On Dec. 10, the members of the Industrial Consortium SimTech e. V. (IC SimTech) met virtually for their annual general meeting. In addition to classic topics, the Simulation Technology study program, the Junior Academy, a brief presentation of the project networks of the Cluster of Excellence and a presentation by the company Altair were on the agenda.
In addition, as in previous years, IC SimTech had offered a prize for outstanding final theses, endowed with 500 EUR. This year, the winners were selected in three categories: Bachelor, Master and doctoral theses. They were selected by the IC SimTech board consisting of Claus-Dieter Munz, Alexander Verl and Jörg Fehr.
In the category "Bachelor Thesis" Philipp Scholze was honored for his work on "Stability of Neural Network Architectures for Optical Flow Estimation under Adversarial Attacks".
In the justification for this decision, it was emphasized above all that Mr. Scholze's work is outstanding due to several aspects. "On the one hand, it shows an extremely high technical standard in the theoretical elaboration, and on the other hand, the practical implementation required, in addition to a deep understanding of existing solutions, a very fast working method, as well as great detailed knowledge and a very high degree of care. From my point of view, Mr. Scholze's work is clearly one of the best bachelor theses (out of about 50) I have supervised in the last 10 years, if not the best. It is indeed rare to see such a strong scientific achievement at such a comparatively early stage in one's studies (even before starting the master's degree). That's why I've also employed him as a research hiwi in my group since his bachelor's degree. The aim is to develop the results of his bachelor thesis into a publication," explained Andres Bruhn.
Abstract: Neural networks are currently used successfully in many areas. In the context of optical flow estimation, RAFT is a very effective network because it provides fast and also accurate predictions. So far, the focus in the analysis of such networks is primarily on the accuracy of the prediction, the complexity of the architecture, or the computation time. The stability with regard to disturbed input data however, is not yet or only insufficiently investigated for many networks, such as RAFT. The fact that potential weaknesses of a network can be exploited deliberately leads to a significant security risk, for example in the context of autonomous driving. Therefore, we examine two key aspects in this bachelor thesis: On the one hand, we analyze the stability of RAFT when the input images are sabotaged using optimized patches. On the other hand, we modify the existing architecture to make RAFT more robust and less vulnerable to these perturbations. We first transfer the framework developed by Ranjan et al. to perform targeted attacks on RAFT. This shows that by cleverly placing small patches near the the image boundaries, RAFT falsely predicts large areas of zero flow that exceed the dimensions of the actual patch. Therefore, in the next step, we establish different architectural layers of the network that compute the optical flow alternately or successively and compose it from different resolutions. Using our pyramid approach pyRAFTmid, we thereby increase the accuracy of the predicted optical flow by up to 12% on Sintel, while decreasing the vulnerability to attacks with optimized patches by up to 30%. Our results help to identify such components that make existing networks more robust and thus contribute to making new methods more robust in this regard.
In the "Master's Thesis" category, Mohamed Adnen Abdessaied came out on top with his work on "Fast Learning System in Multi-Modal Context".
In the statement for being nominated it was said that “Mr. Abdessaied's work relied on the "differentiable neural computer" (Graves et al 2014), it was entirely implemented in Python, and integrated it into a Deep LSTM-Q VQA architecture. The integration was very challenging, required a deep understanding of the underlying mathematics, as well as creativity since such an integration has never been attempted before. Mr. Abdessaied evaluated his new architecture extensively, especially for two other tasks in addition to VQA. Mr. Abdessaied summarized the results of his work in a very well structured, written and illustrated paper. The results were so promising that the paper will be submitted to the ACL conference - the top conference in Natural Language Processing."
Abstract: Inspired by the human brain's complementary learning system, in this thesis we propose a fast-learning neural model that tries to emulate the interplay between the Medial Temporal Lobe (MTL) and the neocortex. We evaluate this new model on the multimodal visual question answering task. We investigate different ways of integrating the external memory-enhanced differential neural computer (DNC) into a Deep-LSTM-Q network. Through extensive experiments, we show that all our models outperform the baseline by a considerable margin on the VQA v2.0 test-dev dataset with our best model scoring a relative accuracy improvement of over 2%. Furthermore, we demonstrate that our models are better at answering notoriously difficult questions, i.e. long questions that require deeper reasoning as well as question with rare ground-truth answers.
With his doctoral thesis on "Scalable Biophysical Simulations of the Neuromuscular System" Benjamin Maier was awarded in the category "Doctoral Thesis".
The reasons for the decision were as follows: “The thesis work has both provided a new level of efficiency and embedding for electrophysiological skeletal muscle simulations. The development of a completely new high performance computing simulation framework required a lot of numerical and implementational skills. In addition, the framework was enhanced by a sophisticated parallel geometry and mesh generation tool using imaging data as input. These challenges were met with a very high quality of both thesis and software. In addition, Benjamin contributed a lot also to the interdisciplinary cooperation of dotoral researchers in Stuttgart and in New Zealand unter the umbrella of the IRTG 2198 Soft Tissue robotics. In this context, he made additional substantial contributions in research and teaching to control of robots handling soft objects.”
Abstract: The thesis "Scalable Biophysical Simulations of the Neuromuscular System" considers simulations in the field of bio-mechanics. The contraction of skeletal muscles is the result of a complex interplay of the nervous system with the active muscle tissue, which is difficult to analyze in real-world experiments. A thorough understanding of the involved processes can help, e.g., in personalized healthcare or for the development of exoskeletons. Detailed simulations can open the door for new insights, however, they usually come with high computational loads and increased complexity. This thesis summarizes state-of-the art models of the neuromuscular system and how to solve them numerically. It discusses the numeric, algorithmic and software engineering related aspects that are required to achieve fast simulations, even with detailed, multi-scale and multi-physics models. The developed software can efficiently simulate the functioning of skeletal muscles using different compute hardware. The presented simulations were computed on typical institute-scale compute servers with several dozen cores up to tens of thousands of cores on the supercomputer Hawk at the High Performance Computing Center Stuttgart.
About the Industrial Consortium SimTech e. V.
In order to make the exchange with industry sustainable, the Industrial Consortium SimTech e. V. was initiated – a non-profit organization serving as a platform for the direct exchange with experienced industrial partners. The IC SimTech works closely together with the Vice Recorate for Knowledge and Technology Transfer as well as the Cluster of Excellence on Integrative Computational Design and Construction for Architecture (IntCDC).
The association promotes cooperation between the Cluster of Excellence and companies,creates links for applications and strengthens the awareness of SimTech’s scientists for industrial needs and requirements.
For more information click here.