Immunization of neural networks against hardware faults

A methodology and set of computer-aided design tools for measuring and improving the fault tolerance characteristics of neural networks are presented. Two analysis programs developed using realistic fault networks appropriate for emulating potential hardware failures are discussed. It is demonstrated how functionally identical neural networks can have significantly different reliability characteristics. Criteria for selection of an optimal architecture and trained state are discussed. A modified training strategy which reduces the network´s sensitivity to faults is introduced. The process involves the deliberate injection of faults into the network during its training phase. The proposed scheme is analogous to viral immunization in the biological domain because it is the neural network´s own adaptive capability which is utilized to improve its fault tolerance characteristics