Jonathan KERN Thesis defense

Signal processing and machine learning algorithms are now at the central stage of our digital world, from navigation to digital communication, or even health monitoring. This has led to a strong increase in the global energy consumption of the Information and Communication Technology sector. To address this challenge, we explore reducing energy consumption in these algorithms by using energy-efficient, but unreliable, memories. In this regard, we develop an energy-reduction methodology applicable to different algorithms and memory technologies, starting with a study of Kalman filtering using voltage-scaled SRAM.We then investigate deep neural networks using in-memory computing architectures based on resistive memories also known as memristors. Using our theoretical analysis, we propose equations linking the algorithm's performance to the parameters controlling the memory's error level and energy consumption. We formulate and solve an optimization problem to find the optimal set of parameters that minimize the energy usage of the memory while satisfying a performance constraint. Our results show that this approach can lead to energy gains, with a gain of up to 50% for the Kalman filter case.