PhD opportunities

Mapping of convolutional neural networks onto hybrid architectures under resource constraints

Thesis proposal

Area of expertiseInformatique temps réel, robotique et automatique - Fontainebleau
Doctoral SchoolISMME - Ingénierie des Systèmes, Matériaux, Mécanique, Énergétique
SupervisorANCOURT Corinne
Research unitMathématiques et Systèmes
KeywordsCompilation, Neural networks, Code optimization, Programming languages
AbstractCNNs are now widely used for image recognition/segmentation, object detection and are executed on a large variety of architectures. Optimizing their mapping on memory-limited architectures, such as drones and robots, is essential. The efficient execution of CNNs on these architectures, during the learning or inference phases, is only possible by adapting their execution to the available hardware resources: parallel components and memories.
When it is necessary to respect the memory constraint, the classic optimizations of the CNNs concern the reduction of the precision of the data, the reduction of the models, and the adaptation of the hyper-parameters. Some techniques are emerging and focus on inter-layer optimization for particular networks. But as mentioned by the authors of article[1], future work must focus on the optimization at the global level of the application and not only at the level of a layer of the network [2, 3].
Additional research on the mapping of CNNs, at different levels of operation granularity, must be carried out for an implementation of the application adapted to the architectural resources, and more particularly to the memory constraints.
The objective of this thesis is to propose techniques to automatically map CNN-like applications on architectures with constraints due to limited memory resources. We'll focus on the memory aspects, but the same techniques can benefit other optimizations like latency.
ProfileMaster 2 in computer science
Knowledge in compilation, AI, linear programming
FundingConvention CIFRE