Unsupervised Representation Learing with Deep Convolutional Generative Adversarial Networks

Recently, I am reading papers which is refer to Deep learning and Convolutional Networks(CNNs). The aim to write the blog is better understanding of above methods.

Please forgive me there are full of wrong ideas and understanding and don’t hesitate to contract with me, which make full of grateful.

Why the authors want to write the paper?

At present, the attracting CNNs, a effective method that has a wide rang of applications in deep learning, has seen huge usage with supervised learning in computer visions. Comparatively, less attention has been payed to the unsupervised learning with CNNs.

What is the starting point or breakthrough point to solve the above problem?

In order to deal with the above problem, the author introduce one way that is based on the Generative Adversarial Networks which provides an attractive alternative to maximum likelihood technique instead of facing intractable probabilistic computations that raise in maximum likelihood estimation and related strategies, as well as the difficulty of leveraging benefits of piecewise liner units in the generative context. The method in this paper is called Deep Convolutional Generative Adversarial Networks (DCGANs)

The architecture

1 the genrator used the fractional-strided convolutions

2 the discriminator used the strided convolutions

3 remove fully connected hidden layers for deeper architecture

4 use ReLU activation in generator for all layers except for the output, which uses Tanh

5 use LeakyReLU activation in the discriminator for all layers

4 The modeling