Learning Generative Models across Incomparable Spaces

Authors: Charlotte Bunne, David Alvarez-Melis, Andreas Krause, Stefanie Jegelka

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we empirically demonstrate the effectiveness of the GW GAN formulation and regularization, and illustrate its versatility by tackling various novel settings for generative modeling, including learning across different dimensionalities, data types and styles.
Researcher Affiliation Academia 1Department of Computer Science, Eidgen ossische Technische Hochschule (ETH), Z urich, Switzerland 2Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology (MIT), Cambridge, USA.
Pseudocode Yes Algorithm 1 Training Algorithm of the Gromov-Wasserstein Generative Model.
Open Source Code No The paper does not provide any specific links to code repositories or explicit statements about public code release.
Open Datasets Yes To illustrate the ability of the GW GAN to generate images, we train the model on MNIST (Le Cun et al., 1998), fashion-MNIST (Xiao et al., 2017) and gray-scale CIFAR10 (Krizhevsky et al., 2014).
Dataset Splits No The paper mentions 'validation' (e.g., 'Empirical validation (see Appendix A) confirms this') but does not provide specific dataset split information (percentages, sample counts, or references to predefined splits) needed to reproduce the data partitioning for training, validation, and test sets.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments (e.g., specific GPU/CPU models, memory, or cloud computing instance types).
Software Dependencies No The paper mentions using Adam optimizer and ReLU activation functions but does not specify software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow, CUDA versions).
Experiment Setup Yes We train the model using Adam with a learning rate of 2 10 4, β1 = 0.5, β2 = 0.99 (Kingma & Ba, 2015). For the experiments on synthetic datasets, generator and adversary architectures are multilayer perceptrons (MLPs) with Re LU activation functions. Algorithm 1 Training Algorithm of the Gromov-Wasserstein Generative Model. Require: α: learning rate, ng: the number of iterations of the generator per adversary iteration, m: mini-batch size, N: number of training iterations, θ0: initial parameters of generator gθ, ω0 = (ˇω0, ˆω0): initial parameters of adversary fω. We found that a total variation regularization (Rudin et al., 1992) induces the right bias here and hence greatly improves the results (see Figures 2b, c, and d). The adversary was constrained to approximate an orthogonal operator.