Set the number of neurons in the output layer
The number of neurons in the output layer is related to the number of categories of the training set samples. If you don't know the number of categories, set the number of nodes as much as possible. Too fine classification can appropriately reduce the output nodes and reduce the dead nodes that have never updated the weights.
Design the arrangement of output layer nodes
The arrangement of output layer nodes depends on actual needs. The arrangement is as intuitive as possible to reflect the physical meaning of the actual problem. For example, for general classification problems, an output node represents a pattern class, and the structure of a one-dimensional linear array is simple and clear. For color and path problems, a two-dimensional plane is simple and intuitive.
Initialize weight
It can be initialized randomly, try to ensure that the initial position of the weight value coincides with the approximate distribution area of the input sample, and avoid a large number of dead nodes. A simple method is to randomly select m input samples from the training set as the initial weight.
Design topology area
The design principle of the topological domain is to keep the domain shrinking so that the weight vectors corresponding to adjacent neurons on the output plane are both different and similar. It is guaranteed that when the winning node produces the greatest response to a certain type of pattern, its domain node can also produce a larger response. The field can be square, hexagon, diamond, etc.
Design learning rate
The learning rate should be a decreasing function. At the beginning of training, the learning rate can be a larger value, and then decrease at a faster rate, which is conducive to quickly fitting the approximate input vector structure. Then the learning rate slowly drops to 0 at a smaller value, so that the weight can be finely adjusted to conform to the sample distribution structure of the input space .