Integration of knowledge acquired by different neural networks

摘要

The author describes a set of experiments on decomposing a problem into smaller ones, training a network for each smaller problem and integrating the learned weight settings into a system capable of solving the original problem. Several network structures are suggested and performance comparisons are made. Integration of knowledge acquired by different neural networks not only can reduce the training time, but also can provide other benefits like ease of modification and possible incorporation of domain knowledge.