AI PhD: Improving Artificial Intelligence with Insights from Neuroscience

AI PhD: Improving Artificial Intelligence with Insights from Neuroscience

Quick Decision

This computational PhD tests two neuroscience-inspired ideas in deep learning: whether astrocyte–neuron interactions can motivate constraints on network behaviour, and whether plasticity associated with spontaneous retinal waves can inform weight initialization. The comparison targets generalization, robustness to damage, explainability, learning speed and data efficiency. It is a good conceptual fit for quantitative applicants interested in neuroscience-to-ML translation, but the biological rules, model families, datasets, success measures and relationship between the two strands remain unspecified.

Research Problem and Doctoral Work

What the project is trying to test

The proposal starts from two biological observations. First, astrocytes respond to neural activity and feed back onto neuronal activity. The project will test whether an astrocyte-inspired rule can constrain deep networks, make them simpler and improve generalization to unseen data. It also proposes testing whether such networks are more resistant to “injury” and easier to explain than standard networks. Second, coordinated retinal waves occur before an animal first sees images and influence downstream synaptic plasticity. The project will apply plasticity rules associated with these waves to initialize deep networks before training, then test whether this speeds learning and reduces the amount of data needed. The stated longer-term aim is better, more explainable and less costly AI, with computational results also generating testable neuroscience hypotheses. The PDF does not assign wet-lab experiments to the student. Official project description.

The broad objectives leave key choices open. The project does not say which astrocyte signaling mechanism will become a computational constraint, what “simpler” means, or whether the rule will alter weights, architecture or the training objective. It does not define whether “injury” means removing units, connections or inputs, nor what explanation method will be scored. For the retinal-wave strand, the description says to use plasticity rules to initialize weights but does not explain the mapping from wave statistics and plasticity to network parameters, or how this differs from using waves as a pre-training signal. It also does not name datasets, architecture families, conventional baselines, compute budgets or a way to measure “less costly” AI.

What recent research contributes

Recent work makes both ideas plausible enough to investigate, but it does not establish the outcome in advance.

• In 2025, May, Dauphin and Gjorgjieva trained convolutional recurrent networks on simulated retinal waves and natural videos. Retinal-wave pre-training sped up next-frame prediction on motion-rich visual tasks, including when total training exposure was matched. This is a useful comparison for the project's second strand, but pre-training is not identical to deriving initial weights from wave-driven synaptic-plasticity rules, and the results are task- and architecture-specific. PLOS Computational Biology study.

• In 2026, Dong and He proposed slow astrocyte-inspired gating of plasticity in deep spiking networks for online continual learning. They reported competitive benchmark accuracy and less forgetting on neuromorphic and continual-learning tasks. This is direct evidence that astrocyte abstractions can shape learning in deep neural systems, but it concerns spiking networks and online plasticity, not the project's unspecified conventional deep-network regularization, injury resistance or explanation metrics. Frontiers in Neuroscience article record.

• A 2024 Communications Biology paper derived a model in which neuronal endocannabinoid signaling drives astrocytic D-serine regulation and changes synaptic plasticity. The model reproduced a specific reversal-learning deficit in mice and generated testable predictions. It provides a possible mechanistic basis for an astrocyte-inspired learning rule, but it is not a deep-network regularizer. Communications Biology study.

• A 2024 PLOS Computational Biology model used slow neuron–astrocyte feedback as meta-plasticity in context-dependent reinforcement learning. It offers another design contrast: astrocytes could dynamically alter when or how quickly a network learns, which differs from simply penalizing model complexity. The work is a reduced model for reinforcement learning, not a benchmark of conventional deep networks. PLOS Computational Biology study.

• An in-vitro study of neuron–astrocyte cultures reported lower synchronization and lower signal predictability under some astrocyte ratios. This does not invalidate the project hypothesis, but reinforces the need to measure task-specific benefits rather than assume astrocytes always simplify or improve a network. Neuron–astrocyte culture study.

Useful research design questions

These are analytical suggestions, not promised methods in the project description. For the astrocyte strand, a strong test would define the biological abstraction and its computational effect, compare against standard regularization methods, quantify simplicity separately from accuracy, and specify lesion severity and targets for robustness testing. For the retinal-wave strand, compare the proposed plasticity-rule initialization with common initializations and wave pre-training while matching data and compute; evaluate learning curves, sample efficiency and performance on held-out visual tasks. Multiple random seeds, transparent baselines and ablations would help distinguish a biological effect from extra training or model capacity. If “less costly” means energy or compute, measure it separately from data reduction.

The project has two large workstreams. Ask whether both are required, how they fit into one thesis, and which has priority if early results show that one method does not transfer across architectures.

School, Programme and Field Context

The degree-awarding institution is the University of Exeter. The project PDF names a potential PhD in Clinical and Biomedical Sciences, the Department of Clinical and Biomedical Sciences, and study locations at Hatherly Building and the Living Systems Institute on Streatham campus. The award page is for September 2027/2027–28 entry. The formal registration department and structured training plan are not detailed in the public project document. Official project description, department postgraduate research, EPSRC DLA programme page.

Exeter's overall institutional rankings are QS 136 in 2027 and THE equal 157 in 2027. In REF 2021, the broader Psychology, Psychiatry and Neuroscience unit reported 53% of activity at 4-star and 39% at 3-star; Mathematics reported 42%/52%, and Computer Science and Informatics 33%/50%. These are institutional or subject-unit indicators, not evaluations of this particular project, group or supervisor. Clinical and Biomedical Sciences describes an interdisciplinary environment spanning neuroscience and Data Science/AI, including computational modelling, statistics, high-performance computing and machine learning. Its neuroscience theme works across central nervous system function, neurodevelopment and neurodegeneration, using genomics, electrophysiology and neuroimaging. Exeter rankings, REF 2021 results, department research themes, Neuroscience and CNS Disease.

Departmental postgraduate information describes research-led supervision, access to facilities and interdisciplinary research, but not this project's coursework or individualized training schedule. Hatherly Laboratories support several neuroscience groups in neurophysiology and neurodevelopment, and list Joel Tabak among researchers. The project also names the Living Systems Institute as a study location. These pages establish surrounding research context, not guaranteed access to specific equipment or a dedicated GPU allocation. Department postgraduate research, Hatherly Laboratories, Living Systems Institute affiliate investigators.

Supervisor and Lab Assessment

The project names Dr Joel Tabak as lead supervisor and Marc Goodfellow as co-supervisor. Tabak's institutional profile describes combining experiments, computation and mathematics to understand rhythmic activity in biological systems. Goodfellow is listed at Exeter's Centre for Systems, Dynamics and Control for complex networks and intermittency. The project therefore sits in a broader computational-neuroscience and mathematical-dynamics context. The public profiles do not establish an astrocyte-specific AI or deep-learning lab; ask how the computational strands will be supported and divided. Joel Tabak-Sznajder profile, Centre for Systems, Dynamics and Control research, centre people.

Recent supervisor-related studies show complementary methods:

• Goodfellow co-authored a 2025 neural-mass modelling study fitting sleep EEG from children with epilepsy and controls. The model linked inferred excitatory synaptic currents to slow-wave differences. This supports experience in mechanistic modelling with human neuroscience data, but does not demonstrate the project's specific AI work. Epilepsia paper.

• Tabak co-authored a 2025 study combining tadpole-behaviour observations with biomechanical and computational modelling of escape movements. This illustrates an experimental-computational approach, but studies a different biological system and does not demonstrate prior work on astrocytes or deep-network regularization. iScience paper.

The department's stated Data Science/AI and Neuroscience themes, Hatherly's neurophysiology/neurodevelopment groups, the Living Systems Institute and the mathematics/dynamics centre provide cross-disciplinary context. Confirm which resources, programming support, data, compute and supervisor time are committed specifically to this PhD. Department research themes, Hatherly Laboratories.

Funding, Application, and Living Cost

Funding and international eligibility

Award reference 6010 describes an index-linked stipend for up to 3.5 years full-time (part-time pro rata), currently £21,805 per year for 2026/27, payment of Home tuition, and £5,000 total Research Training Support Grant over 3.5 years. The grant is a research-support budget, not personal maintenance. The university says it covers the difference between Home and International tuition for successful international candidates. Central guidance says international full studentships are limited to up to 30% of the cohort and flags complex residency requirements. International applicants must also cover their student visa, healthcare surcharge and other costs of moving to the UK themselves; these costs are outside the described studentship. Check individual residency status early and budget separately for these costs. Award page, central EPSRC DLA eligibility information, UKRI training grant guidance.

The award's detailed funding section calls £21,805 the current 2026/27 rate, matching the UKRI minimum-stipend table. Its summary row labels the same amount as the 2027/28 rate. Treat £21,805 as a published current rate, not a confirmed September 2027 amount, until the university confirms the indexed rate in writing.

Dates and application materials

The application deadline is midday on Tuesday 8 December 2026. The project-specific award page says referees must email references from institutional accounts by 8 January 2027. The central EPSRC DLA page says 15 February 2027. The project-specific page indicates general interviews are likely between 15 and 24 February 2027, while central text still gives interview and outcome dates in 2026. These discrepancies appear to be inconsistent page updates; ask PGR Admissions to confirm, and plan for the earliest reference date. Project-specific award dates, central EPSRC DLA dates.

Applicants must use the application route linked from the project description. The award page asks for a personal statement, CV, transcript(s), two references and English-language evidence where applicable. Documents must be in English; certified translations of academic qualifications are required where relevant. The award allows applications to up to three projects, with a separate application for each and a preferred project named if applying to more than one. The project-specific PDF gives Dr Joel Tabak (j.tabak@exeter.ac.uk) for questions; the award page lists PGR Admissions for process questions. Official project PDF, award page and requirements.

Local cost comparison

Exeter estimates that a single student in Exeter or Cornwall will need about £1,300–£2,000 per month for basic living costs, based on local average spending in April 2026. The current £21,805 stipend works out to about £1,817 per month when divided by twelve. At the lower end that leaves roughly £517 before other costs; at the upper end it is about £183 below the estimate. This is a rough comparison: housing and lifestyle vary, future annual rates are not confirmed here, and one-off or family expenses are not included. Exeter budgeting guidance, UKRI stipend rates.

Applicant Fit and Preparation

Required background and skills

The project PDF specifies a degree in Mathematics, Physics, Natural Sciences, Computer Science or Neuroscience. The general award criteria ask for a First or Upper Second UK Honours degree or equivalent in an appropriate science or technology subject. Applicants with a Lower Second may be considered if they also hold a Master's degree or have significant relevant non-academic experience. English-language proof is required when applicable. The PDF does not state a mandatory programming language or prior machine-learning qualification. Project-specific entry subjects, award entry requirements.

Suggested preparation

A strong application can make the computational bridge legible: describe prior quantitative work, reproducible model implementation and how results were compared against baselines. Useful preparation may include Python or another scientific language, deep-network training, initialization and regularization, retinal-wave development, astrocyte signaling and model evaluation. These are suggestions based on the research tasks, not published admission prerequisites. A compact reproducible project that reports limitations as well as results may be more informative than a list of tools alone.

The likely skills transfer to machine-learning, data-science and computational-neuroscience roles, but this is not a placement or employment promise. The UK government's Digital and Technologies plan projects broad sector demand, while noting uncertainty and that its scenarios do not model AI's labour effects. A DSIT-commissioned 2025 survey reported technical-skills gaps among 119 self-selected organisations, and the National Careers Service lists data scientist, machine-learning engineer and AI data scientist among related role titles. These sources signal skills relevance, not the probability of a particular graduate's outcome. Digital and Technologies Jobs Plan, UK AI Labour Market Survey 2025, National Careers Service data-scientist profile.

Next Actions and Questions to Verify

Before applying

1. Email the project contact and PGR Admissions to reconcile the reference deadline (8 January versus 15 February 2027), confirm interview dates, and ask whether £21,805 is only the current 2026/27 rate or also the intended 2027/28 rate.
2. If applying as an international student, check residence rules, the cohort cap and written fee-coverage terms against the applicant's own circumstances.
3. Submit by noon on 8 December 2026 through the route in the official project PDF, with the personal statement, CV, transcript(s), two references and English evidence where required.

At interview

• Are the astrocyte and retinal-wave strands separate or integrated, and which is the central thesis contribution?
• What molecular mechanism informs the astrocyte-inspired rule, and how will network simplicity, generalization, injury resistance and explanation be measured?
• How will spontaneous retinal waves and their plasticity rules generate initial weights, and how will that differ from wave pre-training?
• Which architectures, datasets, conventional baselines and equal-compute/equal-data comparisons are planned?
• What data, code, computing resources and training will be available at the two named study locations?

Before accepting an offer

Confirm the registered PhD programme and department, supervisor responsibilities, 3.5-year funded duration, annual stipend and indexation, Home tuition and international fee coverage if relevant, and the scope of the £5,000 total RTSG. The scheme requires international applicants to cover visa, healthcare surcharge and UK relocation costs themselves. Ask only whether any separate discretionary support exists beyond the published award.

Evidence and Sources

Project and programme

• Official project description
• EPSRC DLA award, reference 6010
• Central EPSRC DLA information and eligibility
• EPSRC Doctoral Landscape Award programme

University, department and research environment

• Exeter league tables and rankings
• REF 2021 results
• Clinical and Biomedical Sciences research themes
• Neuroscience and CNS Disease
• Clinical and Biomedical Sciences postgraduate research
• Hatherly Laboratories
• Living Systems Institute affiliate investigators
• Living Systems Institute affiliate investigators
• Centre for Systems, Dynamics and Control research
• Centre for Systems, Dynamics and Control people

Publications

• Retinal-wave pre-training in artificial neural networks
• Astrocyte-gated plasticity in deep spiking networks
• Astrocytes and reversal learning
• Astrocyte-mediated context-dependent network dynamics
• Neural-mass modelling of sleep EEG in children with epilepsy
• Tabak co-authored study of tadpole escape behaviour and modelling
• Neuron–astrocyte culture activity and predictability

Funding, costs and employment context

• UKRI stipend and fee guidance
• Exeter living-cost estimate
• Digital and Technologies Jobs Plan
• UK AI Labour Market Survey 2025
• National Careers Service: Data Scientist