Aral Cimcim

Investigating a novel approach to multi-label molecular property prediction by adapting diffusion models to learn conditional distributions of chemical properties. The research explores diffusion-based prediction as an alternative to traditional GNNs, with a focus on capturing uncertainty, improving robustness to class imbalance, and producing more reliable and well-calibrated predictions. Benchmarking these models on the JKU Tox21 dataset to investigate whether diffusion-based approaches can overcome lim

JKU Institute of Computer Graphics, Computational Imaging, Machine Learning & Intelligent Optics Engineered a custom interface and processing pipeline for state-of-the-art Vision Transformer models, including VGGT and DA3, to streamline optical sectioning workflows.