How Effective Can Dropout Be in Multiple Instance Learning ?
Authors: Wenhui Zhu, Peijie Qiu, Xiwen Chen, Zhangsihao Yang, Aristeidis Sotiras, Abolfazl Razi, Yalin Wang
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on five MIL benchmark datasets and two WSI datasets demonstrate that MIL-Dropout boosts the performance of current MIL methods with a negligible computational cost. |
| Researcher Affiliation | Academia | 1Arizona State University, USA. 2Washington University in St. Louis, USA. 3Clemson University, USA.. |
| Pseudocode | Yes | Algorithm 1 MIL-Dropout Mechanism |
| Open Source Code | Yes | The code is available at https://github.com/ Chong Qing No Subway/MILDropout. |
| Open Datasets | Yes | The benchmark datasets include MUSK1, MUSK2, FOX, TIGER, and ELEPHANT, which are commonly used to evaluate and compare the performance of MIL algorithms. The CAMELOYON16 dataset aims to identify metastatic breast cancer in lymph node tissue and consists of high-resolution digital WSIs. The TCGANSCLC dataset primarily identifies two subtypes of lung cancer: lung squamous cell carcinoma and lung adenocarcinoma. |
| Dataset Splits | Yes | all experiments are conducted five times with a 10-fold cross-validation. ... The CAMELOYON16 dataset... It is divided into a training set of 270 samples and a testing set of 129 samples. The TCGANSCLC dataset... 1037 WSIs were divided into a training set of 744 WSIs, a validation set of 83 WSIs, and a testing set of 210 WSIs. |
| Hardware Specification | Yes | We implementation all experiments on a node of cluster with NVIDIA V100 (32GB). |
| Software Dependencies | Yes | We use Pytorch Library (Paszke et al., 2019) with version of 1.13. |
| Experiment Setup | Yes | For these two experiments, we uniformly adopted the ABMIL (Ilse et al., 2018) with cross-entropy loss, and the Lookahead optimizer was employed with a learning rate of 1e-4 and weight decay of 1e-4. ... All experiments used the Adam optimizer with 2e-4 learning rates and 5e-3 weight decay and trained on cross-entropy loss for 40 epochs. |