Multitask Causal Contrastive Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Causal Inference

Background:

• Multitask learning (MTL) enhances performance by sharing knowledge across tasks.
• MTL often faces negative transfer due to useless feature utilization and task interference.
• Confounding factors in features and inadequate task relationship measurement contribute to MTL issues.

Purpose of the Study:

• To propose a novel multitask causal contrastive learning (MT-CCL) approach.
• To address negative transfer and interference in multitask learning.
• To improve the effectiveness and efficiency of multitask learning models.

Main Methods:

• Developed a multitask causal inference method using task-aware causal intervention (TACI).
• Quantified intertask relationships through intertask causal affinity.
• Implemented a dual contrastive learning objective with intratask contrast (Intra-TCS) and intertask contrast (Inter-TCS).

Main Results:

• MT-CCL demonstrated superior performance compared to state-of-the-art methods.
• Experiments were conducted on Multi-MNIST, NYU-v2, CityScapes, and CelebA datasets.
• The effectiveness of intertask causal affinity in improving multitask learning was verified.

Conclusions:

• MT-CCL successfully mitigates negative transfer and task interference in multitask learning.
• The proposed causal inference and contrastive learning methods enhance feature utilization and task synergy.
• MT-CCL offers a promising direction for advancing multitask learning research and applications.