InterLUDE: Interactions between Labeled and Unlabeled Data to Enhance Semi-Supervised Learning

Authors: Zhe Huang, Xiaowei Yu, Dajiang Zhu, Michael C Hughes

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on standard closed-set SSL benchmarks and a medical SSL task with an uncurated unlabeled set show clear benefits to our approach.
Researcher Affiliation Academia 1Department of Computer Science, Tufts University, Medford, MA, USA 2Department of Computer Science, University of Texas at Arlington, Arlington, TX, USA.
Pseudocode Yes Alg. 1 provides pseudocode. Details on each component are introduced below.
Open Source Code Yes Code: https://github.com/tufts-ml/Inter LUDE/
Open Datasets Yes Datasets. We use CIFAR-10 (Krizhevsky et al., 2009), CIFAR-100 (Krizhevsky et al., 2009) and STL-10 (Coates et al., 2011).
Dataset Splits Yes The CIFAR-10 dataset... is divided into 50,000 training images and 10,000 testing images... We select hyperparameters based on the validation set and report test set performance at the maximum validation checkpoint.
Hardware Specification Yes We compare the wall time of Inter LUDE, Free Match (Wang et al., 2023), Flex Match (Zhang et al., 2021) and Flat Match (Huang et al., 2023a) on the TMED2 dataset using the exact same hardware (NVIDIA A100 with 80G Memory).
Software Dependencies No The paper mentions software like "PyTorch" and optimizers like "Adam W" but does not provide specific version numbers for these software dependencies, which is required for reproducibility.
Experiment Setup Yes We use SGD optimizer (Nesterov momentum 0.9) and a cosine learning rate schedule... η0 = 0.03 is the initial learning rate, K = 220 is the total training steps... We set the labeled batch size to 64 and the unlabeled batch size to 448 (i.e., µ = 7). For hyperparameters unique to Inter LUDE, we set λDC to 1.0 and fusion strength α to 0.1.