Ratio Trace Formulation of Wasserstein Discriminant Analysis

Authors: Hexuan Liu, Yunfeng Cai, You-Lin Chen, Ping Li

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

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical experiments on real datasets show promising results of the ratio trace formulation of WDA in both classification and clustering tasks.
Researcher Affiliation Industry Hexuan Liu, Yunfeng Cai, You-Lin Chen, Ping Li Cognitive Computing Lab Baidu Research No.10 Xibeiwang East Road, Beijing 100193, China 10900 NE 8th St. Bellevue, Washington 98004, USA {lhxuan93, yunfengcai09, cyoulin.tw, pingli98}@gmail.com
Pseudocode Yes Algorithm 1 WDA-eig algorithm Algorithm 2 Iterative WDA clustering
Open Source Code No The paper does not contain an explicit statement that the authors' code for the described methodology is open-source, nor does it provide a direct link to a code repository for their implementation.
Open Datasets Yes We extract 1000 samples in the MNIST dataset as the training set and use 10000 samples in the test set. We use four real world datasets to evaluate the proposed method: the MNIST dataset for digits recognition, the 15-scene dataset [19] for multi-class image recognition, the KTH action recognition database [33] for multi-class video recognition, and the 20 newsgroup dataset for text classification.
Dataset Splits No The paper specifies a training and test set split for MNIST but does not explicitly mention a separate validation set or its split percentage for any dataset.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU/GPU models, memory) used for running the experiments. It only mentions experiments were run and evaluated.
Software Dependencies No The paper mentions software like "pymanopt solvers", "Sinkhorn’s fixed-point iterations", "Greenkhorn algorithm", "APDAMD", "K-Nearest-Neighbors classifier (KNN)", "K-means", "FDA", "PCA", "LFDA" but does not specify version numbers for these components, which would be necessary for reproducible software dependencies.
Experiment Setup Yes For each λ, we run each algorithms for 100 randomly-initialized trials, and the results are shown in Table 1. In implementation of FDA/WDA-eig we add a small perturbation term ϵIp on Cw to make the denominator positive definite, and we choose ϵ = 2 in this setting... The Wasserstein regularizer λ is coarsely tuned, where we choose λ = 0.01 for MNIST and 15-scene, λ = 10 for KTH, and λ = 5 for 20ng. For (3) and (4) we use the subspace obtained by PCA as initialization and p = Nc 1 as the subspace dimensions. No regularization term is added to Cw.