Three-Way Trade-Off in Multi-Objective Learning: Optimization, Generalization and Conflict-Avoidance

Authors: Lisha Chen, Heshan Fernando, Yiming Ying, Tianyi Chen

JMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on various multi-task learning benchmarks are performed to demonstrate the practical applicability.
Researcher Affiliation Academia Lisha Chen EMAIL Department of Electrical, Computer & Systems Engineering Rensselaer Polytechnic Institute, United States Heshan Fernando EMAIL Department of Electrical, Computer & Systems Engineering Rensselaer Polytechnic Institute, United States Yiming Ying EMAIL School of Mathematics and Statistics University of Sydney, NSW, Australia Tianyi Chen EMAIL Department of Electrical, Computer & Systems Engineering Rensselaer Polytechnic Institute, United States
Pseudocode Yes Algorithm 1 Regularized MGDA Algorithm 2 Mo Do Stochastic MGDA Algorithm 3 SMG (Liu and Vicente, 2021) Algorithm 4 Mo Co (Fernando et al., 2023)
Open Source Code Yes Code is available at https://github.com/heshandevaka/Trade-Off-MOL.
Open Datasets Yes We further verify our theory in the NC case on MNIST image classification (Le Cun, 1998) using a multi-layer perceptron and three objectives: cross-entropy, mean squared error (MSE), and Huber loss.
Dataset Splits Yes The training, validation, and testing data sizes are 50k, 10k, and 10k, respectively.
Hardware Specification Yes Experiments are done on a machine with GPU NVIDIA RTX A5000.
Software Dependencies Yes We use MATLAB R2021a for the synthetic experiments in strongly convex case, and Python 3.8, CUDA 11.7, Pytorch 1.8.0 for other experiments.
Experiment Setup Yes The default parameters are T = 100, α = 0.01, γ = 0.001. In other words, in Figure 3a, we fix α = 0.01, γ = 0.001, and vary T; in Figure 3b, we fix T = 100, γ = 0.001, and vary α; and in Figure 3c, we fix T = 100, α = 0.01, and vary γ.