Optimization frameworks for deep kernel machines

From 01-10-2018 to 30-09-2022

Description

Deep learning and kernel-based learning are among the very powerful methods in machine learning and data-driven modelling. From an optimization and model representation point of view, training of deep feedforward neural networks occurs in a primal form, while kernel-based learning is often characterized by dual representations, in connection to possibly infinite dimensional problems in the primal. In this project we aim at investigating new optimization frameworks for deep kernel machines, with feature maps and kernels taken at multiple levels, and with possibly different objectives for the levels. The research hypothesis is that such an extended framework, including both deep feedforward networks and deep kernel machines, can lead to new important insights and improved results. In order to achieve this, we will study optimization modelling aspects (e.g. variational principles, distributed learning formulations, consensus algorithms), accelerated learning schemes and adversarial learning methods.

Team

• Panagiotis Patrinos, Co-promoter
• Johan Suykens, Promoter

Financing

Funding: KU Leuven - Internal Funding KU Leuven

Program/Grant Type: C1 - KU Leuven Category 1 Funding