Multi-Task Learning as a Bargaining Game

Authors: Aviv Navon, Aviv Shamsian, Idan Achituve, Haggai Maron, Kenji Kawaguchi, Gal Chechik, Ethan Fetaya

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we show that Nash-MTL achieves state-of-the-art results on multiple MTL benchmarks in various domains.
Researcher Affiliation Collaboration 1Bar-Ilan University, Ramat Gan, Israel 2Nvidia, Tel-Aviv, Israel 3National University of Singapore.
Pseudocode Yes Algorithm 1 Nash-MTL
Open Source Code Yes To support future research and the reproducibility of the results, we make our source code publicly available at: https://github.com/Aviv Navon/nash-mtl.
Open Datasets Yes We evaluate Nash-MTL on predicting 11 properties of molecules from the QM9 dataset (Ramakrishnan et al., 2014)... We use 110K molecules for training, 10K for validation, and 10K as a test set. We follow the protocol of (Liu et al., 2019b) and evaluate Nash-MTL on the NYUv2 and Cityscapes datasets (Silberman et al., 2012; Cordts et al., 2016)... We consider a multi-task RL problem and evaluate Nash MTL on the MT10 environment from the Meta-World benchmark (Yu et al., 2020b).
Dataset Splits Yes We use 110K molecules for training, 10K for validation, and 10K as a test set.
Hardware Specification No The paper does not explicitly state the specific hardware used (e.g., GPU models, CPU types) for running its experiments.
Software Dependencies No The paper mentions using Py Torch Geometric, Seg Net architecture, and Soft Actor-Critic (SAC) but does not provide specific version numbers for these software components or any other libraries.
Experiment Setup Yes We train each method for 300 epochs and search for the best learning-rate (lr)... Each method is trained for 200 epochs with the Adam optimizer (Kingma & Ba, 2015) and an initial learning-rate of 1e 4. The learning-rate is halved to 5e 5 after 100 epochs. ... We use a batch size of 2 and 8 for NYUv2 and City Scapes respectively. ... Each method is trained over 2 million steps with a batch size of 1280.