Extending the compressive statistical learning framework : quantization, privacy and beyond
(2021)
Files
thesis.pdf
Open Access - Adobe PDF
- 12.28 MB
Details
- Authors
- Supervisors
- Jacques, Laurent
- Abstract
- Over the last few years, machine learning–the discipline of automatically fitting mathematical models or rules from data–revolutionized science, engineering, and our society. This revolution is powered by the ever-increasing amounts of digitally recorded data, which are growing at an exponential rate. However, these advances do not come for free, as they incur important computational costs, such as memory requirements, execution time, or energy consumption. To reconcile learning from large-scale data with a reasoned use of computational resources, it seems crucial to research new learning paradigms. A particularly promising candidate is the compressive statistical learning framework. In a nutshell, the idea of this method is to first compress the learning data, in an efficient manner, as lightweight sketch vector (given by random feature moments of the data). The desired learning methods are then carried out using only this sketch, instead of the full dataset, which can sometimes save orders of magnitude of computational resources. This thesis broadens the scope of the compressive learning framework by exploring three extensions of it. First, the quantization of sketch contributions is studied, which allows to further reduce the computational burden associated with computing the sketch. Second, the addition of a privacy-protecting layer on top of the sketch is considered, which allows to learn from the sketch while ensuring the privacy of the data contributors. Finally, generalizations of the framework to novel tasks are discussed.
- Affiliations
UCLouvain SST/ICTM - Institute of Information and Communication Technologies, Electronics and Applied Mathematics UCLouvain SST/ICTM/ELEN - Pôle en ingénierie électrique UCLouvain SST/ICTM/INGI - Pôle en ingénierie informatique UCLouvain SST/ICTM/INMA - Pôle en ingénierie mathématique
Citations
Schellekens, V. (2021). Extending the compressive statistical learning framework : quantization, privacy and beyond. https://hdl.handle.net/2078.5/165616