• paper

Current techniques for training generative models require access to fully-observed samples.
It is expensive or even impossible to obtain fully-observed samples, but economical to obtain partial, noisy observations.
We show that the true underlying distribution can be provably recovered even in the presence of per-sample information loss for a class of measurement models.
We call this AmbientGAN.

• GAN에 대한 일반적인 설명이 나옵니다.
• train을 위해서는 large number of fully-observed samples 이 필요하다.
• 어떤 분야에서는 samples의 획득 비용이 너무 높거나 실용적이지 않은 경우가 있다.(many sensing and tomography problems, MRI, CT Scan)
• 그런데 처음부터 sensing이 비싸다면? 어떻게 훈련이 가능할 정도로 데이터를 모을 수 있을 것인가?

This work solves this chicken-and-egg problem by training a generative model directly from noisy or incomplete samples.
We show that our observations can be even projections or more general measurements of different types and the unknown distribution is still provably recoverable. A critical assumption for our framework and theory to work is that the measurement process is known and satisfies certain technical conditions.

• The idea is simple: rather than distinguish a real image from a generated image as in a traditional GAN, our discriminator must distinguish a real measurement from a simulated measurement of a generated image.

• Our work is also closely related to [Gadelha et al. (2016)] where the authors create 3D object shapes from a dataset of 2D projections. We note that their setup is a special case of the AmbientGAN framework where the measurement process creates 2D projections using weighted sums of voxel occupancies.

* 출처: https://ml-dnn.tistory.com/9#2-1-Notation

• Block-Pixels: 확률 p에 따라 독립적으로 각각의 픽셀 값이 0
• Convolve+Noise: k를 convolution kernel 이라고 하고, 이라고 하면,
   measurements는
      
• Block-Patch: 임의로 선정된 k x k patch가 0으로 설정.
• Keep-Patch: 임의로 선정된 k x k patch 바깥의 값들이 0으로 설정.
• Extract-Patch: 임의로 선정된 k x k patch가 추출됨.
• Pad-Rotate-Project: 4면을 모두 패딩하고 회전을 시킴.
• Pad-Rotate-Project-세타: 위와 같은데 대신 회전각을 정해줌.
• Gaussian-Projection: random Gaussian vector로 project함.

• MNIST, CelebA, CIFAR-10
• MNIST:
    first model- conditional DCGAN, second model- unconditional Wasserstein GAN with gradient penalty.
• celebA:
    unconditional DCGAN
• CIFAR-10:
    Auxiliary Classifier Wasserstein GAN with gradient penalty(ACWGANGP)

• "ignore" baseline: A crude baseline is to ignore that any measurement happened at all.
• "stronger" baseline: If the measurement functions were invertible, and we observed for each measurement y_i in our dataset, we could just invert the functions to objain full-samples
   
  

1. We may not observe
2. Functions may not be invertible.

자세한 설명은 논문으로 대체하겠습니다.

자세한 설명은 논문으로 대체하겠습니다.