Gradually understand the deep belief network

1. Deep belief network （DBN） It's through layer by layer training , Solve the optimization problem of deep neural network , Through layer by layer training, a better initial weight is given to the whole network , So that the network can reach the optimal solution as long as it is fine tuned .
2. Training DBN The most important thing is “ Limited Boltzmann machine ”（RBM）
   1. Boltzmann machine （BM）：
      1. It is a kind of stochastic neural network , There are only two states of neurons in the network （ not active , Activate ）, In binary 0,1 Express , The value of the state is determined according to the law of probability and statistics . The key ： When the energy is minimum , The network is the most stable , At this time, the network is optimal .
      2. BM It is a feedback neural network composed of random neurons , Symmetrical connection , From the visible layer 、 Hidden layer composition , It can be seen as an undirected graph ：（x1、x2、x3 Is the visible layer ,x4、x5、x6 For hidden layer ）
      3. Boltzmann machine （BM） It can be used in supervised learning and unsupervised learning . In supervised learning , Visible variables can be divided into input and output variables , Implicit variables implicitly describe the complex constraint relationship between visible variables . In unsupervised learning , Implicit variables can be regarded as the internal feature representation of visible variables , Be able to learn complex rules in data . The price of Boltzmann machine is that the training time is very long, very long .
   2. Limited Boltzmann machine （RBM）：
      1. It's right BM simplified , Original BM The visible and hidden elements of are fully connected , And hidden elements and hidden elements are also fully connected , It increases the amount and difficulty of calculation
      2. RBM There is also a visible layer , A hidden layer , But there is no connection in the layer , Layer to layer full connection , It's a bipartite graph , The feeling is full connection ：
      3. RBM Essence is the sharp weapon of unsupervised learning , It can be used for dimensionality reduction （ Less hidden layer settings ）、 Learning to extract features （ Hidden layer output is the feature ）, Self encoder （autoencoder) And deep belief network （ Multiple RBM）
      4. RBM Express the data as a probability model through learning , Once the model is trained or converged to a stable state through unsupervised learning , It can also be used to generate new data .
3. Deep belief network ：

os： It feels like the predecessor of the full connection layer , The above is just the perspective gradually proposed by previous people to learn .