Adaptive Information Allocation in Machine Learning

Stochastic gradient methods represent a key computational engine of large-scale modern machine learning. To estimate gradient directions, they sample training data uniformly at random, independently of their informational value. However, real-world datasets are heterogeneous, containing noise, outliers, imbalance, and unknown varying levels of deviations from the i.i.d. assumption. Consequently, gradient estimates obtained from uniformly sampled examples can vary dramatically in quality, leading to inefficient optimisation, unnecessary computation, and reduced robustness.

This project addresses a foundational question in machine learning: How should a learning system allocate its stochastic computational effort across data of unequal informational value?

We propose a new framework in which the sampling process in stochastic optimisation is not fixed, but learned jointly with model parameters - that is, an adaptive data selection or learned sampling policy. This transforms stochastic gradient descent into a coupled learning system consisting of: (1) a model learning process driven by stochastic gradients, and (2) a data selection process that adapts to data imperfections and to the evolving state of learning.

The key idea is that learning efficiency is determined not only by model capacity, but by how the optimisation process allocates attention across training data. The project will develop a principled theory, algorithms, and analysis of this coupled system. Specifically, its scope includes the following:

- develop a PAC-Bayes theory for learning under adaptive data selection,

- identify the conditions under which adaptive sampling provably improves optimisation efficiency, robustness, and generalisation,

- design scalable algorithms for joint optimisation of model parameters and sampling policies,

- evaluate performance in noisy, imbalanced, heterogeneous, and non-i.i.d. data regimes.

The expected outcome is a new theoretical and algorithmic foundation for stochastic optimisation in which computational effort is allocated adaptively according to the informational value of the data. This perspective provides a principled route to more data-efficient, robust, and computationally scalable machine learning systems operating in realistic heterogeneous environments.

The project will be based at the School of Computer Science at the University of Birmingham. The ideal candidate should have background in a numerate discipline, such as computer sience, mathematics, engineering, and must have excellent problem-solving and programming skills along with high integrity standards. Shortlisted candidates will be interviewed prior to a decision.

Funding notes:

Self-funded or externally sponsored students are invited to apply. Formal applications for research degree study should be made online through the University’s website: https://www.birmingham.ac.uk/schools/computer-science/postgraduate-research/applying-for-phd-in-computer-science.aspx

If you have interest related to this project but require funding, you should submit your original research proposal with your application. Scholarships are awarded on a very competitive basis.