论文阅读-solving-ill-posed-inverse-problems-using-iterative-deep-neural-networks

less than 1 minute read

Published:

论文阅读-solving-ill-posed-inverse-problems-using-iterative-deep-neural-networks

论文阅读-Solving ill-posed inverse problems using iterative deep neural networks

inverse problem:


给定数据g,得到
通常方法:

是对f的先验信息的总结。
问题:需要计算量,需要显式地将S表达出来,正则化参数的选择
机器学习方法:
直接找一个pseudo-inverse:

监督性学习:

是分布,d是距离函数

非监督性:

as learning an optimiser for the variational rgularization

fully learned reconstruction :
通过data构建一个来逼近T的逆
缺点:需要大量的数据来满足各种的data manifold

sequential data and knowledge driven reconstruction:

损失函数:

找到最好的f使得损失最小。

假设右侧是Frechet differentiable,and convex

用普通的梯度下降法

用一个更新算子代替梯度部分:

Alt text
用CNN拟合更新算子

Alt text
加入一个memory ,用拟牛顿法

Alt text
问题:这里是怎么来的?

Alt text
用CNN,是3*3的卷积核


Deep Bayesian Inversion

a sampling based method using
a Wasserstein GAN with a novel mini-discriminator and a direct approach that trains
a neural network using a novel loss function.