MSc. Thesis Defense: Sormeh Serpoosh

A CURRICULUM-BASED METHOD FOR ROBUST DOMAIN

GENERALIZATION IN NEURAL NETWORKS

Sormeh Serpoosh

MSc. Thesis, 2025

Thesis Jury

Assoc. Prof. Dr. Öznur Taştan (Thesis Advisor)

 Assoc. Prof. Dr. Huseyin Ozkan

Prof. Dr. Pınar Duygulu Şahin

Date & Time: July 17th, 2025 – 11:00 AM

Place: FENS G025

Keywords : Domain generalization, Domain Shift, Curriculum Learning,

Progressive Feature Alignment, Feature Remixing, Computer Vision

Abstract

Domain generalization aims to train models to perform well on unseen domains without access to data from those domains during training. ADRMX (Additive Disentanglement of Domain Features with Remix Loss) is an augmentation based design to improve generalization to unseen domains. ADMRX disentangles domain-invariant and domain-specific features via an additive architecture and applies a latent-space remix loss, mixing same-class representations across source domains to generate synthetic samples. Building on the ADRMX method, which mixes feature representations of same-class samples across different domains, this thesis introduces Progressive Feature Alignment (PFA). PFA is a curriculum-driven remixing strategy. Remixing proceeds from the closest to the most distant pairs of domains, with mixing coefficients dynamically adjusted based on class centroid distances  across domains to prevent unrealistic blending of dissimilar features and reduce noise in the resulting synthetic examples. By organizing feature remixing according to semantic proximity, PFA enables a gradual adaptation to increasingly challenging shifts.

Under the leave-one-domain-out protocol on the PACS and OfficeHome benchmarks, PFA consistently outperforms ADRMX and other state-of-the-art domain generalization techniques, yielding especially strong gains on the more challenging OfficeHome dataset. These results demonstrate that a curriculum-driven approach to feature remixing can substantially enhance the robustness of computer vision models to complex domain variation, suggesting new directions for tackling severe shifts in unseen data.