Domain-Guided Weight Modulation for Semi-Supervised Domain Generalization

TL;DR

A novel method is proposed that can facilitate the generation of accurate pseudo-labels under various domain shifts by retaining the domain-level specialism in the classifier during training corresponding to each source domain by utilizing a domain-aware mask for modulating the classifier's weights.

Abstract

Unarguably, deep learning models capable of generalizing to unseen domain data while leveraging a few labels are of great practical significance due to low developmental costs. In search of this endeavor, we study the challenging problem of semi-supervised domain generalization (SSDG), where the goal is to learn a domain-generalizable model while using only a small fraction of labeled data and a relatively large fraction of unlabeled data. Domain generalization (DG) methods show subpar performance under the SSDG setting, whereas semi-supervised learning (SSL) methods demonstrate relatively better performance, however, they are considerably poor compared to the fully-supervised DG methods. Towards handling this new, but challenging problem of SSDG, we propose a novel method that can facilitate the generation of accurate pseudo-labels under various domain shifts. This is accomplished by retaining the domain-level specialism in the classifier during training corresponding to each source domain. Specifically, we first create domain-level information vectors on the fly which are then utilized to learn a domain-aware mask for modulating the classifier's weights. We provide a mathematical interpretation for the effect of this modulation procedure on both pseudo-labeling and model training. Our method is plug-and-play and can be readily applied to different SSL baselines for SSDG. Extensive experiments on six challenging datasets in two different SSDG settings show that our method provides visible gains over the various strong SSL-based SSDG baselines.

Figures (6)

• Figure 1: Left: Pseudo-labeling (PL) accuracy when source domains (Art, Cartoon, and Product) are gradually added to the set of training domains in baseline sohn2020fixmatch and ours on OfficHome dataset. Our method tends to maintain a higher PL accuracy while the baseline's PL accuracy drops upon gradually adding source domains. Right: Top-1 Accuracy on the target domain (Real-world).
• Figure 2: Overall architecture with our domain-guided weight modulation method.
• Figure 3: PL accuracy during training for baseline sohn2020fixmatch and ours.
• Figure 4: PL accuracy when computed using using features in $J_+$ for classes $c$ with $v_{\mathrm{cls}}[c] > 0$ and features in $J_-$ for classes $c$ with $v_{\mathrm{cls}}[c] < 0$ on OH dataset (10 labels).
• Figure 5: (Left) Pseudo-label accuracy upon varying the confidence threshold in FixMatch and our method. (Right) Unlabelled data utilization i.e. the percentage of unlabelled data that passes the confidence threshold for both FixMatch and our method. These results are shown on the OfficeHome dataset with 10 label settings.