Yun Sing Koh

Yun’s co-authored research tackles a recurring reliability risk in machine learning: using past experience that no longer fits current conditions. SELeCT and FiCSUM offer evaluated methods for recognising changing patterns and identifying relevant learning. The public SELeCT implementation gives researchers and engineers a way to inspect and experiment with adaptation, supporting practical scrutiny of prediction quality over time.

Yun Sing Koh is Professor of Computer Science at the University of Auckland and Deputy Head of the School of Computer Science, specialising in adaptive machine learning systems that adjust as incoming data changes. A central concern is concept drift, where changing conditions make previously learned patterns less reliable and undermine prediction accuracy when the world no longer matches the data used to train a model.

Yun co-authored SELeCT, a method that continually estimates which past learning remains relevant and reuses earlier models when familiar conditions return. Comparative experiments reported improved accuracy, and a public implementation includes tests and demonstration code, making the approach available for inspection and experimentation.

In further collaborative research, Yun helped introduce FiCSUM, an approach that builds detailed fingerprints of data patterns using information both with and without known answers. Evaluated across 11 real-world and synthetic datasets with stronger results than comparison methods, the technique supports ongoing investigations into how learning systems recognise change and retain useful experience.