Title: Leveraging Synthetic Noise for Semi-Supervised Domain Adaptation

Abstract: Transfer learning typically relies on a source domain populated with semantically rich data, such as images, to effectively impart knowledge to a target domain. However, recent research has revealed an unexpected capability: a noise domain generated from simple distributions, like Gaussian noise, can act as a viable surrogate source in semi-supervised scenarios where labeled target data is scarce. Capitalizing on this finding, we introduce a new problem formulation called Semi-Supervised Noise Adaptation (SSNA), designed to utilize synthetic noise domains to enhance the generalization capabilities of the target domain. To tackle this challenge, we derive a generalization bound that quantifies the impact of the noise domain on performance, and subsequently propose the Noise Adaptation Framework (NAF). Our comprehensive experiments show that NAF successfully harnesses the noise domain to tighten the target domain’s generalization bound, resulting in superior performance. The source code is publicly accessible at https://github.com/AIResearch-Group/SSNA.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC