Sebastian Hönel

About Me

I am currently a postdoctoral researcher and co-Pi in the project In-line visual inspection using unsupervised learning.

• You will find me in building D, room D2254.
• I have a calendar where you can find my public appearances, talks, meetings, office hours, and other public activities at bit.ly/2XY1RZ1.

Undervisning

I am usually involved in teaching courses around (Deep) Machine Learning (e.g., 4DV652, 4DV660, 4DV661).

I was previously a teaching assistant in courses for agile product development, both first- and second-level cycles (e.g., 1DV508, 4DV611).

Forskning

Previously, I emphasized the usage of Machine- and Deep Learning, Optimization, as well as distributed computing for the analysis, calibration, and optimization of models that are related to software applications and -processes. I was (and still am) interested in obtaining quantitative data from software evolutionary processes and applying it through models used for the greater goal of organizational learning.

_____________________________________________________________

Currently, I attempt to apply and improve models for unsupervised and zero-/few-shot learning for the detection of manufacturing defects and other material anomalies. I am particularly interested in Deep Density Estimation (e.g., Normalizing Flows) and Representation Learning. In greater detail, you find what interests me most below.

I usually have concrete problems in some of the following that need solving and are suitable for collaborations or writing a degree project. If you are interested in any of the following, please get in touch!

Anomaly Detection :

• Industrial anomaly detection and -localization.
• One-class classification and outlier exposure.
• Reconstruction vs. classification vs. density estimation methods.
• Contrastive learning and its applications.
• Robustness and invariance in anomaly detection.

Normalizing Flows (NF) :

• Density estimation using Normalizing Flows.
• Challenges and limitations in using NFs for OOD detection.
• Improvements and adjustments in NF architectures (e.g., coupling layers, attention mechanisms).
• Comparison with other probabilistic and generative models.

Feature Extraction and Representation Learning :

• Using pre-trained feature extractors (FEs) like Vision Transformers (ViTs) and ordinary convolutional (equivariant) networks.
• Hierarchical and multi-scale feature extraction.
• The role of semantic features versus low-level features.
• Self-supervised learning to improve model robustness and uncertainty.

Architectural Innovations :

• Autoencoders, especially variations like stacked convolutional AEs and zero-bias AEs.
• Hybrid models with deep and invertible features.
• Specialized AutoEncoders (AEs) and custom losses (e.g., Mutual Information)

Methodological Considerations :

• Evaluation metrics (e.g., AUROC, MI, WAIC, MMD, Typicality).
• Synthetic and variational data augmentation, as well as synthesizing points of interest (e.g., anomalies, novelties)
• Proxy tasks
• Runtime regularization (noise & temperature scaling; ratio models; etc.)

Key Algorithms and Methods :

• Independent Component Analysis (ICA) and Transform-based methods.
• Change-of-Variable (Gaussianization) and probability integral transform.
• The use of various loss functions (e.g., focal loss, margin ranking loss).
• Transfer learning and the importance of auxiliary data.