56974 Escort Service
 
dokumen.tips_chapter-4-transient-heat-conduction-mehmet-kanoglu.ppt
dokumen.tips_chapter-4-transient-heat-conduction-mehmet-kanoglu.pptdokumen.tips_chapter-4-transient-heat-conduction-mehmet-kanoglu.ppt
dokumen.tips_chapter-4-transient-heat-conduction-mehmet-kanoglu.ppt
 
Predictive Analysis - Using Insight-informed Data to Determine Factors Drivin...
Predictive Analysis - Using Insight-informed Data to Determine Factors Drivin...Predictive Analysis - Using Insight-informed Data to Determine Factors Drivin...
Predictive Analysis - Using Insight-informed Data to Determine Factors Drivin...
 
Beautiful Sapna Vip Call Girls Hauz Khas 9711199012 Call /Whatsapps
Beautiful Sapna Vip  Call Girls Hauz Khas 9711199012 Call /WhatsappsBeautiful Sapna Vip  Call Girls Hauz Khas 9711199012 Call /Whatsapps
Beautiful Sapna Vip Call Girls Hauz Khas 9711199012 Call /Whatsapps
 
9654467111 Call Girls In Munirka Hotel And Home Service
9654467111 Call Girls In Munirka Hotel And Home Service9654467111 Call Girls In Munirka Hotel And Home Service
9654467111 Call Girls In Munirka Hotel And Home Service
 
9711147426✨Call In girls Gurgaon Sector 31. SCO 25 escort service
9711147426✨Call In girls Gurgaon Sector 31. SCO 25 escort service9711147426✨Call In girls Gurgaon Sector 31. SCO 25 escort service
9711147426✨Call In girls Gurgaon Sector 31. SCO 25 escort service
 
꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call
꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call
꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call
 
VIP Call Girls Service Miyapur Hyderabad Call +91-8250192130
VIP Call Girls Service Miyapur Hyderabad Call +91-8250192130VIP Call Girls Service Miyapur Hyderabad Call +91-8250192130
VIP Call Girls Service Miyapur Hyderabad Call +91-8250192130
 
20240419 - Measurecamp Amsterdam - SAM.pdf
20240419 - Measurecamp Amsterdam - SAM.pdf20240419 - Measurecamp Amsterdam - SAM.pdf
20240419 - Measurecamp Amsterdam - SAM.pdf
 
Call Girls In Mahipalpur O9654467111 Escorts Service
Call Girls In Mahipalpur O9654467111  Escorts ServiceCall Girls In Mahipalpur O9654467111  Escorts Service
Call Girls In Mahipalpur O9654467111 Escorts Service
 
GA4 Without Cookies [Measure Camp AMS]
GA4 Without Cookies [Measure Camp AMS]GA4 Without Cookies [Measure Camp AMS]
GA4 Without Cookies [Measure Camp AMS]
 
Amazon TQM (2) Amazon TQM (2)Amazon TQM (2).pptx
Amazon TQM (2) Amazon TQM (2)Amazon TQM (2).pptxAmazon TQM (2) Amazon TQM (2)Amazon TQM (2).pptx
Amazon TQM (2) Amazon TQM (2)Amazon TQM (2).pptx
 
VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...
VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...
VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...
 
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
 
04242024_CCC TUG_Joins and Relationships
04242024_CCC TUG_Joins and Relationships04242024_CCC TUG_Joins and Relationships
04242024_CCC TUG_Joins and Relationships
 
Decoding Loan Approval: Predictive Modeling in Action
Decoding Loan Approval: Predictive Modeling in ActionDecoding Loan Approval: Predictive Modeling in Action
Decoding Loan Approval: Predictive Modeling in Action
 
{Pooja: 9892124323 } Call Girl in Mumbai | Jas Kaur Rate 4500 Free Hotel Del...
{Pooja:  9892124323 } Call Girl in Mumbai | Jas Kaur Rate 4500 Free Hotel Del...{Pooja:  9892124323 } Call Girl in Mumbai | Jas Kaur Rate 4500 Free Hotel Del...
{Pooja: 9892124323 } Call Girl in Mumbai | Jas Kaur Rate 4500 Free Hotel Del...
 
Call Girls In Dwarka 9654467111 Escorts Service
Call Girls In Dwarka 9654467111 Escorts ServiceCall Girls In Dwarka 9654467111 Escorts Service
Call Girls In Dwarka 9654467111 Escorts Service
 
Saket, (-DELHI )+91-9654467111-(=)CHEAP Call Girls in Escorts Service Saket C...
Saket, (-DELHI )+91-9654467111-(=)CHEAP Call Girls in Escorts Service Saket C...Saket, (-DELHI )+91-9654467111-(=)CHEAP Call Girls in Escorts Service Saket C...
Saket, (-DELHI )+91-9654467111-(=)CHEAP Call Girls in Escorts Service Saket C...
 

Image anomaly detection with generative adversarial networks

  • 1. Image Anomaly Detection with Generative Adversarial Networks Master-Seminar 1: Data Analytics Presented by: Sakshi Singh (305238) Sakshi Singh (Universität Hildesheim) Seminar 1: Data Analytics Image Anomaly Detection with Generative Adversarial Networks
  • 2. OUTLINE • Abstract • Introduction • Related Work • Motivation & Hypothesis • Proposed Algorithm • Experiments & Results • Conclusions • References Seminar 1: Data Analytics Sakshi Singh (Universität Hildesheim) 2Image Anomaly Detection with Generative Adversarial Networks
  • 3. Paper Introduction • Title: Image Anomaly Detection with Generative Adversarial Networks • Authors: Lucas Deecke, Robert Vandermeulen, Lukas Ruff, Stephan Mandt, Marius Kloft • Publication date: 2018/9/10 • Published by: Springer, Cham Seminar 1: Data Analytics Sakshi Singh (Universität Hildesheim) 3Image Anomaly Detection with Generative Adversarial Networks
  • 4. ABSTRACT • Detection of anomalies in high dimensional space is difficult to achieve. • The authors proposed a novel approach based on deep learning: Anomaly Detection using Generative Adversarial Networks (GAN). • The method is build on searching a good representation of that sample in the latent space of generator and if not found, then the sample is considered anomalous. • State-of-the-art is achieved on standard benchmark image datasets. Seminar 1: Data Analytics Sakshi Singh (Universität Hildesheim) 4Image Anomaly Detection with Generative Adversarial Networks
  • 5. INTRODUCTION • Anomaly detection  Identifying unusual instances  Application areas: Astronomy, medicines, fault detection, intrusion detection • Traditional Algorithms for Anomaly detection: (drawbacks)  Focus on low dimensional problem  Require manual feature engineering Sakshi Singh (Universität Hildesheim) 5Image Anomaly Detection with Generative Adversarial Networks Seminar 1: Data Analytics
  • 6. INTRODUCTION • Deep Learning  Omits manual feature engineering  Works well with High Dimensional data like images  Application performance areas: Image classification, Natural language processing, Speech recognition. • State-of-the-art  Achieved by GAN in high-dimensional generative modelling  GAN has two neural networks- Generator and Discriminator  Generator and Discriminator compete with each other 6 Seminar 1: Data Analytics Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
  • 7. INTRODUCTION • Generator  Maps random samples from low-dimensional space to high- dimensional space.  If generator has learnt a good approximation of training data, then for a new point there already exists a point in the latent space which closely resembles this point. • Discriminator  Learns to separate actual samples from the samples generated by the Generator. 7 Seminar 1: Data Analytics Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
  • 8. GENERATOR Vs DISCRIMINATOR 8 Seminar 1: Data Analytics Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks Reference[2]
  • 9. RELATED WORK 1. Anomaly Detection (Approach 1)  Generative models estimate anomalies through estimation of the data distribution p.  In given data, estimate p̂≈p and declare those samples which are unlikely under p̂ to be anomalous.  Methods used:  Kernel density estimation(KDE)  Gaussians for active learning of anomalies  hidden Markov models  dynamic Bayesian networks Sakshi Singh (Universität Hildesheim) 9Image Anomaly Detection with Generative Adversarial Networks Seminar 1: Data Analytics
  • 10. RELATED WORK 2. Deep Generative Models (Approach 2)  Variational Autoencoders(VAEs) have been proposed as deep generative models.  Two approaches followed:  Optimizing over variational lower bound, parameters are tuned such that samples resembling data is generated from a Gaussian prior.  Train pair of deep convolutional in an autoencoder setup and produce random samples on compression manifold.  None of the approach estimates p.  The approach in paper uses deep generative models in anomaly detection. 10Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks Seminar 1: Data Analytics
  • 11. MOTIVATION • Authors addressed the drawbacks of non-parametric anomaly detection methods:  Suffer from curse of dimensionality  Inadequate for analysis and interpretation of high dimensional data • Authors overcome these limitations by using Deep Learning for anomaly detection which is:  End-to-end deep learning approach  Aims specifically at task of anomaly detection 11 Seminar 1: Data Analytics Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
  • 12. STATE-OF-THE-ART • Achieved by GANs • GAN provides framework to generate samples approximately distributed to p. • GAN attempts to learn the parameterization of neural network  Generator(gθ): maps low-dimensional data from noise prior pz (e.g. multi-variate Gaussian) to samples in image space (including qθ ≈ p)  Discriminator(dw): learns to classify the data from p and qθ. • Discriminator becomes better at separating samples from p and samples from qθ, while the generator fools the discriminator by adjusting θ. 12 Seminar 1: Data Analytics Sakshi Singh (Universität Hildesheim)
  • 13. GAN and WGAN • Objective function of GAN where z: vectors lying in latent space of dimensionality d’<<d5; minmax optimization: lower bound of an f-divergence. • Objective function of WGAN(Wasserstein GAN) NOTE: WGAN uses 1-Wasserstein distance(amount of work to pull one density onto other). WGAN training is more stable(as GAN training has vanishing gradients in high dimensions) and is used in the experiments. 13 Seminar 1: Data Analytics Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
  • 14. HYPOTHESIS Given a test sample x, if there exists no z such that gθ(z)≈x, or if such a z is difficult to find, then it can be inferred that x is not distributed according to p, i.e. it is anomalous. where, gθ: the generator x : new test sample z : existing point in latent space p : data distribution through which generative models detect anomalies 14 Seminar 1: Data Analytics Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
  • 15. PROPOSED ALGORITHM: ADGAN • GAN (Drawback):  Discriminator perfectly separates real from fake data, but can’t deal with samples that are unlike training data • ADGAN (Advantage):  The discriminator is never used which is discarded after training, so it is easy to couple ADGAN with any GAN-based approaches like:  Generator network VAEs  Moment matching networks 15 Seminar 1: Data Analytics Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
  • 16. PROPOSED ALGORITHM: ADGAN • ADGAN (Anomaly Detection using GAN) is derived from GAN algorithm. • ADGAN comes into play after GAN training has converged. • Basic workflow of algorithm:  If generator has captured training data distribution with given new sample x~p, there exists a point z(in latent space) such that: gθ(z)≈p  Points that are away from p have no representation or occupy small probability mass in latent distribution.  If z does not exist then we say x is not distributed over p and it is anomalous. 16 Seminar 1: Data Analytics Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
  • 17. ALGORITHM: PSEUDOCODE 17 Assumption: GAN training has converged Goal: Given a new test point x, the algorithm searches for a point z in the latent space, such that gθ(z)≈p Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks Seminar 1: Data Analytics
  • 18. ALGORITHM: EXPLAINATION 1. Initialization:  z0 ~ pz (pz:noise prior from GAN)  θ reset back to original for each new testing point  search is initialized from nseed individual points 2. For t=1,2,…,k steps:  backpropagate the reconstruction loss l(gθ(zt-1),x)  update parameterization of the generator with a small amount (result: series of mapping from latent space that closely resembles x)  Adjust θ to give generator the extra capabilities 18Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks Seminar 1: Data Analytics
  • 19. ALGORITHM • The key idea of ADGAN is that if the generator was trained on the same distribution x was drawn from, then the average over the final set of reconstruction losses will assume low values and high values otherwise. 19Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks Seminar 1: Data Analytics
  • 20. EXPERIMENTS • Datasets  MNIST: contains grayscale scans of handwritten digits  CIFAR-10: contains colored images of real-world objects belonging to 10 classes  LSUN: image datasets showing different scenes(e.g. bedroom, bridges, conference rooms) • Default settings  Training and test splits remain as default  All images are rescaled to assume pixel values in [-1,1] Sakshi Singh (Universität Hildesheim) 20Image Anomaly Detection with Generative Adversarial Networks Seminar 1: Data Analytics
  • 21. METHODS & HYPERPARAMETERS • Tested the ADGAN performance against 4 traditional non- parametric methods of anomaly detection: 1. KDE with a Gaussian kernel 2. One-class support vector machine(OC-SVM) with a Gaussian kernel 3. Isolation Forest(IF) 4. Gaussian mixture model(GMM) • For all methods above, the authors reduced the feature dimensionality before performing anomaly detection. Note: Done via PCA and Alexnet (refer to table 1, slide 26) • AnoGAN: Similar to ADGAN but they use an additional intermediate discriminator layer dw’ 21 Seminar 1: Data Analytics Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
  • 22. METHODS & HYPERPARAMETERS • Report the performance of two end-to-end deep learning approaches: 1. VAEs: scored by evaluating the evidence lower bound (ELBO) 2. DCAE: scored according to reconstructional losses, interpreting a high loss if new samples differ from training samples. • VAE’s performance is better than thresholding directly via prior likelihood in latent space. • Both VAE and DCAE have same architecture as DCGAN, so these two are paired with DCGAN in the experiments. 22 Seminar 1: Data Analytics Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
  • 23. ASSUMPTIONS & RESULTS EXPERIMENTAL ASSUMPTIONS  Set dimensionality of latent space d’=256  Noise prior pz is same used in GAN training  nseed=64 (to maintain non-convexity of optimization problem)  Adam optimizer is used for optimizing the latent vectors and parameters of generator  Squared L2 loss for measuring reconstruction quality EXPERIMENTAL RESULTS  More iterations helped the performance but this gain saturates quickly  K=5 is a good value(where k: no. of iterations) 23Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks Seminar 1: Data Analytics
  • 24. ONE Vs ALL CLASSIFICATION Task 1: Quantify performance of competing methods  Follow the original publications on OC-SVMs[6]  Train models on data from single class from MNIST  Evaluate each model’s performance on 5000 items randomly selected from test set  In each trial, label unseen classes in training data as anomalous 24 Fig. 3. ROC curves for one- versus-all prediction of competing methods on MNIST (left) and CIFAR10 (right), averaged over all classes. KDE and OC-SVM are shown in conjunction with PCA, for detailed performance statistics see Table 1. Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks Seminar 1: Data Analytics
  • 25. Table 1: reports the AUCs that resulted from leaving out each individual class • In these experiments, 1. Average AUC Score according to GAN discriminator: MNIST (0.625) and CIFAR-10 (0.513) 2. Average AUC score according to prior likelihood pz of final latent vector: MNIST (0.554) and CIFAR-10 (0.721) {better} 25 Seminar 1: Data Analytics Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
  • 26. UNSUPERVISED ANOMALY DETECTION Task 2: showcase the use of ADGAN in unsupervised detection  Build a test set from 300 validation sample images  Train the generator on LSUN scenes (in these experiments: only bedroom scene)  Record whether ADGAN assigns low or high anomaly score  The images associated with highest and lowest anomaly scores of three different scene categories: Figure 4,5,6 in next slides 26Sakshi Singh (Universität Hildesheim) Seminar 1: Data Analytics Image Anomaly Detection with Generative Adversarial Networks
  • 27. 27Sakshi Singh (Universität Hildesheim) Seminar 1: Data Analytics Image Anomaly Detection with Generative Adversarial Networks
  • 28. CONCLUSION • ADGAN is able to outperform traditional methods of detecting anomalies in high-dimensional image samples. • ADGAN has ability to discern usual from unusual samples: anomaly detection • ADGAN is able to incorporate many properties of an image (e.g. colors, canonical geometrics, foreign objects(caption)). 28 Seminar 1: Data Analytics Image Anomaly Detection with Generative Adversarial NetworksSakshi Singh (Universität Hildesheim)
  • 29. FUTURE IMPROVEMENTS • Performance can be boosted by additional tuning of the underlying neural network • To achieve strong experimental results jointly optimize the latent vectors and generator parameterization • Improve by initializing from an approximate inversion of the generator[3,4] and substituting the reconstruction loss for a more elaborate variant[5] 29 Seminar 1: Data Analytics Sakshi Singh (Universität Hildesheim) Image Anomaly Detection with Generative Adversarial Networks
  • 30. REFERENCES 1. Image Anomaly Detection with Generative Adversarial Networks: by Lucas Deecke, Robert Vandermeulen, Lukas Ruff, Stephan Mandt, and Marius Kloft 2. https://developers.google.com/machine-learning/gan/gan_structure, Overview of GAN Structure, accessed on 2.12.2019 3. Donahue, J., Krähenbühl, P., Darrell, T.: Adversarial feature learning. In:International Conference on Learning Representations (2017) 4. Dumoulin, V., Belghazi, I., Poole, B., Lamb, A., Arjovsky, M., Mastropietro, O., Courville, A.: Adversarially learned inference. In: International Conference on Learning Representations (2017) 5. Ling, H., Okada, K.: Diffusion distance for histogram comparison. In: Computer, Vision and Pattern Recognition. pp. 246–253. IEEE (2006) 6. Schölkopf, B., Platt, J.C., Shawe-Taylor, J., Smola, A.J., Williamson, R.C.: Estimating the support of a high-dimensional distribution. Tech. Rep. MSRTR-99-87, Microsoft Research (1999) 30Image Anomaly Detection with Generative Adversarial NetworksSakshi Singh (Universität Hildesheim) Seminar 1: Data Analytics