EIANet: A Novel Domain Adaptation Approach to Maximize Class Distinction with Neural Collapse Principles

TL;DR

This work tackles source-free domain adaptation by leveraging neural collapse to maximize class-prototype separation through a fixed simplex Equiangular Tight Frame (ETF) classifier integrated with an attention mechanism (EIANet). The ETF classifier provides equiangular, maximally separated class prototypes, while the attention module concentrates features on discriminative regions to align with these prototypes; the ETF remains fixed during training. Training on the source domain uses conventional supervision, while target-domain adaptation employs a feature-bank based similarity and diversity objective to refine target embeddings without altering the ETF. Empirical results across Office-Home, Office-31, and fine-grained datasets (CUB-Paintings, Birds-31) show state-of-the-art performance, with notable gains over strong baselines and demonstrable ablations confirming the synergy of attention and ETF design.

Abstract

Source-free domain adaptation (SFDA) aims to transfer knowledge from a labelled source domain to an unlabelled target domain. A major challenge in SFDA is deriving accurate categorical information for the target domain, especially when sample embeddings from different classes appear similar. This issue is particularly pronounced in fine-grained visual categorization tasks, where inter-class differences are subtle. To overcome this challenge, we introduce a novel ETF-Informed Attention Network (EIANet) to separate class prototypes by utilizing attention and neural collapse principles. More specifically, EIANet employs a simplex Equiangular Tight Frame (ETF) classifier in conjunction with an attention mechanism, facilitating the model to focus on discriminative features and ensuring maximum class prototype separation. This innovative approach effectively enlarges the feature difference between different classes in the latent space by locating salient regions, thereby preventing the misclassification of similar but distinct category samples and providing more accurate categorical information to guide the fine-tuning process on the target domain. Experimental results across four SFDA datasets validate EIANet's state-of-the-art performance. Code is available at: https://github.com/zichengpan/EIANet.

Figures (3)

• Figure 1: Illustration of the proposed ETF classifier compared with normal classifiers. Normal classifiers cannot effectively distinguish samples if two class prototypes are closely related, for example, the samples of class $u_2$ (orange) and class $u_3$ (green) which are close to their corresponding decision boundary can be easily misclassified. The ETF classifier can maximally separate each class's prototype, thereby improving the classification accuracy.
• Figure 2: The clustering process of the proposed EIANet during target domain adaptation. The ETF classifier is set up at the beginning before source domain training and remains consistent. Input images first undergo feature extraction via an image encoder with the attention mechanism highlighting salient regions. The refined data is then classified based on the proximity of the image's features to the distinct class prototypes in the ETF classifier.
• Figure 3: Components analysis on Office-31.