8 floors in total

Learn structural features from data that are meaningful for classification

Describe structural information in the input image

The description results are stored in 256 6x6 feature response maps

Change the previous convolution kernel to smaller size

Increase the step size to successfully extract features

Increase the number of late convolution kernels

13 convolutional layers

3 fully connected layers

Use smaller 3x3 convolution kernels in series to obtain a larger receptive field

Deeper depth, stronger nonlinearity, and fewer network parameters

Remove the LRN layer of AlexNet

Advantages of small convolution kernel

Reasons for convolution to 512

Propose the inception structure, which can retain more characteristic information of the input signal

Remove the fully connected layer and use average pooling, resulting in only 5 million parameters, 12 times less than AlexNet.

An auxiliary classifier is introduced in the middle of the network to overcome the vanishing gradient problem during training.

The difference between average pooling vectorization and direct expansion vectorization

Deeper networks appear to be inferior to shallower networks

reason

Proposed residual module

proposed batch normalization

Propose an initialization method for ReLU activation function

can be regarded as an integrated model