Learning from noisy data
Project Level: Summer
Project Duration: 6 weeks
Hours of Engagement: 20-36 hours
Project Description:
Machine learning algorithms often assume data is correctly labelled, but in practice the labelling process is often error-prone, and it is reported that the ratio of corrupted labels range from 8%-38.5% in various real-world datasets. This project investigates the robustness of various machine learning algorithms against label noise, and aims to develop highly robust machine learning algorithms.
Expected Outcomes:
• Develop a general understanding of general robust approaches for learning from noisily labelled data
• Develop empirical and/or theoretical understanding of the robustness of various machine learning algorithms, and develop new robust machine learning algorithms
• Develop skills in research design, implementation, experimentation, and communication.
• A report documenting the work done and the findings.
Suitable for:
Essential: knowledge of machine learning, strong programming skills.
Desirable: knowledge of deep learning
Contact for further information:
Dr Nan Ye: nan.ye@uq.edu.au