Testability of artificial neural networks: A behavioral approach

Testability analysis of neural architectures can be performed at a very high abstraction level on the computational paradigm. In this paper, we consider the case of feed-forward multi-layered neural networks. We introduce a behavioral error model which allows good mapping of the physical faults in widely different implementations. Conditions for error controllability, observability and global testability are analytically derived; their purpose is that of verifying whether it is possible to excite all modeled errors and to propagate the error''s effects to the primary outputs (actual test vectors being then technological-dependent). Mapping of physical faults onto behavioral errors is performed for some representative, architectures