Multi-Instance Partial-Label Learning with Margin Adjustment

Authors: Wei Tang, Yin-Fang Yang, Zhaofei Wang, Weijia Zhang, Min-Ling Zhang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate the superior performance of MIPLMA over existing MIPL algorithms, as well as other well-established multi-instance learning algorithms and partial-label learning algorithms.
Researcher Affiliation Academia 1School of Computer Science and Engineering, Southeast University, Nanjing 210096, China 2Key Laboratory of Computer Network and Information Integration (Southeast University), Ministry of Education, China 3School of Information and Physical Sciences, The University of Newcastle, Callaghan, NSW 2308, Australia
Pseudocode Yes Algorithm 1 describes the complete procedure of MIPLMA. First, the algorithm uniformly initializes the weights on the candidate label set (Step 1). In each epoch, the training set is divided into multiple mini-batches (Step 3).
Open Source Code Yes The code of MIPLMA can be found at https://github.com/tangw-seu/MIPLMA. ... the source code of our proposed MIPLMA is included in the supplementary material. ... We submit the code of our algorithm in the supplementary material and will make it publicly available.
Open Datasets Yes Following the experimental setup of DEMIPL [21], we utilize four MIPL benchmark datasets and one real-world dataset. The four benchmark datasets encompass MNIST-MIPL, FMNIST-MIPL, Birdsong-MIPL, and SIVAL-MIPL, spanning diverse application domains such as image analysis and biology [27 30]. Additionally, the real-world CRC-MIPL dataset is annotated by crowdsourced workers for colorectal cancer classification. ... The MIPL datasets can be accessed publicly at http://palm.seu.edu.cn/zhangml/.
Dataset Splits No The dataset partitioning method aligns with that of DEMIPL [21] and ELIMIPL [22]. We execute ten random train/test splits, maintaining a ratio of 7 : 3. No explicit mention of a validation split percentage or size.
Hardware Specification Yes We implement MIPLMA using Py Torch [44] and conduct training with a single NVIDIA Tesla V100 GPU.
Software Dependencies No We implement MIPLMA using Py Torch [44] and conduct training with a single NVIDIA Tesla V100 GPU. While PyTorch is mentioned, a specific version number is not provided, nor are other key software dependencies with their versions.
Experiment Setup Yes Employing the stochastic gradient descent (SGD) optimizer, we set the momentum value to 0.9 with a weight decay of 0.0001. The initial learning rate is chosen from the set {0.01, 0.05} and coupled with a cosine annealing technique. We set the number of epochs to 100 for benchmark datasets and 200 for the CRC-MIPL dataset. The weight of the margin distribution loss is chosen from the set {0.01, 0.05, 0.1, 0.5, 1, 3, 5} for all datasets. For the annealing process of the temperature parameter, the initial temperature parameter τ (0) = 5. Additionally, τm = 0.1 and τm = 0.5 are used for benchmark datasets and the CRC-MIPL dataset, respectively.