Today, energy consumption is the unavoidable optimization factor, which can not be neglected anymore in the quest of ever highly performant, portable, robust and communicating embedded systems. The goal of this thesis is the development of integrated systems, which are enough energetically efficient to operate without the use of a battery. The proposed approach we intend in this thesis is based on event-based signals and, more generally, on event-based signal processing. Many studies haves already been done within the TIMA laboratory in the field of innovative integrated architectures and robust ultra-low power asynchronous circuits. Techniques coupling non-uniform sampling schemes and asynchronous integrated systems have already shown a real potential for ultra-low power. Other techniques based on approximate computing have already shown their potential to reduce consumption, especially in applications that have intrinsic resilient error properties or that does not require a unique and
accurate result. The scientific thesis objectives consist in combining these two approaches in order to obtain a computational unit able to dynamically adjust its accuracy.