Generative adversarial network

cosmos 31st March 2019 at 4:39am
Generative model

An architecture to train generative neural networks, i.e. neural networks which act as Generative models, i.e. their inputs are Latent variables, and their output is the observed data.

https://deephunt.in/the-gan-zoo-79597dc8c347https://github.com/wiseodd/generative-models

NIPS tutorial

GAN on browser

example of GANs for discrete data: https://arxiv.org/pdf/1611.08408.pdf (in area of semantic segmentation) It's funny how people in NLP literature keep talking about how GANs are hard to make work for discrete data. Then, people in image segmentation literature (oblivious of all this?) successfully apply vanilla GANs to discrete data..

Adversarial networks

The way the network is trained is by having the generative network produce images, while training a different discriminative network to discriminate between real and generated images. In this way, the discriminative network can be used as a very good cost function, which penalized generated images which are distinguishably different from the real images that the generative network is train to model. E.g. for NLP literature stuff https://arxiv.org/abs/1611.04051?fbclid=IwAR0MOTB8nsV555rKdNb3EQ8pt-LbTKcxUHP_5ZC-of4zDGf_Et_uz1r39b4

We are in effect learning the cost function

video

Trining GANs

NIPS 2016 Workshop on Adversarial Training - Soumith Chintala - How to train a GAN

Trained via Gradient descent and Backpropagation

I think that the discriminator needs to have a separate cost function which measures the number of images the discriminator missclassified as being real or generated. Discriminator is optimized to not be fooled by the generator

On the other hand, the generative network uses the discriminative network as cost. Generator to fool discriminator, i.e. it is trained to maximize the mistakes the the discriminator does, that's why they are called adversarial

The discriminator is teaching the generator, and it's adapting to the generator's knowledge and flaws. Machine teaching, not just machine learning.

Alternating Optimization

video

Improved Techniques for Training GANs vid

Theoretical properties

If you have an optimal discriminator, the generator minimizes the Jensen-Shanon divergence

Variants

Original GANs were diffucult to train

Class-conditional GANs

They are supervised

Deep Generative Image Models using a Laplacian Pyramid of Adversarial Networks vid

Video-prediction GANs

Deep multi-scale video prediction beyond mean square error vid

DCGANs

Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks

Latent space arithmetic

In-painting GANs

Context Encoders: Feature Learning by Inpainting

video

Applications

Feature learning for Semi-supervised learning

GANs for feature learning.

Two Minute Papers - Image Editing with Generative Adversarial Networks

Disentangling representations

vid

InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets (from OpenAI)


http://www.inference.vc/how-to-train-your-generative-models-why-generative-adversarial-networks-work-so-well-2/

They are similar to Autoencoders but we learn the cost function, instead of just using l2 loss (vid)

vid

Generative Adversarial Networks

https://www.wikiwand.com/en/Generative_adversarial_networks

the future

Connecting Generative Adversarial Networks and Actor-Critic Methods


More variations and others

URL list from Sunday, May. 21 2017 16:41 PM

To copy this list, type [Ctrl] A, then type [Ctrl] C.

1603.08155.pdf https://arxiv.org/pdf/1603.08155.pdf

1703.10593.pdf https://arxiv.org/pdf/1703.10593.pdf

1610.09003.pdf https://arxiv.org/pdf/1610.09003.pdf

RL Course by David Silver - Lecture 5: Model Free Control - YouTube https://www.youtube.com/watch?v=0g4j2k_Ggc4&index=5&list=PL7-jPKtc4r78-wCZcQn5IqyuWhBZ8fOxT

Teaching http://www0.cs.ucl.ac.uk/staff/d.silver/web/Teaching.html

Lecture 1a - Introduction [Phil Blunsom] - YouTube https://www.youtube.com/watch?v=RP3tZFcC2e8&list=PL613dYIGMXoZBtZhbyiBqb0QtgK6oJbpm

[1612.03242] StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks https://arxiv.org/abs/1612.03242

[1703.06412] TAC-GAN - Text Conditioned Auxiliary Classifier Generative Adversarial Network https://arxiv.org/abs/1703.06412

[1703.06676] I2T2I: Learning Text to Image Synthesis with Textual Data Augmentation https://arxiv.org/abs/1703.06676

I2T2I: LEARNING TEXT TO IMAGE SYNTHESIS WITH TEXTUAL DATA AUGMENTATION https://arxiv.org/pdf/1703.06676.pdf

Generative Adversarial Text to Image Synthesis https://arxiv.org/pdf/1605.05396.pdf

Reed: Generative adversarial text to image synthesis - Google Scholar https://scholar.google.co.uk/scholar?start=70&hl=en&as_sdt=0,5&sciodt=0,5&cites=8255440757806230750&scipsc=

Learning What and Where to Draw http://papers.nips.cc/paper/6111-learning-what-and-where-to-draw

Generating Visual Explanations | SpringerLink https://link.springer.com/chapter/10.1007/978-3-319-46493-0_1

[1609.09444] Contextual RNN-GANs for Abstract Reasoning Diagram Generation https://arxiv.org/abs/1609.09444

[1702.03431] Crossing Nets: Dual Generative Models with a Shared Latent Space for Hand Pose Estimation https://arxiv.org/abs/1702.03431

DISCO Nets : DISsimilarity COefficients Networks http://papers.nips.cc/paper/6143-disco-nets-dissimilarity-coefficients-networks

[1704.06933] Adversarial Neural Machine Translation https://arxiv.org/abs/1704.06933

[1702.04125] One-Step Time-Dependent Future Video Frame Prediction with a Convolutional Encoder-Decoder Neural Network https://arxiv.org/abs/1702.04125

[1703.06029] Towards Diverse and Natural Image Descriptions via a Conditional GAN https://arxiv.org/abs/1703.06029

cvpr17_summarization.pdf http://web.engr.oregonstate.edu/~sinisa/research/publications/cvpr17_summarization.pdf


http://make.girls.moe/