Deep Weightless Networks for Verification and Control

Title

Deep Weightless Networks for Verification and Control

Speaker

Christoph H. Lampert - Institute of Science and Technology Austria (ISTA)

Abstract

Weightless Networks (Aleksander, 1983) represent functions not as combinations of linear layer with non-linear activation functions, but as combinations of lookup tables with binary inputs and outputs. Recently, deep (i.e. multi-layer) variants of weightless networks have been developed (Petersen et al. 2022; Bacellar et al. 2024) that can be trained by backpropagation, thereby opening up the possibility of real-world applications, for example in image classification or process control. In my talk, I will give an introduction into the working mechanism and training of deep weightless networks, and then present two recent projects from our group, one in which we formally verify model properties, such as robustness and fairness, and one in which we learn controllers that operate with very low latency and energy requirements.

Bio

Christoph Lampert received the PhD degree in mathematics from the University of Bonn in 2003. After postdoctoral positions at the German Research Center for Artificial Intelligence and the Max-Planck Institute for Biological Cybernetics, he joined the Institute of Science and Technology Austria (ISTA) in 2010, where he leads the research group for Machine Learning and Computer Vision. Since 2019 he is also the director of ISTA's ELLIS unit. His research interests are transfer learning, such as multi-task and meta-learning, and trustworthy machine learning, in particular methods for guaranteeing the robustness, fairness and privacy of learned systems.

When

Monday, February 23rd, 16:00

Where

Room 217, UniGe DIBRIS/DIMA, Via Dodecaneso 35