commonly used in deep learning. It mainly introduces Alex Net, ZF Net, VGG Net,
Google Net, Res Net, spatial transformer module, capsule network, and a comparative
analysis of the various improved models for deep learning. Finally, the problems that
still need to be solved in this field and the future of deep learning are introduced.
2 The Composition and Principle of Neural Network
The basic neural network consists of input layer, hidden layer, and output layer. The
structure is shown in Fig. 1. For the classification problem, the main task of the input
layer is to read the information of the image, the number of neurons is the number of
pixels of the input image, and the number of neurons in the output layer is the number
of categories. For example, if the MNIST handwritten digital image set is a 28pixel  28-pixel grayscale image, the number of neurons in the input layer of the
neural network is set to 28 Â 28 = 784 neurons, and the 10 neurons in the output layer
are derived from 10 categories. (Numbers 0 to 9 for a total of 10 categories) The
activation function converts the sum of the input signals into an output signal. The
activation function mostly uses Relu, and the activation function of the output layer
uses the Softmax function. The convolutional neural network also includes a convolutional layer, a pooled layer, and a fully connected layer. The convolutional layer is
used for feature extraction, and the convolution operation is equivalent to the filter
operation in image processing. Each time a feature is perceived locally in the image and
then synthesized at a higher level to obtain information about the entire image. And in
order to improve efficiency, parameter sharing and local sensing are introduced [1]. The
pooling layer is an operation for reducing the space in the high and long directions and
is disposed behind the convolution layer. This operation can reduce the amount of data
while retaining valid information. Max pooling is generally used in identifying problems. After using multiple convolutional layers and pooling layers, the fully connected
layer is fully connected to all neurons in the previous layer, integrating local information, and linking all features. Convolutional neural networks as the hierarchy deepen, the extracted information is increasingly complex and abstract, from the initial
layer to the subsequent layer in turn respond to image edges, textures, and more
complex object components. In the improved CNN model, the activation function main
uses the ReLu function. Its expression is as follows hðxÞ ¼
x ðx [ 0Þ
0 ðx 0Þ
When the
input is greater than 0, the value is directly output; when the input is less or equal to
zero, 0 is output. Most use dropout technology, and use GPU parallel training technology, the classifier uses Softmax regression technology. And because training
requires large data sets, data enhancements are used to enhance data size in improved
CNN. When data is input, the data is usually packaged and input. This operation is
called batch processing. Batch processing can greatly shorten the processing time of
each image and improve the computing speed of the computer.
Research on System Architecture Based on Deep Learning …
515
Google Net, Res Net, spatial transformer module, capsule network, and a comparative
analysis of the various improved models for deep learning. Finally, the problems that
still need to be solved in this field and the future of deep learning are introduced.
2 The Composition and Principle of Neural Network
The basic neural network consists of input layer, hidden layer, and output layer. The
structure is shown in Fig. 1. For the classification problem, the main task of the input
layer is to read the information of the image, the number of neurons is the number of
pixels of the input image, and the number of neurons in the output layer is the number
of categories. For example, if the MNIST handwritten digital image set is a 28pixel  28-pixel grayscale image, the number of neurons in the input layer of the
neural network is set to 28 Â 28 = 784 neurons, and the 10 neurons in the output layer
are derived from 10 categories. (Numbers 0 to 9 for a total of 10 categories) The
activation function converts the sum of the input signals into an output signal. The
activation function mostly uses Relu, and the activation function of the output layer
uses the Softmax function. The convolutional neural network also includes a convolutional layer, a pooled layer, and a fully connected layer. The convolutional layer is
used for feature extraction, and the convolution operation is equivalent to the filter
operation in image processing. Each time a feature is perceived locally in the image and
then synthesized at a higher level to obtain information about the entire image. And in
order to improve efficiency, parameter sharing and local sensing are introduced [1]. The
pooling layer is an operation for reducing the space in the high and long directions and
is disposed behind the convolution layer. This operation can reduce the amount of data
while retaining valid information. Max pooling is generally used in identifying problems. After using multiple convolutional layers and pooling layers, the fully connected
layer is fully connected to all neurons in the previous layer, integrating local information, and linking all features. Convolutional neural networks as the hierarchy deepen, the extracted information is increasingly complex and abstract, from the initial
layer to the subsequent layer in turn respond to image edges, textures, and more
complex object components. In the improved CNN model, the activation function main
uses the ReLu function. Its expression is as follows hðxÞ ¼
x ðx [ 0Þ
0 ðx 0Þ
When the
input is greater than 0, the value is directly output; when the input is less or equal to
zero, 0 is output. Most use dropout technology, and use GPU parallel training technology, the classifier uses Softmax regression technology. And because training
requires large data sets, data enhancements are used to enhance data size in improved
CNN. When data is input, the data is usually packaged and input. This operation is
called batch processing. Batch processing can greatly shorten the processing time of
each image and improve the computing speed of the computer.
Research on System Architecture Based on Deep Learning …
515