A Constructive Prediction of the Generalization Error Across Scales

Authors: Jonathan S. Rosenfeld, Amir Rosenfeld, Yonatan Belinkov, Nir Shavit

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

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically explore the behavior of the generalization error over a wide range of datasets and models in vision and language tasks.
Researcher Affiliation Collaboration Jonathan S. Rosenfeld1 Amir Rosenfeld2 Yonatan Belinkov13 Nir Shavit145 {jonsr,belinkov,shanir}@csail.mit.edu amir@cse.yorku.ca 1 Massachusetts Institute of Technology 2 York University 3 Harvard University 4 Neural Magic Inc 5 Tel Aviv University
Pseudocode No The paper does not contain any pseudocode or algorithm blocks.
Open Source Code No The paper does not contain an explicit statement about open-sourcing code or a link to a code repository.
Open Datasets Yes Image Net (Russakovsky et al., 2015): a large-scale recognition benchmark... CIFAR10/100 (Krizhevsky et al., 2009)... DTD (Cimpoi et al., 2014)... Aircraft (Maji et al., 2013)... UCF101 (Soomro et al., 2012)... Penn Treebank (Mikolov et al., 2010)... Wiki Text-2 (Bradbury et al., 2017)... Wiki Text-103 (Merity et al., 2016).
Dataset Splits Yes CIFAR10/100 (Krizhevsky et al., 2009): 60K natural RGB images of 10 classes (100 for CIFAR100) with a train/test split of 50K/10K. ... PTB... 900K/70K/80K training/validation/test words.
Hardware Specification No The paper does not specify any particular hardware (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions software like PyTorch, SGD, and Adam but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes Hyper-parameters: For similar reasons we wish to avoid hyper-paramater search at large scales, and thus avoid the temptation to tune hyper-parameters accordingly (learning rate, regularization, etc.). Therefore, we hold all hyper-parameters fixed. ... In the main experiments, training is done via SGD with a momentum of 0.9, weight decay of 1e-4 and initial learning rate of 0.1. For Image Net we train for 90 epochs, decreasing the learning rate by a multiplicative factor of 0.1 after and 30 and after 60 epochs. We use a batch size of 16. For all other vision datasets we use a batch-size of 128. We begin training with a learning rate of 0.1, run for 200 epochs, and reduce by a multiplicative factor of 0.1 after 80, 120, and 160 epochs.