University of Bath PhD in Knowledge Representation and Localised Model Editing in Bio-Inspired Neural Networks

University of Bath PhD in Knowledge Representation and Localised Model Editing in Bio-Inspired Neural Networks

Project Assessment

Where does a spiking neural network represent a particular piece of information, and can that information be changed without retraining the whole model? This Bath Electronic and Electrical Engineering project makes representation and localised editing the research problem. Replacing a face profile is an illustrative example, not an established dataset, application deployment or access to personal records. Dr Dorian Florescu's work offers relevant mathematical and signal-processing foundations; their transfer to semantic knowledge editing remains to be demonstrated.

The opportunity is strong but its fit needs checking for research combining quantitative modelling, machine learning and careful evaluation. It competes with other Faculty of Engineering and Design projects for URSA support open to Home and Overseas students, with an unspecified international cap. The saved project information gives 29 November 2026 as its deadline; the saved official Engineering URSA information specifies 23:59 GMT. Current acceptance has not been rechecked. The expected start is 4 October 2027, with another date possible by agreement. The main immediate constraint is English: non-UK applicants not currently studying in the UK must meet the language requirement before the application deadline, without exceptions. Individual equivalence and award terms still require decisions. Current URSA competition

A topic-specific university page remains unconfirmed in the available evidence. The official programme, supervisor and URSA sources below verify their own roles and policies, rather than independently proving every task in the project description. A useful first contact should establish the initial SNN model, what counts as knowledge and a successful edit, and the available resources.

Research Problem and Doctoral Work

Representation, Editing and Evidence

The advertised motivation is that rapidly expanding AI continues to demand more data and power despite efficiency improvements. It proposes bio-inspired models, particularly spiking neural networks, to investigate knowledge representation and localised model editing. The notice also frames characterising knowledge representation in SNNs as a way to develop better hypotheses about how the brain represents knowledge. Sparse event-based activity provides a reason to study efficiency; it is not proof that a particular training-and-editing pipeline already saves energy.

The intellectual challenge is to identify the part of a model responsible for information of interest, change that information and explain the consequences. A changed output alone does not show that a stored concept was located. Classification readout selection, reconstruction of an encoded signal and editing a semantic fact are different operations. Likewise, findings about large-language-model editing cannot automatically establish the same capability in a spiking network.

An informative evaluation would examine the requested change, retention of unrelated behaviour, locality, generalisation and the effect of repeated edits. These are analytical suggestions, not a prescribed project protocol. A correlation between a neuron and a class is weaker evidence than showing how a controlled change affects the relevant representation. The exact model, learning rule, task, dataset and success criteria remain unspecified, so they should be agreed before treating the advertisement as a ready-made technical plan.

Ilievski and colleagues' 2025 Nature Machine Intelligence Perspective, “Aligning generalization between humans and machines”, provides conceptual context. It distinguishes notions and evaluation of generalisation across AI and cognitive science. It is an external Perspective, not a new Florescu-team experiment or proof that editing improves human-like abstraction. The advertisement also lists Li et al.'s 2024 “Brain-inspired computing: A systematic survey and future trends” (Proceedings of the IEEE) and Fu et al.'s 2025 “Constructing artificial neurons with functional parameters comprehensively matching biological values” (Nature Communications 16(1), 8599). Their treatment here is limited to the advertised bibliographic information; no findings are attributed to them, and they are not counted among the three recent original studies. Generalisation Perspective

Likely Work and Research Value

The advertisement names SNNs and knowledge representation/localised editing, rather than a fixed hardware implementation or compulsory experiment. Possible doctoral outputs include a well-defined representation question, a reproducible editing comparison and an explanation of which behaviours are preserved or changed. They are plausible outputs inferred from the question, not promised software, publications or performance targets. The advertisement hopes for major AI and computational-neuroscience publications; acceptance cannot be guaranteed.

An efficiency claim needs a defined comparator and accounting boundary. Inference sparsity does not account automatically for training, data preparation, repeated editing, memory use or hardware overhead. The required CPU/GPU allocation, spiking simulator, access to neuromorphic equipment and measurement arrangements were not established. Software-only work and hardware evaluation imply different schedules, dependencies and evidence, so the expected balance is a high-value early question.

This direction can develop skills in scientific programming, mathematical reasoning, experimental design and interpretable ML. Bath's wider AI community covers topics such as inverse problems, signal processing and interpretation, offering relevant intellectual context. It does not establish project membership, a funded industrial placement or a reserved computing resource. Current project-specific jobs, graduate salaries and measured carbon benefits were not verified. The field's appeal rests on the unresolved representation question and the quality of evaluation, rather than an assumed AI-employment guarantee. Bath AI research community

School, Programme and Field Context

The University of Bath awards the PhD, through Electronic and Electrical Engineering in the Faculty of Engineering and Design at Claverton Down. This is not a Computer Science or Chemical Engineering registration. The programme says most students complete in four years, with a two-year minimum and normally four-year maximum. Registration, induction, probation/confirmation, thesis and viva form the academic route; an advertised studentship does not create an employment contract. EEE doctoral programme

Florescu's current institutional profile lists the Centre for Bioengineering and Biomedical Technologies (CBio) and the Bath Institute for the Augmented Human. These are verified affiliations. The departmental environment includes interdisciplinary computational, sensing and engineering research; its other health, energy or materials examples should not become work packages in this doctorate. A visiting affiliation elsewhere does not imply a joint degree or second funding route. Current supervisor profile, Departmental research environment

Bath is 125th in the QS World University Rankings 2027; the 2026 figure was 132. These are overall institutional positions, not an SNN or computational-neuroscience ranking. Bath's institutional PGR account also reports 92% of research as world-leading or internationally excellent in REF 2021, an institution-wide measure, not an EEE-only result. A current relevant subject ranking, group completion statistics and project graduate outcomes were not established. QS institutional ranking, Bath doctoral context

MyDoctoralDevelopment provides workshops, research and transferable-skills development, coaching and communication support. These are institutional opportunities, not a confirmed personal curriculum or placement. Ask how mathematical training, experimental evaluation, scientific software and any hardware induction will be combined for this particular thesis, who supervises them and how probation/confirmation will assess progress. The broad AI community is useful context, but formal project roles and access should be confirmed separately. Doctoral professional development

Supervisor and Lab Assessment

Current Appointment and Collaboration

Dr Dorian Florescu is a Lecturer in Electronic and Electrical Engineering at Bath and the named project contact. His current profile and recent papers connect mathematical signal recovery, event-based encoding and constrained interpretable models. They provide a coherent foundation for asking how information is represented, although they do not themselves establish successful semantic editing in an SNN. Publication counts or a former appointment are not substitutes for this assessment.

Matthew England and Daniel Coca appear as collaborators on relevant work; neither is automatically a formal co-supervisor for this advertised doctorate. Likewise, funding acknowledged in an earlier paper is not the applicant's 2027 award. Daily meeting arrangements, the co-supervision plan, student outcomes and software/data publication expectations remain unknown. The institutional contact is dmf36@bath.ac.uk; confirm capacity and the first bounded research question before preparing a broad multi-direction proposal.

Three Recent Original Contributions

Florescu, 2025, “A Generalized Approach for Recovering Time Encoded Signals with Finite Rate of Innovation”, IEEE Transactions on Signal Processing 73, 1862–1876; DOI 10.1109/TSP.2025.3567443. Its complete institutional abstract reports broader reconstruction conditions for time-encoding-machine measurements, based on local derivatives, with noisy-signal guarantees, simulations and a TEM hardware test. The evidence covers that original abstract and an indexed accepted-manuscript abstract, not substantive full methods or results. This supports mathematical/time-encoding expertise; it does not locate semantic knowledge or demonstrate SNN editing. The paper's “Sensing Beyond Barriers” funding is historical research support, not this studentship. Original time-encoding study

Florescu and Bhandari, 2025, “Multi-dimensional unlimited sampling and robust reconstruction”, Applied and Computational Harmonic Analysis 79, 101796; DOI 10.1016/j.acha.2025.101796. The study uses multidimensional structure and modulo-hysteresis encoding for reconstruction, with noiseless guarantees, Gaussian-noise bounds and numerical tests. The evidence covers the complete institutional abstract and substantive publisher-indexed introduction, model definitions, selected limitations in Theorem 3 and conclusion; it does not cover the full 27-page paper or a reproduction. Original multidimensional study

The qualifications are decisive: arbitrarily large fixed Gaussian noise at sufficiently high sampling rates does not imply robustness to large outliers. Recovering a noisy input is distinct from denoising it. The theorem's probability bound is not a universal practical accuracy estimate, and proposed multidimensional hardware is future work, not an implemented prototype. This is evidence of precise assumptions and theoretical reconstruction work, not semantic memory editing or cost-free robustness.

Florescu and England, 2024, “Constrained Neural Networks for Interpretable Heuristic Creation to Optimise Computer Algebra Systems”, ICMS 2024, LNCS 14749, 186–195; DOI 10.1007/978-3-031-64529-7_19. This is an original published conference paper, not merely a meeting abstract. The evidence covers PDF sections 3–4 (methods), section 5 (results) and section 6 (final thoughts), not a complete reproduction. It expresses a human variable-ordering heuristic for cylindrical algebraic decomposition as a tightly constrained neural network, then investigates alternative features and weights. Original constrained-network paper

Training uses random three-variable polynomial problems; NLSAT evaluation is consulted at each training step during weight selection. That makes it unsuitable to describe the evaluation data as a completely untouched final test for a future adaptation. A fresh independent evaluation would be needed. The work makes the heuristic human-interpretable, but explicitly leaves new mathematical understanding unproven. It illustrates constrained model modification rather than a demonstrated SNN knowledge-editing method.

An additional Florescu–Coca author preprint, first posted in 2018 and revised in 2019, discusses precise spike timing and readout-neuron selection in liquid-state machines. The evidence is limited to its abstract and version history. It is useful SNN background, not the third recent original, a fully read journal paper or proof that all stored knowledge has been causally localised. Across the portfolio, the practical gap between mathematical representation and semantic editing remains a legitimate thesis question. Earlier spike-timing foundation

Funding, Application and Living Cost

Current URSA Terms and Their Limits

The saved URSA Engineering and Design information lists the October 2027 competition, closing 29 November 2026 at 23:59 GMT; current acceptance has not been rechecked. It supports Home or Overseas tuition, maintenance and up to £1,000 annual training support. The project's dated funding description gives £21,805 at the 2026/27 rate and states an international cap without a percentage. Neither that older reference nor the UKRI 2026/27 minimum establishes the personal 2027/28 stipend. Training support is research/training money, not extra living income. URSA current round, Dated UKRI reference

For Engineering and Design, the scheme gives up to 3.5 years as a general maximum. It is not a personal promise of 42 funded months, and the programme's normal four-year completion is not four years of maintenance. The award letter should establish the actual term, fee coverage, indexation, continuation conditions and costs if registration outlasts support. Visa and Immigration Health Surcharge costs are personal responsibilities; study is in person at Bath. No allocation probability or fixed overseas quota was verified.

Application and English Deadline

The expected start is 4 October 2027, with another date potentially agreed by the intended supervisor and Doctoral College. Apply to the Electronic and Electrical Engineering PhD, naming Florescu. At most two projects may be selected: choose “University of Bath URSA” under Funding Your Studies, enter the first project title and intended supervisor, and use Project 2 Title and the second-choice supervisor field if applicable. Incorrect entries can prevent automated consideration. The programme's general international three-month-before-start deadline does not override this earlier funding deadline.

The programme requires two references, at least one academic from the most recent place of study. A personal statement should explain research motivation, examples of undergraduate/Master's research and career plans. Its 1,000-word proposal instruction applies to applicants bringing their own idea, not automatically to an advertised predefined topic. Confirm the required documents and any additional statement with the supervisor; the programme requires a named supervisor and recommends prior contact. Programme application requirements

For a non-UK applicant not currently studying in the UK, English must be satisfied before the application deadline, with no exceptions. The baseline is IELTS 6.5 overall, each component 6.0. TOEFL iBT/Home Edition before 21 January 2026 requires 90 overall, each 21; on or after that date it requires 4.5 overall, each 4.0. Tests must normally be within 24 months of course start. These are not grounds to assume a later pre-sessional course will rescue an unmet competition deadline. Doctoral English rules

A qualifying English-taught degree can provide an exemption route, including certain degrees outside majority-English-speaking countries, subject to the sole-English-medium, recency and applicable residence conditions and university assessment. A Chinese medium-of-instruction letter is not automatic approval. Obtain an explicit decision before the hard deadline rather than treating a possible exemption as already granted.

The project announcement does not specify when non-English certificates or transcripts must be translated, or whether a particular certification is required; confirm this with admissions. The Faculty encourages applications from under-represented groups. Applicants may explain circumstances that affected educational attainment in a short paragraph at the end of the personal statement. If an applicant declares a disability on the application, the Disability Service will contact them to discuss support needs.

Costs Beyond the Headline Stipend

Published 2027/28 overseas Engineering research tuition is £31,350 annually, while Home fees await UKRI confirmation; continuing fees may rise by up to 8% annually. This is the underlying exposure if an award is incomplete, not an additional bill where URSA covers the relevant tuition. Confirm coverage of increases, any training/bench fee and the period after maintenance ends. 2027/28 tuition

The official 2025/26 budget is historical: £405 weekly or £21,060 yearly, including average rent of £230 weekly on a 51-week contract. Its published total includes an off-campus utilities line even under a halls-labelled scenario. Removing that £37 line leaves £368 weekly before one-offs if the actual contract already includes those bills; this is an analytical no-double-counting scenario, not a corrected university price. Off-campus £750 monthly excludes bills. The document is not a 2027 quotation, housing guarantee or basis for assuming a comfortable surplus. Budget the remaining week, deposits and relocation. Historical postgraduate budget, Dated doctoral FAQ

UK visa references read in October 2026 give £558 for a Student visa and £776 per year for student IHS. Actual visa validity and part-year rules determine the upfront charge, and 2027 prices may change. The EEE programme is ATAS-covered; China is not nationality-exempt. The subject classification and individual clearance must be assessed, rather than interpreted as a blanket application ban or assured approval. Student visa, IHS, ATAS rules

Applicant Fit and Preparation

The project requires an achieved or expected first-class or 2:1 honours degree, or international equivalent, in a directly relevant discipline; a Master's is advantageous. The general EEE programme also describes a Merit-Master's alternative, but individual project requirements apply. That broader alternative cannot be treated as automatic permission to bypass this project's stated degree condition. Mathematical reasoning, coding and a clear understanding of representation/evaluation would be useful evidence, not newly imposed requirements to know every SNN method.

There is a positive Chinese-applicant pathway: the programme welcomes worldwide equivalents and URSA supports Overseas fees. Bath's indexed country research guidance mentions a first-class Bachelor's with research experience or a Master's, while the current directly displayed China page does not repeat that PGR subsection. Its taught-Master's 62–75% thresholds cannot be converted into a definitive doctoral cutoff. Send the actual degree, transcript and research experience for a current EEE decision; neither an unspecified institution tier nor a Master's automatically settles equivalence. International qualification guidance

Suggested preparation is a small, reproducible example showing how a changed representation is assessed against retained behaviour and independent evaluation. Distinguish simulation accuracy, robustness and energy measurement. Applicants expecting a turnkey LLM-editing implementation, guaranteed neuromorphic hardware or a purely biological wet-lab doctorate should clarify scope early; the advertisement centres on computational models and representation.

Next Actions and Questions to Verify

Before applying, ask Florescu which SNN, task and definition of knowledge will anchor the thesis, whether hardware evaluation is planned, and what preparation is expected. Obtain the Chinese qualification and English/exemption decision before the funding deadline. Confirm the project-selection fields, complete documents and personal award expectations without substituting a general programme deadline.

At interview, ask how requested edits, unrelated retention, locality, sequential changes and generalisation will be judged. Clarify a genuinely untouched final evaluation set, the compute/data budget and full energy-accounting boundary. Discuss daily supervision, co-supervisors, technical support, software release and access. The papers' assumptions and existing evaluation-selection procedures should inform questions, not become unqualified methodological guarantees.

Before accepting, obtain the actual 2027 stipend, indexation, funded months, Overseas tuition and increase coverage, continuation rules and any completion-period costs in writing. Confirm training support use, housing, relocation, ATAS/CAS and visa timing. The institutional award framework is credible; the personal financial package and first research scope still need to be made concrete.

Recruitment, policy and cost information reflects October 2026 sources; later availability should be confirmed with admissions.

Evidence and Sources

Supervisor, Research and Field Context

• Dorian Florescu current Bath institutional profile
• Bath Electronic and Electrical Engineering research environment
• Florescu 2025: recovering time-encoded finite-innovation signals
• Florescu Bhandari 2025: multi-dimensional unlimited sampling
• Florescu England 2024: constrained neural-network heuristics
• Florescu Coca 2018/2019 spike-timing foundation author preprint
• Ilievski et al. 2025: human and machine generalisation perspective
• Bath Centre for Artificial Intelligence interdisciplinary environment

Programme and Training

• Bath Electronic and Electrical Engineering PhD programme
• Bath doctoral professional development
• QS Bath overall 2027 table
• Bath postgraduate research school context

Funding, Qualifications and Costs

• Current University Research Studentship Award policy and 2027 Engineering deadline
• Bath doctoral English rules
• Bath Chinese research-degree qualification route
• Bath Engineering research tuition 2027/28
• Bath postgraduate budget 2025/26
• Bath doctoral admissions FAQ 2026/27
• UKRI training grant current stipend
• UK Student visa official route
• UK immigration health surcharge
• ATAS current nationalities and subject rules