Learning Task-Preferred Inference Routes for Gradient De-Conflict in Multi-Output DNNs.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Multi-output neural networks (MONs) feature shared filters across tasks, leading to entangled inference routes.
• Divergent optimization objectives in MONs cause task gradient interference along shared routes, hindering overall model performance.

Purpose of the Study:

• To propose a novel gradient de-conflict algorithm, DR-MGF (Dynamic Routes and Meta-weighted Gradient Fusion), for MONs.
• To address the issue of inter-task interference in MONs by learning task-preferred inference routes.

Main Methods:

• DR-MGF learns task-specific importance variables to evaluate filter importance for different tasks.
• Task dominance over filters is adjusted proportionally to task-specific filter importance, reducing inter-task interference.
• Task-specific importance variables dynamically determine task-preferred inference routes.

Main Results:

• DR-MGF effectively reduces inter-task interference in MONs.
• Experimental results on CIFAR, ImageNet, and NYUv2 datasets show DR-MGF outperforms existing de-conflict methods.
• The proposed DR-MGF method is extendable to general MONs without structural modifications.

Conclusions:

• DR-MGF offers an effective solution for gradient de-conflict in MONs.
• The dynamic route learning approach significantly improves MON performance by mitigating task interference.
• DR-MGF provides a flexible and generalizable method for enhancing multi-output deep learning models.