A Graph-Theoretic Framework for Understanding Open-World Semi-Supervised Learning

Authors: Yiyou Sun, Zhenmei Shi, Yixuan Li

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, SORL can match or outperform several strong baselines on common benchmark datasets, which is appealing for practical usage while enjoying theoretical guarantees.
Researcher Affiliation Academia Yiyou Sun, Zhenmei Shi, Yixuan Li Department of Computer Sciences University of Wisconsin, Madison {sunyiyou,zhmeishi,sharonli}@cs.wisc.edu
Pseudocode No The paper describes the Spectral Open-world Representation Learning (SORL) algorithm but does not present it in a pseudocode block or explicitly labeled algorithm section.
Open Source Code Yes Our code is available at https://github.com/deeplearning-wisc/sorl.
Open Datasets Yes Following the seminal work ORCA [7], classes are divided into 50% known and 50% novel classes. We then use 50% of samples from the known classes as the labeled dataset, and the rest as the unlabeled set. [...] on standard benchmark image classification datasets CIFAR-10/100 [35].
Dataset Splits Yes Following the seminal work ORCA [7], classes are divided into 50% known and 50% novel classes. We then use 50% of samples from the known classes as the labeled dataset, and the rest as the unlabeled set.
Hardware Specification Yes We run all experiments with Python 3.7 and Py Torch 1.7.1, using NVIDIA Ge Force RTX 2080Ti and A6000 GPUs.
Software Dependencies Yes We run all experiments with Python 3.7 and Py Torch 1.7.1, using NVIDIA Ge Force RTX 2080Ti and A6000 GPUs.
Experiment Setup Yes For CIFAR-10, we set ηl = 0.25, ηu = 1 with training epoch 100, and we evaluate using features extracted from the layer preceding the projection. For CIFAR-100, we set ηl = 0.0225, ηu = 3 with 400 training epochs and assess based on the projection layer s features. We use SGD with momentum 0.9 as an optimizer with cosine annealing (lr=0.05), weight decay 5e-4, and batch size 512.