Learning algorithm based on moderationism for multi-layer neural networks

We have proposed the idea of moderationism, which is the hypothesis that neurons try to moderate both their inputs and outputs against incoming stimuli. In this paper, we apply the moderationism concept to the input and output of neural networks, and present an unsupervised learning algorithm for multilayer networks. Then we apply it to a simple system containing a feedback loop and an artificial arm with 3-layer network. Within our algorithm, the neural networks adapt to the changeful environment but do not adapt to the changeless one. This point means learning no vain action, and it is reasonable and interesting.