DCGAN

DCGAN

1. Related Works

   • Representation Learning From Unlabeled Data
     • Clustering
     • Auto-encoder
     • Deeb beleif
   • Generative Models
     • Parametric
       • GANs
       • Laplacian Pyramid GAN
     • Non-parametric
       • Often do matching from a database of existing images, often matching patches of images

2. Architecture

   • No max pooling
   • No fully connected layers
   • Batch Normalization
   • ReLU

3. Training

   • LSUN
   • Faces
   • IMAGENET-1K

4. Validation

   • Classifying CIFAR using GAN as feature extractor
     • By applying the GAN  as feature extractor on a supervised dataset, and evaluate the performance of linear models fitted on top of these features, the DCGAN has been trained using Imagenet and then use the discriminators for feature extraction with L2-SVM as a classifier
   • Classifying SVHN using GAN feature extrator

5. Visualizing the internal of the network

   • This section helps to explore the internal of the network to make sure there is no memorization and to explore which filters are responsible for what
   • Latent SPace
     • The smooth transition of the generated samples by slightly change the latent space means there is no memorization
   • Discriminator features
     • This helps to explore features learnt by the discriminator, like beds and windows
   • Forgetting to draw certain objects
   • Vector Arithmetic on Faces samples