Semi-supervised Conditional Density Estimation with Wasserstein Laplacian Regularisation

Authors: Olivier Graffeuille, Yun Sing Koh, Jörg Wicker, Moritz K Lehmann6746-6754

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the effectiveness of the WLR framework by evaluating and comparing the performance of MDN and our proposed WMDN. All experiments are implemented in Py Torch (Paszke et al. 2019), running on a Ge Force RTX 3080 GPU.
Researcher Affiliation Academia Olivier Graffeuille1, Yun Sing Koh1, J org Wicker1, Moritz Lehmann2 1 School of Computer Science, The University of Auckland 2 Xerra Earth Observation Institute, The University of Waikato
Pseudocode Yes The pseudocode for the WMDN is provided in Algorithms 1 and 2; Lines 9 to 11 in Algorithm 1 describe the WLR framework.
Open Source Code Yes A full list of algorithm parameters, datasets, and source code are available online1. 1https://github.com/OGraffeuille/Wasserstein-Laplacian Regularisation
Open Datasets Yes on seven regression datasets from the UCI repository (Dua and Graff 2017). UCI Machine Learning Repository. http://archive.ics.uci.edu/ml. University of California, Irvine, School of Information and Computer Sciences. Accessed 2021-06-01.
Dataset Splits Yes A validation set of size 1,000 was used for early stopping, and up to 10,000 data points were used for the test set. For each dataset, we use 3,000 training data points, of which 100, 300 or 1,000 are labelled. use a 2,000-300-312 for a train-validation-test data split for the Real dataset.
Hardware Specification Yes All experiments are implemented in Py Torch (Paszke et al. 2019), running on a Ge Force RTX 3080 GPU.
Software Dependencies No The paper mentions 'Py Torch (Paszke et al. 2019)' but does not specify a version number for PyTorch or other software dependencies.
Experiment Setup Yes We used a neural network architecture of three fully connected layers of 32 Re LU activated neurons to allow for sufficient representation capacity to model datasets of various dimensions... We used c = 5 Gaussian mixture components as the output parameter vector of MDNs and WMDNs... We set d = 1 as L1 distance retains meaningfulness in high dimensional space, and set q = 2, b = 20, k = 5 for all experiments as default values. To tune η, MDNs were trained on each dataset with η {10n|n = 5, 4.5, . . . , 2}. We selected ηbest as the η which produced the lowest average validation NLL. WMDNs were then trained with η = ηbest, λu {10n|n = 2, 1.5, . . . , 2} to select λu. Experiments were run with 50 seeds for each parameter combination.