Multi-Task Learning Model Outperforms Traditional Approach in Application Scenarios

Strategic Insights -

Researchers at Beijing University of Technology have made a breakthrough in machine learning, developing a multi-task learning (MTL) model that outperforms the traditional approach in various application scenarios. The new model, known as Multi-Task Parallel Model (MTPM), uses fuzzy neural networks and joint gradient descent algorithm to optimize parameters across parallel tasks, achieving an average RMSE improvement of 24.2% across three scenarios.

Key Takeaways:

  • The researchers proposed a multi-task parallel model (MTPM) based on fuzzy neural networks (FNNs) and joint gradient descent algorithm (JGDA) to solve the problem of tuning parameters that equally benefit all tasks.
  • The MTPM model extracts interaction features and specificity features from multiple related tasks using FNN, and uses JGDA to optimize parameters across parallel tasks.
  • The model avoids conflict between correlation between tasks and independence of each task by embedding shared neurons and specific neurons into individual FNN rather than single-type neurons.
  • An adaptive learning rate strategy is designed to further improve the model's performance, coordinating the training process of different tasks.
  • The experimental results show that the MTPM model outperforms the traditional method in MTL effects, achieving an average RMSE improvement of 24.2% across three application scenarios.
  • The model demonstrates its effectiveness and practicality in real-world applications, showcasing its potential in addressing complex tasks through multi-task learning.

Statistics:

  • 24.2% average RMSE improvement of the MTPM model across three application scenarios compared to single-task benchmark experiments.
  • The MTPM model used fuzzy neural networks (FNNs) to extract interaction features and specificity features from multiple related tasks.
  • The joint gradient descent algorithm (JGDA) was used to optimize parameters across parallel tasks.
  • 15.19:1038 (ISSN) of the journal Applied Sciences in which the research article was published.
  • 3 application scenarios in which the MTPM model was tested and outperformed the traditional method.

Sources:

  • Design of Multi-Task Parallel Model Based on Fuzzy Neural Networks and Joint Gradient Descent Algorithm, Applied Sciences, 2025,15(19):10386.
  • Beijing University of Technology, Faculty of Information Technology, Beijing 100124, People's Republic of China.
  • Xiaolong Wu, Yan Zhao, Yanxia Yang, researchers at Beijing University of Technology, co-authors of the research article.