The role of each layer in convolutional neural network

A complete neural network model often contains convolution layer 、 Pooling layer 、 Fully connected layer 、 Output layer . But the specific role and meaning of each layer , Sometimes I don't understand myself , Here is the relevant information , Take a brief note of , Hope to be helpful to relevant students .

1. Convolution layer

The neurons of the convolution layer feature map are connected with some neurons of the previous layer feature map through a group of filters or weight matrices , The locally connected region is also called the acceptance domain . After convolution of the neuron above the receiving threshold and the weight matrix , The characteristic diagram of this layer is generated through nonlinear excitation , And as input to the next level . Convolution , All accepted domains on the same characteristic graph share a set of weight matrices , Called weight sharing . Different weight matrices are used for different characteristic graphs of the same layer network , The number of characteristic graphs can also be understood as the number of channels . Each set of weight matrix detects the specific characteristics of the input data , therefore , Each feature map represents a specific feature at different locations of the previous layer . One of the advantages of local connection and weight sharing greatly reduces the free parameters of the network , To some extent, the over fitting of the network is avoided , At the same time, reduce the storage capacity . This convolution structure is based on the spatial correlation of image data and the displacement invariance of target features . let me put it another way , If a feature appears in a part of the image , Then it can also appear anywhere else . This explains why neurons at different locations share a weight matrix to detect image features .

2. Pooling layer

The function of the pooling layer is to fuse the same features detected by the convolution layer . Each characteristic graph of convolution layer is divided into several local slices , The pooling function calculates the statistical characteristics of each slice . The number of characteristic graphs of pooling layer is the same as that of convolution layer . The two common pooling methods are maximum pooling and average pooling , That is, take the maximum or average value of the slice as the input of the pool layer characteristic graph . therefore , The pooling layer is also called the lower sampling layer . In addition to reducing the dimension of the feature graph , Pooling is invariant to small-scale displacement and distortion of features . When the focus is on the feature itself rather than its location , Displacement invariance is a good characteristic .

3. Fully connected layer

The top of a convolutional neural network usually contains one or more fully connected layers . Different from convolution layer , The neurons of the whole connective layer line up , These neurons are interconnected with the previous layer of neurons through weights , In a fully connected structure . The number of layers of the whole connection layer and the number of neurons in each layer are not fixed . Usually the higher the number of layers , The fewer neurons .

4. Output layer

After multi-layer feature extraction , The last output layer can be regarded as a classifier , Predict the category of input samples .
notes ： The article is excerpted from 《 Synthetic aperture radar intelligent interpretation 》 Xu Fengzhu