SEAL: Simultaneous Label Hierarchy Exploration And Learning
Authors: Zhiquan Tan, Zihao Wang, Yifan Zhang
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method on several standard benchmarks and show that it achieves improved results in semi-supervised image classification scenarios. ... Empirical evaluation demonstrates that adding the SEAL framework consistently and significantly improves the classification performance on supervised learning and various semi-supervised learning methods (Sohn et al., 2020; Zhang et al., 2021; Wang et al., 2022a). |
| Researcher Affiliation | Academia | Zhiquan Tan EMAIL Department of Mathematical Sciences Tsinghua University Zihao Wang EMAIL Department of CSE HKUST Yifan Zhang EMAIL IIIS Tsinghua University |
| Pseudocode | No | The paper does not contain any clearly labeled pseudocode or algorithm blocks. It describes methodologies using mathematical formulations and descriptive text. |
| Open Source Code | No | The paper does not contain an explicit statement about releasing code or a link to a code repository. |
| Open Datasets | Yes | We evaluate our proposed method on three popular datasets, namely CIFAR10, CIFAR100, and STL-10. CIFAR10 and CIFAR100. CIFAR10 (Krizhevsky et al., 2009) ... Similarly, CIFAR100 (Krizhevsky et al., 2009) ... STL-10. STL-10 (Coates et al., 2011) is a semi-supervised benchmark... |
| Dataset Splits | Yes | CIFAR10 (Krizhevsky et al., 2009) contains 60,000 colored images in 10 different classes... The training set consists of 50,000 labeled images and the test set consists of 10,000 labeled images. ... STL-10 (Coates et al., 2011)... contains 500 labeled training samples and 800 labeled testing samples per class. Additionally, it has 10,000 unlabeled images... |
| Hardware Specification | Yes | Table 6 shows the results of these experiments, which were conducted using an Nvidia GeForce RTX 2080 Ti. |
| Software Dependencies | No | The paper mentions using a "(batch) stochastic gradient descent (SGD) optimizer" and "Rand Augment" and refers to models like "Wide Res Net-28-2" and "Vision Nystromformer", but it does not specify any software libraries with version numbers (e.g., PyTorch 1.x, TensorFlow 2.x). |
| Experiment Setup | Yes | Specifically, we use a (batch) stochastic gradient descent (SGD) optimizer with a momentum of 0.9. We set the learning rate scheduler as the cosine decay scheduler, where the learning rate β can be expressed as β = β0cos( 7π / 16 s / S ). Here, β0 is the initial learning rate set to 0.03, s is the current optimization step, and S is the total number of optimization steps set to 220. We set the batch size of the labeled training data to 64, and the ratio of unlabeled training data to labeled data µ is set to 7. We set the threshold τ to 0.95, and the weak and strong augmentation functions used in our experiments are based on Rand Augment(Cubuk et al., 2020). |