Fault Tolerance Improvement through Architecture Change in Artificial Neural Networks

摘要: This paper presents a technique for improving the fault tolerance capability of Artificial Neural Networks. characteristic distributed systems, which is usually pointed out as one advantages this structure hasn't been deeply studied and can be improved in most networks. The solution implemented here consists changing architecture feedforward artificial neural networks after training stage while maintaining its output unchanged. It involves evaluating elements Network are more sensible to duplicating inputs, bias, weights or neurons, according evaluation done before. very interesting because it allows pre-trained network, but cost need additional hardware resources implement same network. also an example application illustrate effectiveness.