This repository is a comprehensive, up-to-date list of resources for Multi-Task Learning (MTL), curated from a machine learning perspective. It serves researchers and practitioners by providing categorized links to datasets, codebases, papers, and surveys covering various aspects of MTL, from fundamental architectures to advanced optimization techniques. The goal is to offer a centralized hub for understanding and implementing MTL across diverse domains like computer vision, NLP, and reinforcement learning.

How It Works

The repository is structured into key areas of MTL research: Datasets, Codebases, Architectures (e.g., Hard/Soft Parameter Sharing, MoE, Adapters), Optimization strategies (e.g., Loss/Gradient strategies, Task Interference, Pareto optimization), Task Relationship Learning, and Theory. It compiles seminal and recent works, offering a broad overview of the field's evolution and current state.

Quick Start & Requirements

This is a curated list, not a runnable codebase. Users will need to follow links to individual projects for installation and execution.

Highlighted Details

• Extensive categorization of datasets and codebases across Computer Vision, NLP, RL, Graph, and Recommendation domains.
• Detailed breakdown of MTL architectures, including parameter sharing, adapters, MoE, and task representation methods.
• Comprehensive coverage of optimization techniques, addressing gradient conflicts, task interference, and adaptive weighting.
• Includes a significant section on theoretical underpinnings and task relationship learning.

Maintenance & Community

The project is actively maintained and welcomes contributions. Contact information for the curator is provided. Links to relevant surveys and foundational papers are included.

Licensing & Compatibility

This repository itself is a list and does not have a specific license. Individual linked codebases and datasets will have their own licenses, which users must adhere to.

Limitations & Caveats

As a curated list, it does not provide direct functionality or code. Users must navigate to external resources for implementation. The rapidly evolving nature of MTL means some information may become dated, though the list aims for recency.