11. Basic concepts of neural network

One 、 Neuron

• Neuron Is the neural network algorithm Basic unit , Its It is essentially a function , Accept external stimulation and generate corresponding output according to input

• The internal structure of neurons can be seen as Linear function and activation function The combination of , The result of linear function operation is passed to the activation function , Finally, the output of the neuron is produced

• Typical neurons are perceptron and S Type neurons

1.1 perceptron

• perceptron It is sometimes called a perceptron , It's an artificial neural network ; Regarded as a form, the simplest forward is artificial neural network ,

• perceptron It is a binary linear classifier , It is mainly used for Solve classification problems

• The perceptron accepts multiple binary inputs and generates a binary output

• How the perceptron works ：

  1. The perceptron accepts multiple binary inputs , Each input corresponds to a weight

  2. The weighted value of the binary input of the perceptron has a significant impact on the output

  3. Compare the weighted value of the perceptron with the threshold , Determine the final binary output value

     The above process can be expressed in the following algebraic form ：

     

  • Sometimes for simplicity , Write the perceptron rule in the following general form ：

    

    • If the set offset b more , Then the final output is 1 Relatively easy
    • If the set offset b smaller , Even the larger plural , The final output 1 It is more difficult
    • You can adjust the output of the sensor by setting different weights and offsets

1.2 S Type neurons

• S Type neurons Compared with perceptron , The advantage is ： Small changes in weight and bias will only lead to small changes in output

• S Type neurons The biggest difference with perceptron lies in its Input and output are no longer binary discrete values , It is 0~1 Continuous values of

• S The difference between type a neurons and perceptrons is ：S Type a neuron is a smooth function , The perceptron is a step function . in other words , The perceptron can only output 0 perhaps 1, and S Type a neurons can output 0～1 Any value of

• S Type neurons Characteristics ：

  • S Type a neurons have multiple input values , These input values are 0~1 Any value of
  • S The input weights of type a neurons are activated sigmoid After the function , Output 0~1 The numerical

• S Type neurons The expression of ：

  

Two 、 neural network

2.1 The concept of neural networks

• The connection and interaction between neurons form neural network
• neural network It is a reticular structure formed by simple neurons connected with each other , Change the strength of the connection by adjusting the weight value of the connection , Then realize perceptual judgment
• The most basic 、 The most typical neural network is —— Multilayer perceptron （MLP）
  • Be careful : Although it is called multilayer perceptron , But actually it's made up of S Type neurons Composed of , It is not composed of perceptron

2.2 The structure of neural networks

• A typical neural network structure includes three layers ： Input layer 、 Hidden layer 、 Output layer

  

  • On the far left is the input layer , The neurons are called input neurons
  • On the far right is the output layer , Among them, neurons are called output neurons
  • The middle layer between the input layer and the output layer is the hidden layer , There may be more than one hidden layer in the neural network , Any non input layer and non output layer are called hidden layers

• Input layer : The first layer of neural network , This layer receives input signals （ value ） And pass it to the next layer , For the input signal （ value ） Does not perform any operations , There is no own weight value and offset value

• Hidden layer ： Hidden layer is the composite layer between input layer and output layer of neural network , Neurons in the hidden layer switch layer by layer , Finally, pass the hidden value to the output layer

• Output layer ： It is the last layer of neural network , Receive the input of the last hidden layer and generate the final prediction result , Get the value of the expected number in the ideal range . There can be only one neuron in this layer , It can also be as much as the result

• Be careful ：

  • Each neural network consists of an input layer and an output layer , And the number of neurons in the input layer = The number of input variables contained in the processing object data , The number of neurons in the output layer = Expected number of outputs
  • However, the number of hidden layers and the number of neurons need to be manually set , This is also a difficulty of neural network

2.3 Basic classification of Neural Networks

• Neural networks include ： Forward neural network 、 Common neural network models such as reverse neural network and self-organizing neural network
• Forward neural network ： It is a one-way multi-layer network structure , That is, information starts from the input layer , Pass layer by layer in one direction , Until the end of the output layer .
• Reverse neural network ： Compared with forward neural network , There is feedback between neurons in the feedback neural network , It can be represented by an undirected complete graph
• Self organizing neural network ： also called Kohonen The Internet , The characteristic of the neural network is that when it receives external signal stimulation , Different areas automatically produce different responses to signals . This neural network was first discovered on biological neurons , If neurons are synchronously active, the signal intensifies , If asynchronism is active, the signal weakens .