Domain-Generalizable Multiple-Domain Clustering

TL;DR

This work tackles unsupervised domain generalization for clustering across multiple source domains, aiming to predict semantically related cluster assignments for unseen domains without any labeled data. It introduces a two-phase framework: (1) self-supervised pre-training to obtain domain-invariant representations, and (2) a multi-head clustering predictor trained with pseudo labels derived from both semantic embeddings and a common domain transformation, augmented by a prediction-based label smoothing scheme. Key contributions include style transfer augmentations, adversarial domain losses, domain-specific queues and balancing during pre-training, a BCD-based pseudo-labeling process, and diversification of clustering heads to stabilize learning. The approach achieves strong performance compared to baselines that rely on target-domain data or supervision, demonstrating practical potential for fully unsupervised cross-domain clustering with broad applicability across scientific and real-world data domains.

Abstract

This work generalizes the problem of unsupervised domain generalization to the case in which no labeled samples are available (completely unsupervised). We are given unlabeled samples from multiple source domains, and we aim to learn a shared predictor that assigns examples to semantically related clusters. Evaluation is done by predicting cluster assignments in previously unseen domains. Towards this goal, we propose a two-stage training framework: (1) self-supervised pre-training for extracting domain invariant semantic features. (2) multi-head cluster prediction with pseudo labels, which rely on both the feature space and cluster head prediction, further leveraging a novel prediction-based label smoothing scheme. We demonstrate empirically that our model is more accurate than baselines that require fine-tuning using samples from the target domain or some level of supervision. Our code is available at https://github.com/AmitRozner/domain-generalizable-multiple-domain-clustering.

Figures (6)

• Figure 2: The proposed pre-training procedure. Each image is transformed using strong augmentations or style transfer augmentation. The features $u^s$ (strong augmentation) and $u^{st}$ (style augmentation) are extracted using $f_e$. Then we use a domain head $f_d$ to classify the domain identity of each sample, minimizing the domain loss $\mathcal{L}_{f_d}$; we use gradient reversal to update the feature extractor $f_e$ to fool the domain head in an adversarial fashion. The contrastive loss $\mathcal{L}_{f_{proj}}$ is minimized based on the output of the projection head $u^s$, $u^{st}$, and $u^-$ (negative samples).
• Figure 3: Clustering head training. The image is passed through the feature extractor $f_e$ in its original ($u$), strongly augmented ($u^s$), and BCD form ($u^{bcd}$). The weights of $f_e$ are frozen and used to produce the features. Representatives are selected from the original image features based on the clustering head's predictions over the BCD images. The class representatives are used as pseudo labels for the CE loss.
• Figure 4: Sample images from BCD domain for Officehome dataset venkateswara2017deep. The right and left columns show the original image and its BCD transform.
• Figure 5: Sample images from the datasets used in the paper.
• Figure : Training Data Required