The rapid growth of deep learning has come at an extraordinary environmental and computational cost, yet the standard training paradigm remains remarkably unchanged. Every sample is passed through the network, and every synapse is updated, on every epoch. Recent work has begun to challenge both halves of this independently. Progressive Data Dropout has shown that progressively reducing the training set across epochs can cut effective training epochs by up to 90% while actually improving accuracy. Separately, research grounded in neuroscience has demonstrated that restricting synaptic updates to only the most informative weights, motivated by the metabolic cost of learning in biological systems, can reduce the energetic burden of training by orders of magnitude. What has not yet been explored is what happens when these two ideas are unified under a single energy-aware framework. This PhD will develop principled methods for jointly optimising which data and which weights are updated during training, using metabolic energy as the governing design constraint. The student will build on an established energy model for synaptic plasticity to derive theoretically grounded training schedules and will evaluate these across a range of architectures (CNNs, Vision Transformers, and language models) on standard benchmarks through to large-scale settings.
The project sits at the intersection of machine learning, computational neuroscience, and cognitive science, and the student will work closely with both supervisors to move between these perspectives. We are looking for a candidate with a strong foundation in either machine learning or mathematical/computational neuroscience, demonstrable programming experience (Python/PyTorch), and the curiosity to work across disciplinary boundaries. A background in optimisation theory or an interest in the energy and sustainability implications of AI would be particularly welcome.
Supervisors: Prof. Mark van Rossum (School of Psychology), Dr. Shreyank Narayana Gowda: shreyank.narayanagowda@nottingham.ac.uk (School of Computer Science)
For more information and to apply, click the apply button above.
Annual tax-free stipend based on the UKRI rate (£21,805 for 2026/27), Home tuition fee, and £3000 p.a. Research Training Support Grant.
#J-18808-Ljbffr