Foundation and research content of artificial neural network

Artificial neural network is a simulation of the biological nervous system. Its information processing function is determined by the input and output characteristics (activation characteristics) of the network Unit (neuron), the topology of the network (the connection mode of the neuron), the connection weight (synaptic contact strength), and the threshold of the neuron (which can be considered a special connection right).

Compared with the digital computer, the artificial neural network is closer to the human brain in terms of its composing principle and functional features, and it is not the execution of the operation by a given procedure, but the ability to adapt itself to the environment, summarize the law, and complete some kind of operation, identification or process control.

Figure 1 Typical three-layer artificial neural network

1. All nodes in the graph are layered, and each layer of nodes can be directed to the previous layer of nodes, but no arcs are connected between nodes of the same layer, and each node cannot go over a node connected to the upper layer. Theoretically, the number of layers of an artificial neural network can be arbitrary. In practice, no one can design more than five layers of network, because the more layers the network, the more complex the calculation.

2. Each arc has a value (called a weight or weights), based on these values, you can use a very simple formula to calculate the value of the node they are referring to.

The network learning criterion should be: If the network makes the wrong judgment, then through the network study, should make the network to reduce the next time to make the same mistake possibility. First, to the network of the various connection weights to the (0,1) interval of random values, the "A" corresponding to the image mode input to the network, the network input mode weighted sum, compared with the threshold, and then the non-linear operation, to obtain the output of the network. In this case, the probability of the network output being "1" and "0" is each 50%, which means it is completely random. If the output is "1" (the result is correct), the connection weights are increased so that the network encounters the "A" mode input again and can still make the correct judgment.

If the output is "0" (that is, the result is wrong), the network connection weights to reduce the overall input weighted value of the direction of adjustment, the purpose is to make the network next encounter "A" mode input, reduce the likelihood of making the same error. So operation Adjustment, when the network turns into a number of handwritten letters "A", "B" after the internet according to the above learning method after a number of studies, the correct rate of network judgment will be greatly improved. This shows that the network's learning of these two patterns has been successful, and it has distributed the two patterns to the network's various connection weights. When the network encounters any of these patterns again, it is able to make rapid, accurate judgments and recognition. In general, the more neurons the network contains, the more it can remember and recognize the pattern.

Main contents of research and application of artificial neural network

1. Research of artificial neural network model

Neural network prototyping, simulation of the biological characteristics of neurons, etc.

2. Basic theoretical study of neural networks

Nonlinear characteristics of neural networks, stability, convergence, robustness, etc.

3. Application of neural network intelligent Information Processing system

Cognitive and artificial intelligence, optimization and control, signal processing, multi-sensor data fusion, etc.

4. Software simulation and hardware implementation of neural network

Software simulations, neural chips, etc.

5. Realization of neural network computer

Foundation and research content of artificial neural network