Extreme Learning Machine pour réseau de neurones embarqué
Vision sensor, CNN, ELM, Random matrices, software optimization
Based on the literature, this work will study the possibilities of deriving from an existent network an approximate one with a set of learned parameters and a set of pseudo-random parameters in the intermediate layers.
The proximity to the sensors and the specifics of the intended application will be essential assets for meet the constraints of memory, computing power and consumption while preserving the better detection or classification performance.
This point is important as the methodology put in place should facilitate the adaptation of the network to different sensors or to development databases. The effectiveness of the solution will be evaluated, not only in terms of the number of MAC operations, but also in terms of suitability for the architecture of the sensors.
Thesis director: Stéphane MANCINI
Thesis supervisor: Stefanie HAHMANN (LJK)
Thesis started on: Aug. 2020
Doctoral school: MSTII