Oral Presentation by Abdullah-Al-Zubaer Imran for the paper entitled "Multi-Adversarial Variational Autoencoder Networks" at the 2019 IEEE International Conference on Machine Learning and Applications (ICMLA) which was held in Boca Raton, FL, USA.
4. 12/17/2019
Multi-Adversarial Variational Autoencoder Networks
4
Deep Generative Models: GAN
[Goodfellow et al. 2014, Radford et al. 2015]
• Mini-max game
• Generator maps latent variables to data samples
• Discriminator distinguishes generated and real samples
• Sharpest image generation
• Unstable and difficult to optimize
• Losses
𝐷
𝑚𝑎𝑥
𝑉 𝐷 = 𝐸 𝑥~𝑝 𝑑𝑎𝑡𝑎(𝑥) 𝑙𝑜𝑔𝐷 𝑥 + 𝐸 𝑥~𝑝 𝑧(𝑧) log(1 − 𝐷 𝐺(𝑧)
𝐺
𝑚𝑖𝑛
𝑉 𝐺 = 𝐸 𝑥~𝑝 𝑧(𝑧) log(1 − 𝐷 𝐺(𝑧)
5. 12/17/2019
Multi-Adversarial Variational Autoencoder Networks
5
Deep Generative Models: PixelRNN
[Oord et al. 2016]
• Autoregressive model
• Simple and stable training process
• Inefficient sampling
• Assign probability to every pixel in the image
• Softmax loss
𝑝 𝑥 =
𝑖=1
𝑛2
𝑝(𝑥𝑖|𝑥1, … , 𝑥𝑖−1)
𝑝 𝑥𝑖, 𝑅|𝑥<𝑖 𝑝(𝑥𝑖, 𝐺|𝑥<𝑖, 𝑥𝑖,𝑅)𝑝(𝑥𝑖, 𝐵|𝑥<𝑖, 𝑥𝑖,𝑅, 𝑥𝑖,𝐺)
6. Aim
12/17/2019
Multi-Adversarial Variational Autoencoder Networks
6
VAE-GAN
PixelGAN Autoencoder
• Improving the deep generative models
• Evaluation measures
[Larsen et al. 2016, Makhzani et al. 2017]
Primary Aim
(Efficient and stable generative modeling
for medical image analysis)
✓ Combining generative models
✓ High quality image generation
✓ Learning from limited labeled data
7. Proposed: MAVENs
• Highlights
• Ensemble of multiple discriminators in VAE-GAN
• Joint image generation and classification
• Motivation
• Instability in generative models
• Mode collapsed generation
• Poor image quality in VAE
• Small labeled data
• Objective
• Improve samples and semi-supervised classification
• Unified generative model
• Variational inference with adversarial learning
12/17/2019 7
Basic comparisons of MAVEN with GAN, VAE, and VAE-GAN
Multi-Adversarial Variational Autoencoder Networks
GAN-mode collapsed generation
10. MAVENs: Implementation Details
• Datasets
• SVHN (32 x 32 x 3) [street view digits]
• CIFAR10 (32 x 32 x 3) [outdoor natural images]
• Chest X-ray (128 x 128 x 1) [normal, bacterial and virus-pneumonia]
• Baselines: DC-GAN and VAE-GAN
• MAVENs with 2, 3, and 5 discriminators
• Feedback as mean or random selection
• Merely with 10% training data with their corresponding label information
12/17/2019 Multi-Adversarial Variational Autoencoder Networks 10
11. MAVENs: Evaluations
12/17/2019
Multi-Adversarial Variational Autoencoder Networks
11
• Image quality
• Fréchet Inception Distance (FID)
• Activation from pool3 of inception-v3 model
𝐹𝐼𝐷 = µ 𝑑𝑎𝑡𝑎 − µ 𝑓𝑎𝑘𝑒
2
+ 𝑇𝑟 𝞢 𝑑𝑎𝑡𝑎 + 𝞢 𝑓𝑎𝑘𝑒 − 2(𝞢 𝑑𝑎𝑡𝑎 𝞢 𝑓𝑎𝑘𝑒)1/2
• Descriptive Distribution Distance (DDD)
• Comparing first four moments of the two distributions
𝐷𝐷𝐷 =
𝑖=1
𝑖=4
−𝑙𝑜𝑔𝑤𝑖 µ𝑖 𝑑𝑎𝑡𝑎
− µ𝑖 𝑓𝑎𝑘𝑒
• Classification
• Overall accuracy
• Class-wise F1 scoring
𝐹1 =
2 ∗𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 ∗𝑟𝑒𝑐𝑎𝑙𝑙
𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑟𝑒𝑐𝑎𝑙𝑙
19. Model Vs Real Distributions: Not-So-Good Match
12/17/2019 Multi-Adversarial Variational Autoencoder Networks 19
CXR
20. Conclusions & Future Work
Significance
New generative model
Improved image quality and classification
Evaluation measure for deep generative models
Limitation
Performance for medical image data
Execution time
What’s Next
Hyper-parameters for medical images
Constrained generation
Complex image analysis tasks
Generative multi-tasking
12/17/2019 Multi-Adversarial Variational Autoencoder Networks 20