Learning with Noisy Labels
In machine learning, labels are the foundational "truths" on which algorithms are trained. However, a significant number of real-world datasets contain noisy labels. When algorithms affected by label noise are deployed in critical environments—be it medical diagnostics, financial predictions, or safety-critical systems—the consequences can be not only costly but also life-altering or even life-threatening. Recognizing and addressing label noise is, therefore, not just a technical necessity but a matter of utmost responsibility.
(See our Paper)
Understanding
We have delved deep into understanding the mechanics of how label noise is generated and how such noise influences the training and generalization capabilities of models through both experiments and rigorous theoretical analyse.
Our Research
​Our research is centered on understanding and identifying label errors, ensuring that machine learning models are built upon a foundation of accuracy and trustworthiness.
(See our Paper)
Detection
Based on our insights into label noise, we have developed a series of tools and algorithms that assist researchers and engineers in effectively detecting label noise in datasets. These tools offer a quantitative noise assessment, enabling users to identify and pinpoint problematic data.
(See our Paper)
Correction
Beyond just detecting noise, we have also explored ways to correct these noises to optimize model performance. We've proposed a series of methodologies aimed at correcting or reducing the impact of label noise, thus enhancing the training stability and predictive accuracy of machine learning models.
Empowering Trustworthy Machine Learning for All
Advancing Academic Value
Our research stands as a significant contribution to the academic landscape of machine learning and data science. By addressing the complex issue of label noise, we've filled a crucial gap in current academic understanding. Our findings and methodologies, recognized and accepted by respected academic journals and conferences, add depth to existing knowledge and provide a clear direction for future research. Our detailed exploration of label noise, from its origins to methods for detection and correction, sets a standard for scholarly investigations in this area. Through our ongoing engagements and collaborations, we aim to strengthen the academic discourse on label noise, promoting rigorous research and informed discussions.
Democratization of Machine Learning
Our research not only offers a profound understanding of label noise but also brings forth new perspectives and methodologies for data processing. Crucially, our methods make machine learning development accessible and beneficial for entities of all sizes. Recognizing that not every organization, especially smaller companies, has access to flawless datasets, we present methods that provide a more systematic and scientific approach to addressing label noise. These invaluable tools empower even the smallest players to confidently step into the machine-learning landscape and make informed, data-driven decisions.
Potential for Broad Applications
Our methodologies offer cutting-edge solutions to the pervasive challenge of label noise. Designed with adaptability in mind, they find relevance across a myriad of sectors, from healthcare to finance and from natural language processing to computer vision. While real-world validations are ongoing, early indications suggest our methods will substantially shape best practices and redefine data quality improvement strategies across various domains.
​
Dr. Tongliang Liu
Director of Sydney AI Centre
Director of Trustworthy Machine Learning Lab (TML Lab)
ARC Future Fellow; ARC DECRA Fellow
School of Computer Science
Facult of Engineering
The University of Sydney, Australia
Visiting Associate Professor in Machine Learning
Department of Machine Learning
Mohamed bin Zayed University of Artificial Intelligence, United Arab Emirates
Dr. Tongliang Liu has published more than 200 papers at leading ML/AI conferences and journals. He has been widely recognised by his research. For example, he was ranked among the Best Rising Stars of Science in Australia by Research.com in 2022; he was ranked among the Global Top Young Chinese Scholars in AI by Baidu Scholar in 2022; he was named in the Early Achievers Leaderboard by The Australian in 2020. Tongliang received the ARC DECRA Award in 2018, ARC Future Fellowship Award in 2022, and IEEE AI's 10 to Watch Award in 2023.