Abstract: FCAI, CSC, and NVIDIA have launched a joint research center (https://fcai.fi/nvaitc) to accelerate AI research in Finland. This presentation showcases the first collaborative project on optimising differentially private learning, that was conducted under the joint center.

Differential privacy provides a strong theoretical foundation for machine learning algorithms that guarantee that the result cannot be used to violate the privacy of the data subjects. FCAI researchers have successfully applied differential privacy to a range of tasks, including data anonymisation using generative models. The algorithmic basis for this research is the differentially private version of stochastic gradient optimisation. An initial implementation of this algorithm was found to suffer from unexpectedly poor performance relative to standard learning. By correctly harnessing GPU-acceleration and introducing a new GPU-optimised random permutation generator, we were able to deliver significant performance improvements to the original implementation. Joint work between Lukas Prediger (Aalto University) and Niki Loppi (NVIDIA AI Technology Center).

Speakers:  Niki Loppi & Lukas Prediger

Affiliation: NVIDIA & Department of Computer Science, Aalto University

Place of Seminar:  Zoom (Now available on YouTube)