Seminar: Marcel Ochocki
During the KIiAED weekly seminars, PhD student Marcel Ochocki presented his work entitled “Toward Robustness to Domain Shift: Domain Generalization for Models Based on Whole-Exome Sequencing Data”.
The research focuses on reducing the impact of technical variability (e.g., sequencing protocols) in WES data, aiming to preserve meaningful disease-related signals while improving model generalization.
The presentation introduced a hybrid architecture combining a regression branch with a domain generalization autoencoder, leveraging KL-divergence–based latent alignment and reconstruction constraints to learn invariant feature representations across domains.
This work contributes toward more reliable and transferable machine learning models in bioinformatics, especially in scenarios where access to target-domain data is limited.
