Realistic Evaluation of Deep Semi-Supervised Learning Algorithms

Authors: Avital Oliver, Augustus Odena, Colin A. Raffel, Ekin Dogus Cubuk, Ian Goodfellow

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental After creating a unified reimplementation of various widely-used SSL techniques, we test them in a suite of experiments designed to address these issues.
Researcher Affiliation Industry Google Brain {avitalo,augustusodena,craffel,cubuk,goodfellow}@google.com
Pseudocode No The paper describes algorithms but does not provide structured pseudocode or algorithm blocks.
Open Source Code Yes To help guide SSL research towards real-world applicability, we make our unified reimplemention and evaluation platform publicly available.2... 2https://github.com/brain-research/realistic-ssl-evaluation
Open Datasets Yes We tested each SSL approach on the widely-reported image classification benchmarks of SVHN [40] with all but 1000 labels discarded and CIFAR-10 [31] with all but 4,000 labels discarded.
Dataset Splits Yes We optimized hyperparameters to minimize classification error on the standard validation set from each dataset, as is common practice (an approach we evaluate critically in section 4.6).
Hardware Specification No For every SSL technique in addition to a fully-supervised (not utilizing unlabeled data) baseline, we ran 1000 trials of Gaussian Process-based black box optimization using Google Cloud ML Engine s hyperparameter tuning service [18].
Software Dependencies No The paper mentions software components like 'Adam optimizer' and 'Wide Res Net' but does not specify their version numbers or the versions of underlying libraries or programming languages.
Experiment Setup Yes We chose a Wide Res Net [52], due to their widespread adoption and availability. Specifically, we used WRN-28-2... For training, we chose the ubiquitous Adam optimizer [29]. For all datasets, we followed standard procedures for regularization, data augmentation, and preprocessing; details are in appendix B. ... An enumeration of these hyperparameter settings can be found in appendix C.