Mr. Daniel Coquelin successfully completed his dissertation on 31.10.2024. His research focus in recent years has been on distributed machine learning.
Mr. Coquelin has made valuable contributions to distributed machine learning, particularly in the area of data parallel neural networks. His early work focused on communication-avoiding approaches to data parallel training. He then investigated the behavior of neural networks during training, observing that the orthogonal basis of weight matrices tends to stabilize during training. Based on this insight, he developed a novel method for training low-rank neural networks on distributed memory systems. This method allows for efficient scaling of training across multiple devices, leading to compressed neural networks that can outperform their full-rank counterparts. His work has implications for large-scale applications like natural language processing and computer vision. Mr. Coquelin's work was funded by the Helmholtz Analytics Framework (HAF) project and the Helmholtz Artificial Intelligence Cooperation Unit (Helmholtz AI) platform.
SCC congratulates Mr. Coquelin on the successful completion of his doctorate and wishes him all the best for his future career.
[1] Prof. Ina Schäfer (KIT, Chair of the Examination Board), Prof. Achim Streit (KIT, first supervisor), Prof. Håkan Grahn (Blekinge Tekniska Högskola (BTH), Sweden, external examiner/opponent), Prof. Ralf Reussner (KIT, examiner) , Prof. Thorsten Strufe (KIT, examiner)