Scalable Global Optimisation Framework for Machine Learning

Artificial Intelligence (AI) and Machine Learning (ML) have emerged as transformative technologies, with ML models often relying on optimization algorithms for constructing predictive models. In particular, algorithms like stochastic gradient descent (SGD) are central to training ML models, minimizing data-fitting errors to determine optimal parameters. However, current optimization methods face critical challenges, and improving them can significantly advance ML technology.

This project aims to develop a new scalable global optimization framework capable of addressing key bottlenecks in modern ML optimization problems: Large Datasets, High Dimentionality and Complex Model Structure. Existing methods address these issues only partially. For example, SGD reduces computation costs but struggles with escaping local minima. Conversely, traditional global optimization methods are infeasible due to their high computational demands in large-scale problems.

This project seeks to overcome these limitations by establishing a novel, scalable global optimization framework. This framework will address the challenges of large datasets, high dimensionality, and complex structures, fundamentally improving the efficiency and effectiveness of optimization in modern ML applications.