A general principle adopted by most of these techniques is the decomposition of the problem into smaller problems that are easier to handle.
For each smaller problem, a single node is employed. Its parameters are determined either iteratively using appropriate learning algorithms.
From the way these algorithms build the network, they are sometimes referred to as constructive techniques. (查看原文)
The other direction we will follow to design a multilayer perceptron is to fix the architecture and compute its synaptic parameters so as to minimize an appropriate cost function of its output.
This method not only overcomes the drawback of the resulting large networks of the previous section but also makes these networks powerful tools for a number of other applications, beyond pattern recognition.
However, this method is confronted with a serious difficulty. This is discontinuity of the step(activation) function, prohibiting differentiation with respect to the unknown parameters(synaptic weights). (查看原文)
木子三水
电子工业出版社2010版
还没人写过短评呢