MSc research project (mid-program study, not the final thesis) comparing coreset selection strategies for efficient ML model training.
The aim of this project is to investigate the potential of coreset-based techniques as a means of improving the efficiency of machine learning model training. In particular, the study seeks to understand how different coreset selection strategies influence both model accuracy and computational cost. To achieve this, several well-known strategies such as random sampling, k-Center greedy selection, and herding will be implemented and systematically compared. The project further aims to evaluate the trade-offs that arise when reducing dataset size, with a particular focus on the balance between faster training times and the possible loss of predictive performance. Through these experiments, the study intends to identify the conditions under which coreset-based training provides the greatest benefits, offering insights into practical scenarios where such methods can be applied effectively in resource-constrained environments.
The objectives mapped to this aim are:
- To implement three established coreset selection strategies (Random Sampling, k-Center Greedy, Herding).
- To compare their performance against a full-dataset baseline in terms of accuracy, training time, and compression ratio.
- To analyze how dataset complexity (simple vs. complex datasets) influences coreset effectiveness.
- To position the results relative to recent literature (e.g., CRAIG, CREST, CCS) to highlight strengths and limitations of classical methods.