- Most Cited Deep Learning Papers # Unsupervised | Github
- 从OpenAI看深度学习研究前沿 | 知乎专栏
- Generative Models | OpenAI博客
- A Path to Unsupervised Learning through Adversarial Networks
- Generative Adversarial Networks(GAN)的现有工作 | 程序媛的日常
- 生成内容
- 生成对抗网络Generative Adversarial Network
- 变分自编码机Variation Auto Encoder
- Pixel RNN类模型
- 自编码机Auto Encoder
- 梯子网络Ladder Network
- One-shot Learning
- Zero-shot Learning
- Biologically Plausible Learning
- 其他
A Neural Algorithm of Artistic Style
- 传说中的Neural Style
Image Completion with Deep Learning in TensorFlow (代码)
- 用GAN做图像修复(image inpainting任务),主要思想是同时优化两个目标:
- 1.原图中有完好区域和丢失区域,要让生成的修复图与原图在对应的完好区域尽可能接近(所谓Contextual Loss)
- 2.要让生成的修复图尽可能被GAN的判别器判定为真实图片,尽可能像真的(所谓Perceptual Loss)
- 论文:Semantic Image Inpainting with Perceptual and Contextual Losses
Generative Adversarial Networks (代码)
- Goodfellow的GAN开山之作
Conditional Generative Adversarial Nets
Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks (代码)
-
生成房间图片
-
戴眼镜男人-不戴眼镜男人+不戴眼镜女人=戴眼镜女人
-
从一张人脸渐变到另一张人脸
-
“这篇论文的提出看似并没有很大创新,但其实它的开源代码现在被使用和借鉴的频率最高……这些工程性的突破无疑是更多人选择 DCGAN 这一工作作为 base 的重要原因”
Improved Techniques for Training GANs (代码)
- 改变架构,解决GAN训练不稳定的问题
- 半监督学习,少量标注样本,效果比Ladder Network还好一些
InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets (代码)
- 对representation code空间施加一些要求,使其更具结构化,而非混沌一团
- 结果在representation向量的单个维度上获得了非常好的可解释性,例如渐变一个维度的数值,生成的人脸图谱从“抬头姿态”到“低头姿态”渐变,非常像流形学习里面的一些例子
Pixel Recurrent Neural Networks
Conditional Image Generation with PixelCNN Decoders
Stacked What-Where Auto-encoders
From neural PCA to deep unsupervised learning
- 提出Ladder架构,但还未做半监督学习
Semi-Supervised Learning with Ladder Network
- 半监督学习,MNIST用100个标注数据达到约99%,CIFAR用4000个标注数据达到约80%
Deconstructing the Ladder Network Architecture
- 深入挖掘Ladder Network的原理
One-Shot Generalization in Deep Generative Models
Towards Biologically Plausible Deep Learning
Towards a Biologically Plausible Backprop
Feedforward Initialization for Fast Inference of Deep Generative Networks is Biologically Plausible
Towards Principled Unsupervised Learning
- 用GAN做半监督学习的论文中所定义的新的损失函数与这篇提出的Output Distribution Matching (ODM) cost有紧密联系