Knowledge increasable neural network based on self-containing neural modules

Today´s neural network systems have serious limitations, of which scalability and reliability are the most serious. Also, the one-shot training procedure prevents the network from progressive learning like that of human brains. Destruction of learned knowledge on learning new knowledge and network knowledge capacity limitation are key factors in this. For the problem, Luo (2000) proposed the idea of knowledge inheritance and accumulation for a knowledge increasable artificial neural network (KIANN). We propose a building block for KIANNs. The building block is a self-containing neural network, which consists of a feedforward network and an assessing component based on error correcting coding theory. In this way, a network can not only learn, but also know what it learnt with confidence. We have applied it to character recognition and good performance is achieved.