Introduction to convolutional neural network

introduce

Using fully connected neural network Handwritten digit recognition A high accuracy has been achieved , But this can be improved .
First, let's look at the problems of fully connected neural networks ： It ignores structure , When inputting pictures into neural network , It directly converts it into a column , Ignore the relationship between pixels and other pixels in position .
In this simple handwritten numeral recognition , Ignoring this structural relationship can still maintain a good accuracy , But for more complex image recognition , Ignoring this structural relationship often leads to poor results . Therefore, a new neural network is needed to solve this problem .

Convolutional neural networks

Convolutional neural networks （Convolutional Neural Networks, CNN） It is a kind of feedforward neural network with convolution calculation and depth structure （Feedforward Neural Networks）, It's deep learning （deep learning） One of the representative algorithms of .

The difference with full connection

Compared with full connection , Convolutional neural networks use Convolution layer And Pooling layer Process the features of the image , Then output the result with full connection .
Convolution layer needs to learn parameters , Convolution check with different parameters, different images will make different responses .
The pooling layer does not need to learn parameters , Generally speaking, it takes the maximum value or average value . But the existence of pooling layer can greatly increase the robustness of neural network , Because even if there are some small changes , It has little effect on the output of the pooling layer .

Why still use full connection

Because in most cases , We don't know the relationship between these parameters , And full connection provides a good platform .
It assumes that all parameters are related , Through constant learning , The edge power of some relationships will continue to grow , It shows that this relationship is real ; But the edge power of some relationships will continue to decrease , It shows that this relationship does not exist or is very weak , It won't have much impact on the final result .