A novel dynamic structural neural network with neuro-regeneration and neuro-degeneration

This paper presents a novel dynamic structural neural network (DSNN) and a learning algorithm for training DSNN. The performance of a neural network system depends on several factors. In that, the architecture of a neural network plays an important role. The objective of the developing DSNN is to avoid trial-and-error process for designing a neural network system. The architecture of DSNN consists of a three-dimensional set of neurons with input/output nodes and connection weights. Designers can define the maximum connection number of each neuron. Moreover, designers can manually deploy neurons in a virtual 3D space, or randomly generate the system structure by the proposed learning algorithm. This work also develops an automatic restructuring algorithm integrated in the proposed learning algorithm to improve the system performance. Due to the novel dynamic structure of DSNN and the restructuring algorithm, the design of DSNN is fast and convenient. Furthermore, DSNN is implemented in C++ with man-machine interactive procedures and tested on many cases with promising results.