[Paper Review] 06.(improved gan training) Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks

[Paper Review] 06.Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks

Contents

1. Abstract
2. Introduction
3. Related Work
   1. Representation Learning from Unlabeled data
   2. Generating Natural Images
   3. Visualizing the internals of CNNs
4. Approach and Model Architecture
5. Details of Adversarial Training

0. Abstract

• supervised learning with CNNs : HOT
• UNsupervised learning with CNNs : less attention

\(\rightarrow\) bridge the gap between supervised & unsupervised

propose DCGANs ( = Deep Convolutional GANs )

• learns a hierarchy of representations from object parts to scenes in both generator & discriminator

1. Introduction

learn feature representations from large UNlabled datasets

This paper proposes..

• 1) train GAN
• 2) reuse parts of the G & D networks as feature extractors for supervised tasks

Contributions

• (1) introduce DCGAN
• (2) trained discriminator for image classification
• (3) visualize the filters learnt by GANs

2. Related Work

(1) Representation Learning from Unlabeled data

• skip

(2) Generating Natural Images

main 2 categories

• 1) parametric
  • has been explored extensively
  • ex) GAN
    • however, have not leveraged generators for supervised tasks
• 2) non-parametric
  • matching from a database of existing images

(3) Visualizing the internals of CNNs

NN = black-box methods

Solution

• deconvolutions & filtering the maximal activations

  \(\rightarrow\) find the approximate purpose of each convolution filters

3. Approach and Model Architecture

attempts to scale up GANs using CNNs…unsuccessful

\(\rightarrow\) LAPGAN (2018)

• iteratively upscale low resolution generated images

Architecture guidelines for stable DCGANs

Core approach ( = adopt & modify 3 recent changes to CNN )

[1] spatial pooling functions (ex. maxpooling) \(\rightarrow\) STRIDED CONVOLUTONS

• learn its OWN spatial downpooling

[2] ELIMINATE FC layers on top of convolutional features

[3] use BATCH NORMALIZATION

• prevent generator from collapsing all samples to a single point

4. Details of Adversarial Training

• no-preprocessing besides scaling
• SGD with mini-batch size 128
• initial weights ~ \(N(0,0.02^2)\)
• Adam optimizer
• Leakly ReLU’s slope = 0.2
• learning rate : 0.0002