Description
Date depot: 11 mars 2020
Titre: Bayesian Machine Learning on Optical Hardware
Directeur de thèse:
Maurizio FILIPPONE (Eurecom)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Non defini
Resumé:
Artificial Intelligence (AI) has the potential to stimulate changes for the greater good of society and Machine Learning (ML) is a key element to enable this.
While we are experiencing a glimpse of this much anticipated future, there is a growing concern over the alarming figures on power consumption and carbon footprint of ML, which indicate an unsustainable trend.
An emerging trend is the use light and optics principles as alternatives to transistor-based hardware to carry out computations.
Optical Processing Units (OPUs) promise energy-efficient randomized computations at the speed of light, by performing in-hardware multiplications of input vectors with Gaussian random matrices.
While this has the potential to improve the energy efficiency of ML, which largely relies on deterministic computations, the limited set of computations offered by OPUs are so that effective solutions to design and use ML models on OPUs are practically nonexistent.
The overarching ambition is to rethink the design and use of ML models in order to develop the algorithms and the theory necessary to make ML exploit OPUs, and ultimately invert its unsustainable trend.
This project in particular will focus on Bayesian ML, which is an elegant self-contained framework to deal with probabilistic reasoning to aid decision-making, and where randomness is an intrinsic element of model design and inference.
This project will tackle one or more of the following objectives: (i) Energy-efficient large-scale Bayesian ML on OPUs (ii) Energy-efficient Bayesian deep learning on OPUs (iii) Validation of energy efficiency and acceleration in applications.
Doctorant.e: Kozyrskiy Bogdan