8 - Deep Learning of Neural Networks Core Principles and Algorithms - Convolutional Neural Networks
convolutional neural network(CNN)
Neural networks are not a specific algorithm， Rather, it is a model building idea or way of。
We have previously successfully builtBP web，BP Each neuron of the network is composed of a linear classifier+ The excitation function formed by。 Each neuron is connected front-to-back and back-to-tail， Form a network structure。
BP Limitations of the network。 Is it possible to do some new logic unit or build way。 The vast majority of pattern recognition draws on the key components of convolutional neural networks。
Comparison with fully connected networks
fully connected network(Full Connect Network)
The in and out of each neuron node is derived from the output of each neuron in the previous layer。( full connection)
advantages: Information from each dimension is propagated to any subsequent node。( Let the network not miss factors that contribute to this dimension) drawback: w harmonyb Extraordinarily.， update( Too much weight.) slow， slow convergence
It is not very feasible to use fully connected networks for image recognition which has a very high input dimension。 Because of the computer's lack of computing power。
Convolutional neural networks are likewise a feed-forward neural network。
The neurons of a convolutional neural network can respond to a portion of the coverage of the surrounding units。 Great for large-scale pattern recognition， Good performance， Especially for large-scale image processing efficiency， Convolutional neural network it is very efficient。
Two major features:
- Has at least one convolutional layer， Used to extract features
- Convolutional layers share by weights， Reduction of weightsw Number of， Faster convergence。
After the weighting has been reduced， This way the convolutional neural network will be faster thanBP Neural network convergence。
Convolutional networks are mainly used to identify displacement， resizing， and other forms of twisted invariant two-dimensional graphs。 Then since the larger features we introduced earlier。 Convolutional networks learn through training data at the feature detection layer。
So when using convolutional neural networks， You can then avoid the display to extract features。 And it is possible to learn implicitly from the training data。
furthermore， Since neurons on the same feature mapping level， It uses a weight sharing approach， So the network can learn in parallel， This also makes convolutional neural networks， Relative to a network of neurons connected to each other， With greater advantages。
Recommended>>
1、Guangzhou Daily Smart doctors can see eyes as well as lungs City Womens and Childrens Medical Center research results in top international medical journals giving AI systems the ability to migrate and learn
2、Advantages and disadvantages of digital supply chains compared to traditional supply chains
3、Chen Weixing dislikes Zhu Xiaohu again and the wind and fire haunt the Lantern Festival
4、Europe calls for alcohol lockout system to end drink driving vomit
5、Write a tracking motion object system in Python