Tikhonov Regularization is Optimal Transport Robust under Martingale Constraints

Authors: Jiajin Li, Sirui Lin, Jose Blanchet, Viet Anh Nguyen

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

Reproducibility Variable Result LLM Response
Research Type Experimental 5 Numerical Results In this section, we validate the effectiveness of our methods (referred to as martingale DRO) on both linear regression and deep neural networks under the adversarial setting.
Researcher Affiliation Academia Jiajin Li Sirui Lin José Blanchet Stanford University {jiajinli, siruilin, jose.blanchet}@stanford.edu Viet Anh Nguyen Chinese University of Hong Kong nguyen@se.cuhk.edu.hk
Pseudocode Yes Algorithm 1: Martingale Distributionally Robust Optimization with Adversarial Training
Open Source Code No The paper does not provide an explicit link to a source code repository for the methodology described, nor does it state that the code is publicly released within the main text.
Open Datasets Yes We test our method on three LIBSVM regression real world datasets 1. More specifically, we randomly select 60% of the data to train the models and the rest as our test data. Then, we validate our method on the MNIST dataset [15]. Experimental setup for CIFAR-10 [14]:
Dataset Splits Yes More specifically, we randomly select 60% of the data to train the models and the rest as our test data.
Hardware Specification Yes All simulations are implemented using Python 3.8 on: (1) a computer running Windows 10 with a 2.80GHz, Intel(R) Core(TM) i7-1165G7 processor and 16 GB of RAM, and (2) Google Colab with NVIDIA Tesla P100 GPU and 16 GB of RAM. The simulations are implemented using Python 3.8 on Google Colab with TPU v2 and 16GB RAM.
Software Dependencies Yes All simulations are implemented using Python 3.8 on:...
Experiment Setup Yes we choose the same hyperparameter ρ = 0.08 for ridge regression and martingale DRO mode for fair comparison. We optimize using Adam with a batch size of 128 for all methods. The learning rate starts from 0.01 and shrinks by 0.1 epoch total epochs , and each model is trained for 100 epochs.