On Adversarial Mixup Resynthesis

Authors: Christopher Beckham, Sina Honari, Vikas Verma, Alex M. Lamb, Farnoosh Ghadiri, R Devon Hjelm, Yoshua Bengio, Chris Pal

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section we evaluate the classification accuracy of AMR on various datasetss by training a linear classifier on the latent features of the unsupervised variant of the model. We also measure evaluate our model on a disentanglement task, which is also unsupervised. Finally, we demonstrate some qualitative results.
Researcher Affiliation Collaboration 1Mila Québec Artificial Intelligence Institute, Montréal, Canada 2Université de Montréal, Canada 3Polytechnique Montréal, Canada 4Element AI, Montréal, Canada 5Microsoft Research, Montréal, Canada 6Aalto University, Finland
Pseudocode No The paper provides mathematical formulations and descriptions of the model and mixing functions, but does not include any structured pseudocode or algorithm blocks.
Open Source Code Yes Code provided here: https://github.com/christopher-beckham/amr
Open Datasets Yes We employ the following datasets for classification: MNIST (Deng, 2012), KMNIST (Clanuwat et al., 2018), and SVHN (Netzer et al., 2011). ... Lastly, we run experiments on the DSprite (Matthey et al., 2017) dataset...
Dataset Splits Yes We perform three runs for each experiment, and from each run we collect the highest accuracy on the validation set over the entire course of training, from which we compute the mean and standard deviation.
Hardware Specification Yes We thank Compute Canada for GPU access, and n Vidia for donating a DGX-1 used for this research.
Software Dependencies No The paper mentions 'Py Torch' (footnote 4) and 'ADAM' (optimizer), but it does not specify version numbers for these or any other software libraries required for reproducibility.
Experiment Setup Yes Hyperparameter tuning on λ was performed manually (this essentially controls the trade-off between the reconstruction and adversarial losses), and we experimented with a reasonable range of values (i.e. {2, 5, 10, 20, 50}. In terms of training hyperparameters, we used ADAM (Kingma & Ba, 2014) with a learning rate of 10 4, β1 = 0.5 and β2 = 0.99 and an L2 weight decay of 10 5. ... MNIST, KMNIST, and SVHN were trained for 2k, 5k, and 4.5k epochs, respectively.