Optimal Tensor Transport

Authors: Tanguy Kerdoncuff, Rémi Emonet, Michael Perrot, Marc Sebban7124-7132

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we illustrate the interest of OTT on two different tasks. First, following the success of OT in Domain Adaptation (Courty et al. 2016), we propose to predict the genres of recent movies based on labeled older movies by relying only on users preferences. We advantageously use a 3D-tensor formulation to take into account the particularity of each user. In a second experiment, we use the OTT barycenter in a Comparison-Based Clustering task.
Researcher Affiliation Academia Tanguy Kerdoncuff1, R emi Emonet1, Micha el Perrot2, Marc Sebban1 1 Univ Lyon, UJM-Saint-Etienne, CNRS, Institut d Optique Graduate School, Laboratoire Hubert Curien UMR 5516, F-42023, Saint-Etienne, France 2 Univ. Lille, Inria, CNRS, Centrale Lille, UMR 9189 CRISt AL, F-59000 Lille, France
Pseudocode Yes Algorithm 1: OTT
Open Source Code Yes 1The code to reproduce all the experiments is available online: https://github.com/Hv0nnus/Optimal Tensor Transport
Open Datasets Yes We consider a DA task on the Movielens dataset (Harper and Konstan 2015)... We consider some 3-class unbalanced subsamples of the MNIST dataset (Le Cun et al. 1998).
Dataset Splits No The paper discusses hyperparameter tuning and evaluation on source/target domains, and refers to 'target labels available' in experiments. However, it does not explicitly provide specific train/validation/test dataset splits with percentages, sample counts, or references to predefined splits for reproducibility.
Hardware Specification No The paper discusses the computational complexity and efficiency of its algorithm but does not provide any specific details about the hardware (e.g., GPU model, CPU type, memory) used to run the experiments.
Software Dependencies No The paper refers to various algorithms and methods such as 'OT Sinkhorn solver', 'semi-supervised algorithm OTDA', 'SVM', and 'k-means', but it does not provide specific version numbers for any software, libraries, or dependencies used in the experiments.
Experiment Setup Yes The Kullback-Leibler regularization parameter ϵ of the Sinkhorn method (Cuturi 2013) is selected in the range [10 5, 102] and the class regularization η of OTDA (Courty et al. 2016) in [10 4, 101]. The hyperparameters selection is limited to 24 hours for each method and dataset... As the initialization is key to avoid local minima, we take advantage of both the labels and our stochastic algorithm by sampling only the labelled points in the source and target for the first gradient estimation. The squared euclidean loss is used for L and we estimate the gradient of OTT using M = 1000 samples. S is set to 1000 iterations in Algorithm 1... We use default hyperparameters, reported in the supplementary material, for t-STE, Add S3, and OTT with the KL regularization parameter set to ϵ = 0.1. To ensure convergence, we also set the number of samples M = 100 and the number of iteration S = 500 between each of the 20 barycenter updates.