Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]
Random Registers for Cross-Domain Few-Shot Learning
Authors: Shuai Yi, Yixiong Zou, Yuhua Li, Ruixuan Li
ICML 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on four benchmarks validate our rationale and state-of-the-art performance. Codes and models are available at https://github.com/shuaiyi308/REAP. 4. Experiments |
| Researcher Affiliation | Academia | 1School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China 2School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, China. Correspondence to: Yixiong Zou <EMAIL>, Yuhua Li <EMAIL>, Ruixuan Li <EMAIL>. |
| Pseudocode | No | The paper describes the methods in natural language and mathematical equations, without explicitly presenting any pseudocode or algorithm blocks. |
| Open Source Code | Yes | Codes and models are available at https://github.com/shuaiyi308/REAP. |
| Open Datasets | Yes | Following current works (Oh et al., 2022), our model is trained on the mini Image Net dataset (Vinyals et al., 2016) as the source domain and then transferred to four target-domain datasets, including Crop Diseases (Mohanty et al., 2016), Euro SAT (Helber et al., 2019), ISIC2018 (Codella et al., 2019), and Chest X (Wang et al., 2017), using the k-way n-shot classification. |
| Dataset Splits | Yes | Following current works (Chen et al., 2021; Fu et al., 2021), n-way k-shot episodes are sampled for target-domain training and evaluation. It means each episode consists of a support set {IT hj, y T hj}n,k h=1,j=1 with n classes and k samples in each class for training and a query set {IT q } for evaluation. The classification is conducted by the distance between class prototypes and samples |
| Hardware Specification | No | The computation is completed in the HPC Platform of Huazhong University of Science and Technology. This statement does not provide specific hardware details such as GPU/CPU models or memory. |
| Software Dependencies | No | We use the Adam (Kingma & Ba, 2017) optimizer for 50 epochs with a learning rate of 10 5 for the backbone network and 10 3 for the classifier respectively. This mentions the Adam optimizer but does not specify its version or any other software dependencies with version numbers. |
| Experiment Setup | Yes | During the training on the source domain, we take Vi T-S as the backbone network and DINO pretraining on Image Net as the initialization following (Zhang et al., 2022b; Fu et al., 2023). We set the ratio of anchor number and minimum drop ratio to 70% for cluster-dropping to the images, and replace them with Random Registers. We use the learnable standard deviation to generate random Gaussian noise and the initial value we set is 0.1. In addition, we also concatenate an additional 16 random registers to the reconstructed patches as the input sequence of the model. Our model has trained with the Adam (Kingma & Ba, 2017) optimizer for 50 epochs with a learning rate of 10 5 for the backbone network and 10 3 for the classifier respectively. During the few-shot evaluation on target domains, we provide the image with the same number(16) of learnable registers as the input of the Vi T and set a learning rate of 10 3 for registers especially for absorbing target-domain domain-specific information. |