Rethinking Confidence Scores and Thresholds in Pseudolabeling-based SSL
Authors: Harit Vishwakarma, Yi Chen, Satya Sai Srinath Namburi Gnvv, Sui Jiet Tay, Ramya Korlakai Vinayak, Frederic Sala
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments show that, by integrating this framework with modern SSL methods, we achieve significant improvements in accuracy and training efficiency. In addition, we provide novel insights on the tradeoffs between the choices of the error parameter and the end model s performance. ... 5. Experiments: We conduct empirical evaluation over several settings to, C1. Verify that the adaptations of popular pseudolabelingbased SSL methods with Pab LO output models with better test accuracy. C2. Study the effects of choice of error tolerance ϵ on test accuracy of the final model. C3. Understand the role of pseudolabel accumulation in our method and baselines. |
| Researcher Affiliation | Collaboration | 1Department of Computer Sciences, University of Wisconsin-Madison, WI, USA 2Department of Electrical and Computer Engineering, University of Wisconsin-Madison, WI, USA 3GE Health Care 4NYU Courant Institute. Correspondence to: Harit Vishwakarma <EMAIL>, Yi Chen <EMAIL>. |
| Pseudocode | Yes | The detailed steps are outlined in Algorithm 3 in the Appendix. ... B. Detailed Algorithms: Algorithm 1 Estimate Pseudolabeling Thresholds Classwise; Algorithm 2 Estimate Pseudolabeling Threshold Jointly for All Classes; Algorithm 3 Pseudolabeling Based SSL with Pab LO |
| Open Source Code | Yes | First, we briefly describe the experimental setup, with details deferred to Appendix C. The code is available on Git Hub 1. 1https://github.com/harit7/Pab LO-SSL |
| Open Datasets | Yes | We experiment with three datasets: CIFAR-10 (Krizhevsky et al., 2009) is an image dataset with 10 classes. CIFAR-100 (Krizhevsky et al., 2009) is an extended version of CIFAR-10 with 100 classes. SVHN (Netzer et al., 2011) is a 10-class image dataset of digits from Google Street View. |
| Dataset Splits | Yes | Table 1. Details of the dataset we use in our experiments. k is the number of classes. Nl is the number of labeled data points used for training the backbone model h. nu is the number of unlabelled data points used for consistency regularization and pseudolabeling for all the methods. Nval is the number of points used for model selection in all methods. Ntest is the number of test data points. Ncal is the number of points used for learning the g function. Nth is the number of data points used for threshold estimation. ... Unless otherwise mentioned, we use Nl as 250 for CIFAR-10 and SVHN and 2500 for CIFAR-100 in our experiments. |
| Hardware Specification | Yes | We ran all of our experiments on a high-throughput system with various GPUs. Therefore, each individual experiment task may be scheduled among NVIDIA A100 SXM4-40GB, NVIDIA A100 SXM4-80GB, NVIDIA L40, and NVIDIA H100 80GB HBM3. We measured the runtime of our algorithm on a desktop with a single NVIDIA RTX 4090. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies. It mentions 'The code is available on Git Hub' but doesn't list explicit software versions. |
| Experiment Setup | Yes | Table 10. Hyperparameters used for our method. Learning g function: optimizer SGD learning rate 0.01 batch size 64 max epoch 500 weight decay 0.01 momentum 0.9. Estimating t: optimizer SGD learning rate 0.01 batch size 64 max epoch 500 weight decay 0.01 momentum 0.9. |