TaskNorm: Rethinking Batch Normalization for Meta-Learning

Authors: John Bronskill, Jonathan Gordon, James Requeima, Sebastian Nowozin, Richard Turner

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on fourteen datasets demonstrate that the choice of batch normalization has a dramatic effect on both classification accuracy and training time for both gradient basedand gradientfree meta-learning approaches. Importantly, TASKNORM is found to consistently improve performance.
Researcher Affiliation Collaboration 1University of Cambridge 2Invenia Labs 3Microsoft Research.
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Source code is available at https://github.com/ cambridge-mlg/cnaps
Open Datasets Yes We provide experiments with MAML running simple, but widely used few-shot learning tasks from the Omniglot (Lake et al., 2011) and mini Imagenet (Ravi & Larochelle, 2017). Next, we evaluate NLs on a demanding few-shot classification challenge called Meta-Dataset, composed of thirteen (eight train, five test) image classification datasets (Triantafillou et al., 2020).
Dataset Splits No The paper mentions 'validation accuracy' and 'meta-training set' and 'meta-test phase', and that 'A task τ is drawn from p(τ) and randomly split into a context set Dτ and target set T τ', but it does not provide specific percentages or sample counts for train/validation/test splits of the datasets themselves in the main body. Details are deferred to Appendix B, which is not accessible.
Hardware Specification No The paper does not explicitly describe the hardware used to run its experiments, such as specific GPU or CPU models.
Software Dependencies No The paper mentions general frameworks like TensorFlow and PyTorch but does not provide specific version numbers for these or any other software dependencies needed to replicate the experiment.
Experiment Setup No The paper states that 'Configuration and training details can be found in Appendix B' and 'Details provided in Appendix B and Triantafillou et al. (2020)', but these specific experimental setup details (e.g., hyperparameters) are not provided in the main text.