deep learning for the earth sciences 读书笔记

第二章 learning unsuoerxised festure representations of remote sensing data with sparse convolutional

2.1 Introduction

第一段：

内容：

总述了很多领域都是用的结构化的领域（时间空间维度有固定的格式）；

结构化的坐标表示数据在可视化理解和设计算法上有很大的难度由于高维度和协方差矩阵的相关性；

这就是为什么特征表示在科学界有很多兴趣；

这就是特征提取和降维的领域有监督和无监督学习的方法。

好的语句：

1.fields like ****work in the **, in which(解释**)

2.from a geometrical viewpoint, **,but visualizing, understanding and designing algorithms therein is challenging,mainly due to the high dimensionality and correlation between covariates.在可视化理解和设计算法上有很大的难度由于高维度和协方差矩阵的相关性

3. this is why **has captured a lot of interest in the scientific community.这就是为什么**在科学界有很多兴趣

4.this is the field of **,for which ****

第二段

内容：

介绍了一些无监督学习的原始方法：

1.总述：在标签稀疏的情况下，无监督学习是首选

2.介绍一些传统的算法主成分分析以及非线性的流形学习算法（局部和非局部）；

3.深度学习成为一个趋势；

4.自动编码：

an autoencoder is a deep learning architecture in which the input signal is reconstructed/decoded at the output layer through an intermediate layer with reduced number of hidden nodes. Basically, the autoencoder aims at reproducing the inputs at the output layer by using the high-abstraction features learned in the intermediate layer. The use of autoencoders implies, however, the tuning of several free parameters, addressing the regularization issue mainly by limiting the structure of the network heuristically.自动编码器是一种深度学习体系结构，其中输入信号通过隐藏节点数量减少的中间层在输出层进行重构/解码。基本上，自动编码器的目标是通过利用在中间层学习到的高抽象特征来再现输出层的输入。然而，自动编码器的使用意味着要调整几个自由参数，主要通过启发式地限制网络的结构来解决正则化问题。

5.用深度网路存在的问题，是由于自然数据形成的过程导致了强的自相关函数（在空间和时间上）、重尾、强的特征相关性，不一定是线性的，以及（从流形学习的角度来看）数据存在于（很多）低维的子空间中。这可以解释为，观察结果是由一个简单得多的机制产生的，人们应该寻求一个稀疏的表示；

While the issue of non-linear feature extraction can be resolved with deep networks efficiently, there is still the issue of what a sensible criterion should be employed. Natural data, namely, data generated by a physical process or mechanism, often show strong autocorrelation functions (in space and time), heavy tails, strong feature correlations, not necessarily linear, and (from a manifold learning perspective) data lives in subspaces of (much) lower dimensionality. This can be interpreted as that the observations were generated by a much simpler mechanisms, and that one should seek for a sparse representation

6.稀疏编码和字典学习都是很好地无监督学习方法。字典学习：

A dictionary can be understood as a set of bases that sparsely  represent the original data. The main goal in dictionary learning is to find a dataset to best represent the inputs by only using a small subset (also known as atoms) of the dictionary.字典可以理解为一组稀疏地表示原始数据的基。字典学习的主要目标是通过只使用字典的一个小子集（也称为原子）来找到一个最好地表示输入的数据集。

                                   

未解决的问题：

These methods describe the input images in sparse representation spaces but do not take advantage of the high non-linear nature of deep architectures. However, attaining sparse non-linear deep networks is still unresolved in the literature, especially in unsupervised learning。

这些方法描述了稀疏表示空间中的输入图像，但没有利用深度体系结构的高非线性性质。然而，获得稀疏非线性深度网络在文献中仍未得到解决，特别是在无监督学习中

                                                                                                                    

好的语句：

1.** is the prefered approach in cases of *****  在*****的情况下，** 是首选

2.**method can be mainly categorized as either local or global approaches. Local methods retain local geometry of data, and are computationally efficient since only sparse matrix computations are needed. Global manifold methods, on the other hand, keep the entire topology of the dataset, yet are computationally expensive for large datasets. They have higher generalization power, but local ones can perform well on datasets with different sub-manifolds.局部和非局部的优劣势

3.The use of **in *** is wide spread for a wide range of applications, including****.**在**方面有广泛应用比如：

2.2 sparse unsupervised convolutional networks