Mitigating Negative Transfer via Reducing Environmental Disagreement.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Unsupervised Domain Adaptation (UDA) aims to transfer knowledge from labeled source to unlabeled target domains.
• Domain shift presents a significant challenge, often leading to negative transfer and reduced model performance.
• Mitigating negative transfer is crucial for effective UDA.

Purpose of the Study:

• To investigate the root causes of negative transfer in UDA.
• To propose a novel method for reducing negative transfer by addressing environmental disagreement.
• To improve the performance of UDA models facing significant domain shifts.

Main Methods:

• The study employs causally disentangled learning to analyze negative transfer.
• A new method, Reducing Environmental Disagreement (RED), is proposed.
• RED disentangles samples into causal and non-causal environmental features using adversarial training and reduces environmental disagreement.

Main Results:

• Theoretical analysis identifies cross-domain discriminative disagreement on non-causal features as a cause of negative transfer.
• Experimental results demonstrate that RED effectively mitigates negative transfer.
• The proposed RED method achieves state-of-the-art performance in UDA tasks.

Conclusions:

• Environmental disagreement, arising from overreliance on non-causal features, is a key factor in negative transfer.
• RED successfully disentangles features and reduces environmental disagreement, thereby improving UDA performance.
• The findings offer a new perspective and effective solution for addressing domain shift challenges in UDA.