Self-PU: Self Boosted and Calibrated Positive-Unlabeled Training

Authors: Xuxi Chen, Wuyang Chen, Tianlong Chen, Ye Yuan, Chen Gong, Kewei Chen, Zhangyang Wang

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the state-of-the-art performance of Self-PU on common PU learning benchmarks (MNIST and CIFAR-10), which compare favorably against the latest competitors. Moreover, we study a realworld application of PU learning, i.e., classifying brain images of Alzheimer s Disease. Self PU obtains significantly improved results on the renowned Alzheimer s Disease Neuroimaging Initiative (ADNI) database over existing methods. and 4. Experiments
Researcher Affiliation Collaboration 1University of Science and Technology of China 2Texas A&M University 3Nanjing University of Science and Technology 4Banner Alzheimer s Institute.
Pseudocode No The paper describes its methods but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes 1The code is publicly available at: https://github.com/ TAMU-VITA/Self-PU.
Open Datasets Yes We conducted experiments on two common testbeds for PU learning: MNIST, and CIFAR-10; plus a new realworld benchmark, i.e. ADNI (Jack Jr et al., 2008)...
Dataset Splits Yes For standard supervised learning of binary classifiers, both positive and negative classes need to be collected for training purposes. and The batch size of validation examples equals to the batch size of the training examples. and Table 1. Specification of benchmark datasets and models. Dataset #Train #Test Input Size πp Positive/Negative Model MNIST 60,000 10,000 28 28 0.49 Odd/Even 6-layer MLP CIFAR-10 50,000 10,000 3 32 32 0.40 Traffic tools/Animals 13-layer CNN ADNI 822 113 104 128 112 0.43 AD Positive/Negative 3-branch 2-layer CNN
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper mentions 'Adam optimizer' and 'Re LU' as components used, but does not provide specific version numbers for software libraries or dependencies needed to replicate the experiment.
Experiment Setup Yes We use Adam optimizer with a cosine annealing learning rate scheduler for training. The batch size is 256 for MNIST and CIFAR-10, and 64 for ADNI. The γ is set to 1 16. The batch size of validation examples equals to the batch size of the training examples. and in all experiments and as shown in Figure 1, we first apply self-paced learning and self-calibrated loss reweighting from the 10th epoch to the 50th epoch, followed by a selfdistillation period from 50th to 200th epoch.