Xian-Jin Gui

Xian-Jin Gui

Huawei Technologies (China) · Nanjing University
2 previous affiliations
2Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Xian-Jin Gui focuses on machine learning in non-stationary environments, specifically addressing unsupervised domain adaptation and learning with noisy labels. The available evidence indicates a concentrated research trajectory aimed at improving algorithmic robustness when source data is corrupted by label, feature, or open-set noise. Recent work investigates theoretical bounds and curriculum learning strategies for mitigating these multiple noise sources.

Research Trajectory

The available publication record for Xian-Jin Gui is highly concentrated, reflecting a focused research effort between 2020 and 2023 on the problem of distribution shifts in the presence of data corruption. Operating primarily as a second author, Gui has contributed to the research direction of Robust Domain Adaptation for Corrupted Labels. Standard unsupervised domain adaptation typically assumes a clean source domain and an unlabeled target domain. Gui's collaborative work addresses the more realistic and challenging scenario where the source domain itself is compromised by corrupted labels or features.

In 2020, this focus resulted in the conference publication Towards Accurate and Robust Domain Adaptation under Noisy Environments. This paper provided a theoretical analysis of how harmful noise impacts unsupervised domain adaptation. To mitigate these effects, the authors proposed an offline curriculum learning approach designed to minimize a newly defined empirical source risk. This methodology connects closely with the broader area of Robust Learning with Noisy Labels via Sample Selection and themes such as Dynamic Sample Selection for Robust Learning with Noisy Labels, where models must be prevented from memorizing incorrect labels during training.

This line of inquiry was subsequently expanded into a comprehensive journal article in 2023, Towards Accurate and Robust Domain Adaptation Under Multiple Noisy Environments. Also participating as a second author, Gui helped extend the theoretical framework to account for multiple types of noise simultaneously, including label noise, feature noise, and open-set noise. The work demonstrated that different noise types exert disparate impacts on the expected target risk, necessitating more sophisticated curriculum learning strategies to isolate and eliminate the effects of source noise. This research also touches upon work in the area of Temporal Prediction Stability for Robust Learning, as maintaining reliable predictions over time is a core challenge when learning from corrupted, non-stationary data.

Current Directions

Based on the most recent paper in the evidence from 2023, Gui's research investigates the theoretical and practical challenges of multiple noisy environments in machine learning. The primary open problem addressed in this recent work is how to maintain algorithmic robustness when a source domain is simultaneously afflicted by label corruption, feature noise, and open-set noise. By analyzing the disparate impacts of these distinct noise types on expected target risk, the work points toward the development of more advanced offline curriculum learning techniques that can systematically filter out complex source noises before adapting models to new target distributions.

Profile updated

See something wrong? Report a mistake