Feature analysis by neuronal self-regulation

We propose a new learning paradigm for neural networks and apply it to solving the subspace decomposition problem for feature analysis. In this proposed network, each neuron learns about the environment through a process of self-regulation which actively controls the neuron's own learning by perceiving its status in the overall learning effectiveness. Based on this concept of self-regulation, we derive the primary learning rules of the synaptic adaptation in the network. A self-regulative neural network is utilized to explore significant features of the environmental data in an unsupervised way and to implement subspace decomposition of the data space. Numerical simulations demonstrate the efficiency of the learning model and verify the practicability of the concept of individual neuronal self-regulation for learning control.