Dot product operation

Hadamard Product Sum

cross-correlation operation

Filter: represents a certain feature of the image

The depth of the convolution kernel must be consistent with the input

Number of channels: number of convolution kernels

Convolution kernel sliding time interval

Dot product operation using sliding window

Compared with convolution, only the flipping of the convolution kernel is omitted

Sparse interaction, local feeling

weight sharing

translation invariance

Downsampling: feature selection, reducing the number of features, thereby reducing the number of parameters

Reduce feature dimensions to avoid overfitting

Max Pooling: Texture Extraction

Average pooling: background preservation

Convolutional layer:

Pooling layer:

Fully connected layer:

summary

forward propagation

loss function

Backpropagation

Update weights

Depth of the network

The number of convolution kernels

Convolution kernel size

hyperparameters

deconvolution

Transpose: mapping from low dimension to high dimension

General: high-dimensional to low-dimensional mapping

Affine transformation

How to increase the receptive field of the output unit

Insert hole

Handwritten digit recognition

Network characteristics

Network introduction

Network introduction

Network structure

Network characteristics

solved problem

Network introduction

advantage