Fair Resource Allocation in Multi-Task Learning

Authors: Hao Ban, Kaiyi Ji

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that our method can achieve state-of-the-art performance among gradient manipulation methods on a suite of multitask benchmarks in supervised learning and reinforcement learning. Furthermore, we incorporate the idea of α-fairness into the loss functions of various MTL methods. Extensive empirical studies demonstrate that their performance can be significantly enhanced.
Researcher Affiliation Academia 1Department of Computer Science and Engineering, University at Buffalo, New York, United States. Correspondence to: Kaiyi Ji <kaiyiji@buffalo.edu>.
Pseudocode Yes Algorithm 1 Fair Grad for MTL
Open Source Code Yes Code is available at https: //github.com/Opt MN-Lab/fairgrad.
Open Datasets Yes Celeb A (Liu et al., 2015) is a large-scale face attributes dataset...QM9 (Ramakrishnan et al., 2014) is a widely-used benchmark...NYU-v2 (Silberman et al., 2012) contains...Cityscapes (Cordts et al., 2016) contains...
Dataset Splits Yes Following (Navon et al., 2022; Liu et al., 2023), we use the example provided in Pytorch Geometric (Fey & Lenssen, 2019), and use 110k molecules for training, 10k for validation, and the rest 10k for testing.
Hardware Specification No The paper does not provide specific details about the hardware used for running experiments.
Software Dependencies No The paper mentions software like Pytorch Geometric, Soft Actor-Critic (SAC), and MTRL codebase, but does not provide specific version numbers for these software dependencies, which are necessary for reproducible descriptions.
Experiment Setup Yes We train our method for 15 epochs, using Adam optimizer with learning rate 3e-4. The batch size is 256.