A study of cognitive resilience in a JPEG compressor

Many classes of applications are inherently tolerant to errors. One such class are applications designed for a human end user, where the capabilities of the human cognitive system (cognitive resilience) may compensate some of the errors produced by the application. We present a methodology to automatically distinguish between tolerable errors in imaging applications which can be handled by the human cognitive system and severe errors which are perceptible to a human end user. We also introduce an approach to identify non-critical spots in a hardware circuit which should not be hardened against soft errors because errors that occur on these spots are tolerable. We demonstrate that over 50% of flip-flops in a JPEG compressor chip are non-critical and require no hardening.