Choose Wisely, Learn Efficiently: Effective Data and Model Selection Strategies for Efficient Machine Learning

The Hong Kong University of Science and Technology (Guangzhou)

数据科学与分析学域

PhD Thesis Examination

By Mr. Hanmo LIU

摘要

This thesis confronts the challenge of enhancing machine learning efficiency by establishing a symbiotic relationship between data and models through mutual selection. It diverges from traditional, unidirectional training pipelines by proposing a framework where models and data intelligently inform one another to optimize knowledge transfer and performance. The research is bifurcated into two primary streams. The first, data-to-model selection, introduces the Knowledge Benchmark Graph to leverage historical performance data, enabling the strategic selection of appropriate models for new datasets without exhaustive retraining. The second, model-to-data selection, presents novel methods for identifying and utilizing the most informative data subsets in complex learning environments. This includes EDSR (Effective Data Selection and Replay) for unsupervised continual learning, which prioritizes high-entropy data, and LTF (Learning Towards the Future) for dynamic temporal graphs, which manages evolving data distributions and emerging classes. Collectively, these methods advance a unified perspective on adaptive learning systems, demonstrating improvements in efficiency and knowledge retention while maintaining high performance.

TEC

Chairperson: Prof Sihong XIE
Prime Supervisor: Prof Can YANG
Co-Supervisor: Prof Lei CHEN
Examiners:
Prof Qiong LUO
Prof Yongqi ZHANG
Prof Guang ZHANG
Prof Xiaokui XIAO