Chair in Visual Computing
The Visual Computing Group focuses on computer vision and machine learning. One of our main areas of focus is the automated development of neural networks. Our goal is to determine how such networks can become more efficient while remaining robust in the face of image noise. We are also exploring new approaches to better understand how neural networks make decisions—and how these decisions can be made to more closely resemble human decision-making.
Our Research Profile
Development of efficient machine learning approaches for various tasks in the field of computer vision.
Our group's research focuses on machine learning and computer vision methods, with an emphasis on:
-
Efficient search for neural architectures
We focus on advancing the automated design and optimization of neural networks to replace manual trial-and-error methods with efficient search strategies. Our group develops acceleration methods—often using predictive models—to explore extensive search spaces and minimize computationally expensive evaluations, thereby optimizing the path to high-performance architectures.
-
Parameter Generation and Transfer Learning
Our research extends transfer learning beyond traditional pre-trained weights to include the generation of network parameters from learned distributions and the use of large language models. We aim to improve the efficiency and adaptability of neural networks across various domains while carefully considering the number of parameters and resource constraints.
-
Improving the Robustness of Neural Networks
An important aspect of our research is improving the robustness of neural networks against image noise and other disturbances that can lead to incorrect predictions. We are investigating strategies such as emphasizing object detection and controlling frequency distortions in network weights to reduce vulnerabilities and strengthen the reliability of the models.
Research Focus Areas
- Automated Design of Neural Architectures
- Robustness of Deep Learning Models
- Object Re-identification
- Efficiency of Neural Networks
- Joint hardware-software optimization (Learning2Sense)
- Learning and Generating Weights
Courses
Courses
We offer the following courses:
- Automated Machine Learning (Summer Semester)
- Digital Image Processing (Summer Semester)
- Deep Learning Practicum (Winter Semester)
- Unsupervised Deep Learning (Winter Semester)
Master’s and Bachelor’s Theses
If you are interested in writing a bachelor’s or master’s thesis or participating in a project group with us, please feel free to contact us. To write a thesis with us, you must have attended at least one of our lectures. Please note that you must include your current transcript with your inquiry. The following list contains only a small selection of the topics currently offered by our group. Please contact us for additional topics and further information. Your own ideas and interests are also welcome. Please note that we do not supervise external projects that require signing a confidentiality agreement.
- Alexander Auras:
- NAS for Inverse Problems
- Kıvanç Tezören:
- Hardware-Aware NAS
- Neural Network Compression
- Object Re-Identification
- Rachana Tirumanyam:
- Reinforcement Learning
- Automated Machine Learning
- Neural Architecture Search
Publications
An Evaluation of Zero-Cost Proxies - from Neural Architecture Performance Prediction to Model Robustness
An Evaluation of Zero-Cost Proxies - from Neural Architecture Performance Prediction to Model Robustness
Can we talk models into seeing the world differently?
Can we talk models into seeing the world differently?
Transferrable Surrogates in Expressive Neural Architecture Search Spaces
Transferrable Surrogates in Expressive Neural Architecture Search Spaces
Implicit Representations for Constrained Image Segmentation
Implicit Representations for Constrained Image Segmentation
Surprisingly Strong Performance Prediction with Neural Graph Features
Surprisingly Strong Performance Prediction with Neural Graph Features
An Evaluation of Zero-Cost Proxies - From Neural Architecture Performance Prediction to Model Robustness
An Evaluation of Zero-Cost Proxies - From Neural Architecture Performance Prediction to Model Robustness
Improving Native CNN Robustness with Filter Frequency Regularization
Improving Native CNN Robustness with Filter Frequency Regularization
Neural Architecture Design and Robustness: A Dataset
Neural Architecture Design and Robustness: A Dataset
Learning Where to Look – Generative NAS is Surprisingly Efficient
Learning Where to Look – Generative NAS is Surprisingly Efficient
Surrogate NAS Benchmarks: Going Beyond the Limited Search Spaces of Tabular NAS Benchmarks
Surrogate NAS Benchmarks: Going Beyond the Limited Search Spaces of Tabular NAS Benchmarks
Neural Architecture Performance Prediction Using Graph Neural Networks
Neural Architecture Performance Prediction Using Graph Neural Networks
Smooth Variational Graph Embeddings for Efficient Neural Architecture Search
Smooth Variational Graph Embeddings for Efficient Neural Architecture Search
Current Research Projects
KIWI@SIWI
KIWI@SIWI is a project focused on detecting European bison using AI technology. The goal is to use this technology to better detect the animals and thereby support herd management. The project is being carried out by the University of Siegen and NanoGiant GmbH. It has received funding through the NEXT.IN.NRW funding program.
DFG Emmy Noether Group UniCoDe
With funding from the DFG through the Emmy Noether Program, the UniCoDe project investigates how neural network design and training can be improved. In particular, we focus on the goal of effectively designing networks for computer vision and their applicability in various constrained domains.
Contact the Working Group
Postal address
University of Siegen
Visual Computing Group
Hölderlinstraße 3
57076 Siegen
Visitor address
University of Siegen
Visual Computing Group
H-A Level 7
Room: H-A 7107
57076 Siegen
Secretariat
Secretary: Sarah Wagener
Phone: +49 (0)271 / 740-3315
Room: H-A 7107
Email: sarah-chr.wagener@uni-siegen.de