The Associate Teams program is a Bilateral Partnership aiming to strengthen the scientific potential of teams by developing collaborations between Inria and high-level research teams from around the world.
The program is based on a joint research project and supports the mobility of senior researchers, engineers, post-doctoral fellows, PhD students and trainees involved in the team in France and in the partner country. Workshops and working groups on joint research are also organized.
An Associated Team has a life span of three years.
The aim of AxTRADE Associate Team is to design and develop techniques to generate DNN models capable of dynamically adjusting in real-time to adapt to changes in computational quality provided by the underlying hardware platform. The quality level will change depending on the environmental conditions in which the system operates. A naive approach would consist in generating as many DNN variants (i.e., sets of parameters) as the available quality levels. However, this would mean duplicating the necessary storage/power/performance overhead needed to dynamically switch among different models. The challenge lies in generating such variants so that the hardware overhead (in terms of memory, performance, energy consumption, silicon area) needed for the DNN to dynamically adjust is minimized and the quality of the delivered service is maximized. Specifically, our objective is to develop strategies and methodologies to obtain variants of a given DNN model (i.e., different sets of parameters) while minimising the number of required parameter updates to switch from a given variant to another in real-time, thus minimizing the needed storage, memory footprint, power consumption, and delay. The final goal is maintaining a high output accuracy while substantially reducing the associated hardware overheads.
AxTRADE has been created in 2024.
Partnership
Marcello Traiola
Inria / IRISA, Rennes, France
