提及深度学习的优点就要从深度学习的前身来比较。机器学习的历史阶段的划分

Mention the advantages of deep learning from the predecessor of deep learning to compare. The division of the historical stage of machine learning can be divided according to the hierarchical structure of the machine learning model. From the 1980s to the present, the process of machine learning has gone through two stages: shallow learning and deep learning. <br>Shallow learning uses shallow structure. The shallow structure is used for most traditional machine learning and signal processing technology applications (such as Gaussian mixture models, linear or nonlinear dynamic systems, conditional random fields, maximum entropy models, support for vertigo, kernel regression, multi-layer perceptrons, etc. ). These structures usually contain one to two layers of non-linear feature transformations, and can be considered as a structure with one hidden layer or no hidden layer. The shallow structure has shown its effectiveness in solving some easy or limited problems, but because of the limited modeling and representation capabilities of shallow learning, it can be used to deal with more complex practical applications (such as natural speech signals, Natural image signals and visual scene signals are very difficult. <br>Deep learning uses deep networks. Deep networks are networks with multiple hidden layer structures. By introducing a deep network, the system model can achieve an approximation of a complex function by learning a deep nonlinear network, thereby calculating more complex input features. Since each hidden layer can perform a non-linear transformation on the output of the previous layer, the deep network has a better expression ability than the shallow network. For example, more complex functional relationships can be obtained through learning, and it shows from The ability to learn the essential characteristics of data in a few samples. <br>The main advantage of the deep network is that it can represent a much larger set of functions than the traditional shallow network in a simpler way, and the advantage of the multi-layer is that it can use less parameters to represent complex function relationships. It should be noted that in the process of training deep networks, each hidden layer needs to use a nonlinear activation function, because the combination of multiple linear functions is also a linear function in nature. Therefore, if the activation function uses a linear function, it does not increase the expressive power compared to the network with a single hidden layer.<br>When the processing object is an image, the decomposition relationship of "part-whole" can be learned by using a deep network. For example, the first layer can learn how to combine pixels in an image to detect edges, and the second layer can combine edges to detect longer contours or simple "target parts." At a deeper level, these contours can be further combined to detect more complex features. <br>The essence of deep learning is to learn more useful features by building a learning model with multiple hidden layers and massive training data, thereby improving the accuracy of classification or prediction. Therefore, "deep model" is a means, " "Feature learning" is the purpose.

提及深度学习的优点就要从深度学习的前身来比较。机器学习的历史阶段的划分可按机器学习模型的层次结构来划分。从20世纪80年代至今，机器学习的历程经历了两个阶段：浅层学习和深度学习。<br>浅层学习运用浅层结构。浅层结构用于大多数传统的机器学习和信号处理技术的应用（例如高斯混合模型、线性或非线性动力系统、条件随机场、最大熵模型、支持向晕机、核回归、多层感知机等）。这些结构通常包含一层至两层的非线性特征变换，可以当成是含有一层隐含层或者没有隐含层的结构。浅层结构在解决一些容易的或者受限的问题中显示出了有效性，但是因为浅层学习的有限建模和表征能力，使其在处理较为复杂的实际的应用时（例如自然语音信号、自然图像信号和视觉场景信号这些自然信号）十分艰难。<br>深度学习运用深度网络。深度网络是含有多个隐含层结构的网络。通过引入深度网络，系统模型可以通过学习一种深层的非线性网络，来实现复杂函数的逼近，从而计算更为复杂的输入特征。由于每一个隐含层可以对上一层的输出进行非线性变换，因此深度网络拥有比浅层网络更为优异的表达能力，例如可以通过学习得到更为复杂的函数关系，并且表现出了从少数样本中学习数据的本质特征的能力。<br>深度网络最主要的优点在于，它能用更加简单的方式来表示比传统浅层网络大得多的函数集合，而多层的优势是可以利用较少的参数来表示复杂的函数关系。需要注意的是，训练深度网络的过程中，每一层隐含层需要使用非线性的激活函数，这是因为多层的线性函数组合在一起本质上也还是线性函数。因此，如果激活函数使用线性函数，相比于单隐含层的网络，并没有增加表达能力。<br>当处理对象是图像时，通过使用深度网络可以学习到“部分－整体”的分解关系。 例如，第一层可以学习如何将图像中的像素组合在一起来检测边缘，第二层可以将边缘组合起来检测更长的轮廓或者简单的“目标部件”。在更深的层次上，可以将这些轮廓进一步组合起来以检测更为复杂的特征。<br>深度学习的实质，是通过构建具有多隐含层的学习模型和通过海量的训练数据，来 学习更为有用的特征，从而提升分类或预测的正确性，因此，“深度模型”是手段，“特征学习”是目的。

The advantages of deep learning should be compared from its predecessor. The historical stages of machine learning can be divided according to the hierarchical structure of machine learning model. Since 1980s, machine learning has experienced two stages: shallow learning and deep learning.<br>Shallow learning uses shallow structures. Shallow structure is used in most traditional machine learning and signal processing applications (such as Gaussian mixture model, linear or nonlinear dynamic system, conditional random field, maximum entropy model, airsickness support, nuclear regression, multi-layer perceptron, etc.). These structures usually contain one or two layers of nonlinear feature transformation, which can be regarded as structures with or without one layer of hidden layer. Shallow structure shows its effectiveness in solving some easy or limited problems, but because of the limited modeling and representation ability of shallow learning, it is very difficult to deal with more complex practical applications (such as natural voice signals, natural image signals and visual scene signals).<br>Deep learning uses deep networks. Deep network is a network with multiple hidden layer structures. By introducing the depth network, the system model can realize the approximation of complex functions by learning a kind of deep nonlinear network, so as to calculate more complex input characteristics. Because each hidden layer can transform the output of the upper layer nonlinearly, the deep network has more excellent expression ability than the shallow network. For example, it can get more complex functional relationships through learning, and it shows the ability to learn the essential characteristics of data from a few samples.<br>The main advantage of deep network is that it can represent a much larger set of functions than traditional shallow network in a simpler way, while the advantage of multi-layer network is that it can use fewer parameters to represent complex functional relationships. It should be noted that in the process of training depth network, each layer of hidden layer needs to use nonlinear activation function, because the combination of multi-layer linear functions is also linear in nature. Therefore, if the activation function uses linear function, it does not increase the expression ability compared with the network with single hidden layer.<br>When the processing object is an image, we can learn the "part whole" decomposition relationship by using the depth network. For example, the first layer can learn how to combine pixels in an image to detect edges, and the second layer can combine edges to detect longer outlines or simple "target parts". At a deeper level, these contours can be further combined to detect more complex features.<br>The essence of deep learning is to learn more useful features by building a learning model with multiple hidden layers and through massive training data, so as to improve the accuracy of classification or prediction. Therefore, "deep model" is a means and "feature learning" is the purpose.<br>