Multitask Causal Contrastive Learning

Multitask learning (MTL) aims to improve the performance of multiple tasks by sharing knowledge among multiple different tasks, which has attracted increasing interest and shown success in various fields. However, MTL often suffers from negative transfer since the model may utilize useless features and face interference among tasks' optimization objectives. The utilization of useless features can be attributed to the confounding factors in multitask features, while the interference among tasks' optimization objectives is due to inadequate measurement of the relationship among tasks. This article proposes a novel multitask causal contrastive learning (MT-CCL) approach to address the above problem. First, we propose a multitask causal inference method, which removes the confounding factors in features via task-aware causal intervention (TACI) and measures the relationship among tasks from a novel causal perspective by quantifying the intertask causal affinity. Then, we build a dual contrastive learning objective to help the model better learn useful features via intratask contrast (Intra-TCS) and mitigate interference among tasks' optimization objectives via intertask contrast (Inter-TCS). Experiments demonstrate that MT-CCL achieves improved performance over state-of-the-art methods on Multi-MNIST, NYU-v2, CityScapes, and CelebA, verifying the effectiveness of intertask causal affinity.