Fault tolerant training algorithm for multi-layer neural networks focused on hidden unit activities

We propose a new training algorithm that enhances fault tolerance of multi-layer neural networks (MLNs). Faults mean physical defects or noise in MLNs. Some studies on fault tolerance pointed out that faults on the connections that connected to an output unit bring worse damage than other faults, and proposed training algorithms that enhance fault tolerance of MLNs based on this idea. In this paper, we reveal that it is not always true. Based on this idea, we improved our previous method (weight minimization algorithm).