FAMO: Fast Adaptive Multitask Optimization

Authors: Bo Liu, Yihao Feng, Peter Stone, Qiang Liu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct an extensive set of experiments covering multi-task supervised and reinforcement learning problems. Our results indicate that FAMO achieves comparable or superior performance to state-of-the-art gradient manipulation techniques while offering significant improvements in space and computational efficiency.
Researcher Affiliation Collaboration Bo Liu, Yihao Feng, , Peter Stone, Qiang Liu The University of Texas at Austin, Salesforce AI Research, Sony AI {bliu, pstone, lqiang}@cs.utexas.edu, yihaof@salesforce.com
Pseudocode Yes Algorithm 1 Fast Adaptive Multitask Optimization (FAMO)
Open Source Code Yes Code is available at https://github.com/Cranial-XIX/FAMO.
Open Datasets Yes We consider four supervised benchmarks commonly used in prior MTL research [24, 27, 32, 33]: NYU-v2 [31] (3 tasks), City Scapes [8] (2 tasks), QM-9 [1] (11 tasks), and Celeb A [28] (40 tasks). ... We use 110K molecules from the QM9 example in Py Torch Geometric [10], 10K molecules for validation, and the rest of 10K molecules for testing.
Dataset Splits Yes We use 110K molecules from the QM9 example in Py Torch Geometric [10], 10K molecules for validation, and the rest of 10K molecules for testing.
Hardware Specification No The paper does not explicitly describe the hardware used for running its experiments. It states, 'We followed the exact experimental setup from NASHMTL [32]', implying reliance on a referenced paper for such details, but no direct hardware specifications are given within this document.
Software Dependencies No The paper mentions 'Py Torch Geometric [10]' and 'Soft Actor-Critic (SAC) [15]' but does not provide specific version numbers for these or other software dependencies.
Experiment Setup No The paper states: 'The experimental setting and hyperparameters all match exactly with those in CAGRAD.' It refers to an external paper for these details rather than providing them directly within the text.