Continuous Regularized Wasserstein Barycenters

Authors: Lingxiao Li, Aude Genevay, Mikhail Yurochkin, Justin M. Solomon

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the effectiveness of our approach and compare against previous work on synthetic examples and real-world applications.
Researcher Affiliation Collaboration Lingxiao Li MIT CSAIL lingxiao@mit.edu Aude Genevay MIT CSAIL aude.genevay@gmail.com Mikhail Yurochkin IBM Research, MIT-IBM Watson AI Lab mikhail.yurochkin@ibm.com Justin Solomon MIT CSAIL, MIT-IBM Watson AI Lab jsolomon@mit.edu
Pseudocode Yes Algorithm 1: Stochastic gradient descent to solve the regularized barycenter problem (11)
Open Source Code Yes The source code is publicly available at https://github.com/lingxiaoli94/CWB.
Open Datasets Yes We consider Poisson regression for the task of predicting the hourly number of bike rentals using features such as the day of the week and weather conditions.3 (Footnote 3: http://archive.ics.uci.edu/ml/datasets/Bike+Sharing+Dataset)
Dataset Splits No The paper describes splitting the data into 5 equally-sized subsets for posterior aggregation, but does not explicitly mention a 'validation set' or a dedicated validation split in the context of model training or evaluation.
Hardware Specification Yes We ran our experiments using a NVIDIA Tesla V100 GPU on a Google cloud instance with 12 computeoptimized CPUs and 64GB memory.
Software Dependencies Yes The stochastic gradient descent used to solve (11) and (14) is implemented in Tensorflow 2.1 [Aba+16].
Experiment Setup Yes In all experiments below, we use Adam optimizer [KB14] with learning rate 10 4 and batch size 4096 or 8192 for the training. The dual potentials {fi, gi}n i=1 in (11) are each parameterized as neural networks with two fully-connected layers (d ! 128 ! 256 ! 1) using Re LU activations. Every Ti in (14) is parameterized with layers (d ! 128 ! 256 ! d).