Rasmus Kjær Høier Electrical engineering

Rasmus Kjær Høier is a Phd student in the research group Computer vision and medical image analysis, Signal Processing and Biomedical Engineering

Discussion leader is Assoc.professor Pawel Herman, KTH, Stockholm

Examiner is Professor Christopher Zach, Signal Processing and Biomedical Engineering

Abstract

Deep neural networks have found a wide range of applications, each of which places different constraints on successful implementations with respect to efficiency and robustness. This presentation highlights the applicability of bilevel optimization techniques to address three distinct constrained settings at training time: (1) fully localized learning (learning without backprop), (2) neural network quantization and (3) adversarial robustness by design. (1) and (2) are particularly interesting from the point of view of energy efficient neuromorphic and edge computing, while (3) is important in safety-critical applications.