This is the supplementary material of paper "Towards Enabling Meta-Learning from Target Models". We give implementation details, more discussions, and more experiment results in this material.

This paper presents a new algorithm for domain generalization (DG), test-time template adjuster (T3A), aiming to robustify a model to unknown distribution shift. Unlike existing methods that focus on training phase, our method focuses test phase, i.e., correcting its prediction by itself during test time. Specifically, T3A adjusts a trained linear classifier (the last layer of deep neural networks) with the following procedure: (1) compute a pseudo-prototype representation for each class using online unlabeled data augmented by the base classifier trained in the source domains, (2) and then classify each sample based on its distance to the pseudoprototypes. T3A is back-propagation-free and modifies only the linear layer; therefore, the increase in computational cost during inference is negligible and avoids the catastrophic failure might caused by stochastic optimization. Despite its simplicity, T3A can leverage knowledge about the target domain by using off-the-shelf test-time data and improve performance. We tested our method on four domain generalization benchmarks, namely PACS, VLCS, OfficeHome, and TerraIncognita, along with various backbone networks including ResNet18, ResNet50, Big Transfer (BiT), Vision Transformers (ViT), and MLP-Mixer. The results show T3A stably improves performance on unseen domains across choices of backbone networks, and outperforms existing domain generalization methods.

AAppendix1 In the supplementary material, we provide more experimental results summarized as follows:2 In A.1, we use ResNet101 as the backbone network and compare our method with state-of-3 the-art methods, demonstrating that our method achieves consistent top results on different4 In A.2, we provide more t-SNE visualization results for a comprehensive analysis on the6 feature space learned from different models.7 In A.3, we study the effect of the image-to-image translation model on the performance of8 domain adaptive semantic segmentation.9 In A.4, we discuss the limitations of our method and provide the URL link of code to10 reproduce the main experimental results.11 "V" and "R" indicate the method using VGG16 and ResNet101 backbone networks, respectively. In the main paper, we report results using VGG1613 as the backbone for both settings: single-target14 and multi-target domain adaptation.

Then, one can adopt theprinciple of separatingthe presence of an entity and its instantiation parameters into capsule length and orientation, respectively. In particular, we use the lengths of capsules to score the presence of entity classes corresponding to different subspaces, while their orientations are used to instantiate the parameters of entity properties such as poses, scales, deformations and textures.