A General Framework for Learning from Weak Supervision

Authors: Hao Chen, Jindong Wang, Lei Feng, Xiang Li, Yidong Wang, Xing Xie, Masashi Sugiyama, Rita Singh, Bhiksha Raj

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental GLWS not only enhances the scalability of machine learning models but also demonstrates superior performance and versatility across 11 weak supervision scenarios. We conduct the evaluation mainly on CIFAR-10 (Krizhevsky et al., 2009), CIFAR-100 (Krizhevsky et al., 2009), STL-10 (Coates et al., 2011), and Image Net-100 (Russakovsky et al., 2015).
Researcher Affiliation Collaboration 1Carnegie Mellon University 2Microsoft Research 3William & Mary 4Singapore University of Technology and Design 5Peking University 6RIKEN AIP 7The University of Tokyo 8Mohamed bin Zayed University of AI.
Pseudocode Yes We present the pseudo-algorithm of performing the forward-backward algorithms on common weak supervision settings we evaluated. The pseudo-algorithm also corresponds to description of the trellis expanded from the NFA. ... Algorithm 1 Forward-Backward Algorithm for multiple instance (Multi Ins) Learning ... Algorithm 2 Forward-Backward Algorithm for label proportion (LProp) Learning ... Algorithm 3 Forward-Backward Algorithm for pairwise comparison (PComp) Learning
Open Source Code Yes Code is available at: https://github.com/Hhhhhhao/ General-Framework-Weak-Supervision.
Open Datasets Yes We conduct the evaluation mainly on CIFAR-10 (Krizhevsky et al., 2009), CIFAR-100 (Krizhevsky et al., 2009), STL-10 (Coates et al., 2011), and Image Net-100 (Russakovsky et al., 2015). Results on MNIST (Deng, 2012) and F-MNIST (Xiao et al., 2017) are included in the Appendix, where most of the baseline methods were evaluated.
Dataset Splits Yes Table 6. Dataset details Dataset # Classes # Training # Validation # Unlabeled (...) CIFAR-10 10 50,000 10,000
Hardware Specification No The paper does not provide specific details regarding the hardware used to run experiments, such as GPU models, CPU types, or cloud computing specifications.
Software Dependencies No The paper mentions software components like 'Adam W' and 'SGD' as optimizers, and models like 'CLIP', but does not provide specific version numbers for Python, PyTorch, or other libraries and frameworks used for implementation.
Experiment Setup Yes We follow the hyper-parameters from Wu et al. (2022) for training all methods, with more details provided in Appendix C.2.1. A summarize of training parameters is shown in Table 7. Here we present the training hyper-parameters we used for Multi Ins and LProp in Table 9. For pairwise observations (x1), x2), we adopt the same training parameters for the four settings we evaluated, as shown in Table 11. The hyper-parameters are shown in Table 16.