[Study notes]
According to the above study, we already know that when we receive the customer's needs, so we do recognize, judge or predict, we need to final delivery to the customer our neural network model. In fact, we painstakingly trained neural network model, that is, from a mysterious unknown function input to output mapping. In most cases, we do not know what the real function, we just try to fit it. Examples given above, only introduction and description of functions, so only a few linear combination (described below). It is not possible approximation to complex functions. How to make us free to approach the complexity of the neural network function? Artificial intelligence, have given the answer, like the higher mathematics of them, we use the Taylor series approximation of a variety of complex functions, like the introduction of non-linear activation function of neural networks can make us free to approach a complex function .
Reprinted from the original article: https://blog.csdn.net/qq_44594249/article/details/100558806