Cnn-convolutional Neural Networks
In recent years in the field of machine vision is a very fire of acquiescence, first proposed by Yan LeCun.
If you want to learn the details, see Li Feifei cs231n course
How does it work?
Give a picture, each circle is responsible for processing part of the picture.
These circles form a filter.
Filter identifies whether the specified pattern exists in the picture and in which region.
There are 4 filter in the same color, the different points of the filter is responsible for different areas of the picture.
Neurons use convolution techniques to find patterns, simply to understand whether a filter is used to find out if a picture has some form of pattern.
Weights and bias play an important role in the effect of the model.
Change the white circle to a neuron, which is what CNN looks like.
There is no connection between neurons in the convolution layer, each of which is connected only to inputs.
Neurons in the same layer use the same weights and bias, so that neurons in the same layer can crawl the same pattern, but in different areas of the image.
Next is the Relu (rectified Linear Unit) layer and the pooling layer, which are used to construct the pattern found by the convolution layer.
CNN also uses back propagation training, so there is also the possibility of vanishing gradient. and Relu as an activation function, gradients will generally maintain the appearance of constant value, so that it will not be in the key layer has a significant decline.
The pooling layer is used for dimensionality reduction.
After the convolution of the relu, there will be more and more complex forms, so the pooling layer is responsible for extracting the most important patterns, thus improving the efficiency of time space.
These three layers can extract useful pattern, but they don't know what the pattern is.
So then there is the fully connected layer, which can classify the data.
A typical deep CNN consists of several sets of convolution-relu-pooling layers.
But CNN also has a drawback because it is supervised learning, so a lot of tagged data is needed.
Basic knowledge of CNN