PhD Position F/ M Dimensioning probabilistic embedded systems for efficient execution of artificial intelligence algorithms

Contexte et atouts du poste

The PhD thesis is funded by the Paris region program and it is hosted by the Kopernic team in Paris (see more details at 

Travelling is expected in France and in Bresil as well as EU countries, the associated costs being covered following the current public laws. Inria offers an equal opportunity and friendly working environnement, while covering partially the transport and meal costs. AGOS (its commité d'entreprise) provides financial support for holidays or jobbies. 

Mission confiée

The arrival of artificial intelligence methods in the embedded systems area pushes for the inclusion of complex computations in presence of critical constraints like time or energy. For example, in an autonomous vehicle, understanding the impact of automatic recognition of a pedestrian on the reaction in time of that vehicle is an open problem.

In order to perform these complex calculations within a reasonable time delay, designers are integrating multiple cores processors within more hybrid architectures such as CPU-GPU or CPU-FPGA. Although hybrid architectures increase computing capacities, the time validation of the execution of programs running on these architectures is an open problem, especially if communication delays are considered. Within the Kopernic team we propose combining probabilistic and non-probabilistic models to deal with such validations.

The worst-case execution time (WCET) and the worst-case response time are important parameters in the time validation of real-time critical systems because they allow to verify if a program, combined with other programs, can be implemented on a processor while respecting strict time constraints. The WCET can be estimated either by static analysis methods, or by measurement-based methods, or by a combination of both approaches [1]. During this thesis, measurement-based statistical approaches are considered as well as methods combining analytical solutions to these approaches. Depending on this estimation, the response time calculation methods can be analytical or measurement-based. The objective of the thesis is to propose efficient scheduling algorithms of probabilistic embedded systems on hybrid architectures, to compare their energy performances wrt existing non-probabilistic algorithms, while respecting the time constraints. All results are illustrated on the Kopernic benchmarks - KDBench (see 

Principales activités

The thesis is expected to cover the following main activitivies :

All results are expected to be published within real-time conferences and journals. 

Compétences

Technical skills and level required : background on real-time systems is an avantage, but not necessary, while back ground on embedded system is mandatory. Python is the main programming language, but being familiar with C/C++ code is expected. 

Languages : English and French 

Avantages