Safety critical neural networks

There is a lack of trusted techniques which will allow ANNs to be used in safety critical systems. There is a lack of rigid specification methods for ANNs. By “rigid” we mean “not open to misinterpretation”. Research in formal methods is currently very fashionable. Informal methods are well established in industry. However, no similar methods exist for neural networks. It is tempting to believe that, should a rigid specification method be devised for neural networks, that proving conformance to the specification should be relatively easy when compared with software systems. The justification for this is based in the simplicity of the ANN compared to software, and the availability of appropriate statistical methods. Hierarchical methods for the design of ANNs are not sufficiently advanced. These are highly useful in conventional systems because they allow an initially highly complex design to be partitioned into components. Although an embedded ANN may be considered as a component it is difficult to decompose it into useful subcomponents. Therefore it may remain highly complex and difficult to analyse. The nature of problem domains in which neural networks are applied tend to have ill-defined solutions with respect to formal descriptive techniques, and hence existing verification methods are unlikely to be successful. Instead, more complete methods of testing may be required. Reliability can be attained via standard methods where an ANN is emulated. Additional protection at the computational level is also possible given suitable ANN architectures, hardware and training techniques