Constructing knowledge increasable neural network system via vicinal performance

Inheriting the learned knowledge of existing neural networks is a difficult problem. Knowledge increasable neural network system (KINNS) is a parallel neural computation model for this purpose, which consists of multiple neural units and is architecturally scalable. In this paper, we propose a vicinal performance approach for constructing KINNS. The method makes use of performance information of neural units on the vicinal region of the input and selects appropriate neural network(s) for processing. Empirical results on function approximation demonstrate its good performance and feasibility for KINNS. This is meaningful for utilizing the learned knowledge of existing neural networks and for large scale parallel processing.