A Tree-Structured Multitask Model Architectures Recommendation System.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Multitask learning (MTL) utilizes neural networks with branched architectures, specifically tree-structured models, to address multiple vision tasks simultaneously.
• Designing these tree-structured networks involves optimizing the branching points from a shared backbone to balance task accuracy and computational efficiency.

Purpose of the Study:

• To propose an automated recommendation system for generating tree-structured multitask learning architectures.
• To enable the automatic suggestion of architectures that meet user-defined computational budgets while maximizing task performance.

Main Methods:

• Development of a recommendation system that takes a set of tasks and a convolutional neural network (CNN) backbone as input.
• The system automatically suggests optimal tree-structured multitask architectures without requiring model training.
• Evaluation of recommended architectures on popular multitask learning benchmarks.

Main Results:

• The proposed recommendation system effectively suggests tree-structured multitask architectures.
• Evaluated architectures demonstrate competitive task accuracy and computational efficiency compared to existing state-of-the-art MTL methods.
• The system successfully balances performance and computational constraints.

Conclusions:

• The automated recommendation system provides an efficient solution for designing tree-structured multitask learning architectures.
• This approach offers a practical method for optimizing model design for specific computational budgets and task requirements.
• The open-sourced recommender facilitates further research and application in multitask learning.