Shedding a PAC-Bayesian Light on Adaptive Sliced-Wasserstein Distances

Authors: Ruben Ohana, Kimia Nadjahi, Alain Rakotomamonjy, Liva Ralaivola

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

Reproducibility Variable Result LLM Response
Research Type Experimental We provide three types of results: i) PAC-Bayesian generalization bounds that hold on what we refer as adaptive Sliced Wasserstein distances, i.e. SW defined with respect to arbitrary distributions of slices (among which data-dependent distributions), ii) a principled procedure to learn the distribution of slices that yields maximally discriminative SW, by optimizing our theoretical bounds, and iii) empirical illustrations of our theoretical findings.
Researcher Affiliation Collaboration 1Flatiron Institute, USA 2MIT, USA 3Criteo AI Lab, France.
Pseudocode Yes Algorithm 1 PAC-SW: Adaptive SW via PAC-Bayes bound optimization. ... Algorithm A2 PAC-Bayes bound optimization for v MF-based SW
Open Source Code Yes All our numerical experiments presented in Section 5 can be reproduced using the code we provided in https://github.com/rubenohana/PACBayesian_Sliced-Wasserstein.
Open Datasets Yes We consider a generative modeling task on MNIST data (Deng, 2012)
Dataset Splits No The paper mentions training and test data, but does not explicitly describe a validation split or its methodology.
Hardware Specification Yes Timing results of this experiment were obtained with a NVIDIA GPU A100 80 GB, compared to Figure 4 which was on a NVIDIA V100.
Software Dependencies No The paper mentions using Adam (Kingma & Ba, 2015) with its default parameters, but no specific version numbers for software libraries or dependencies are provided.
Experiment Setup Yes We sample n = 500 samples from µ and ν and optimize ρ (µn, νn): the optimization is performed on the space of v MF distributions, using Adam (Kingma & Ba, 2015) with its default parameters. ... For each minibatch of size 512, the distribution ρ is learned by optimizing 100 projections over 100 iterations and the generative model is trained over 400 epochs.