> Academic opportunities > PHD positions

Quantified DNN learning algorithms with limited hardware overhead for Edge implementation

Technological challenge: Artificial intelligence & Data intelligence (learn more)

Department: Département Systèmes et Circuits Intégrés Numériques (LIST)

Laboratory: Laboratoire Systèmes-sur-puce et Technologies Avancées

Start Date: 01-09-2022

Location: Grenoble

CEA Code: SL-DRT-22-0348

Contact: thomas.mesquida@cea.fr

Intelligence at the Edge aims to push the computation performed on the data to the edge for energy and security reasons. This results in the implementation of co-optimized hardware for the inference of artificial neural networks (ANN) of varying depths (DNN) and aims at a calculation as close as possible to the creation of useful information. The supported networks are trained offline and their parameters exported to the support. Two main aspects are necessary for this hardware to adapt to a particular/peculiar environment or to refine its knowledge: the direction of learning, or how to define the target to reach, and the optimization of the network parameters, or how to minimize the error with respect to the defined target. The most commonly used learning algorithms have the drawback of requiring a much larger amount of memory than during the inference phases. Indeed, all the intermediate results of the network must be stored so that the gradient back-propagation can be performed. The extra cost of learning compared to pure inference is consequent and the goal of this PhD is to minimize it in the framework of quantized artificial neural networks. The goal of this PhD is to propose, implement and validate DNN learning algorithms by optimizing the associated memory and energy requirements. In addition to this, there is the constraint of strong quantification of DNN parameters for frugal inference, which is not yet taken into account in the literature in this context. These algorithmic studies can be integrated into the laboratory's hardware simulation platforms.

See all positions Download the offer (.zip)

Email Bookmark and Share