Efficient and Adaptive AI PhD | Learning Under Resource Constraints

Efficient and Adaptive AI PhD | Learning Under Resource Constraints

Research cut-off: 6 November 2026

Quick Decision

This is a four-year UCL EPSRC Doctoral Landscape Award (UELA) collaborative PhD with MediaTek Research UK Ltd. It targets a substantive research problem: how a learning system should choose data, computation and update decisions when compute, memory, energy and data access are limited. The advertised work combines mathematical modelling, theory and computational experiments, with possible directions in efficient generative modelling, reinforcement learning and adaptive data acquisition. It is most closely aligned with applicants who can demonstrate strong mathematics or computer science, solid programming and interest in foundational machine learning. The project catalogue lists an additional £2,500-per-year partner stipend enhancement and at least three months of industrial placement.

The main decision condition is funding clarity. UCL’s UELA page gives £26,022 as the 2027/28 estimate in its summary, but a detailed funding line labels the same figure 2026/27; a separate UCL studentship-rates page lists £24,522 as the projected 2027/28 baseline. The partner enhancement is separately listed at £2,500/year. Ask UCL for the year-by-year award schedule and confirmation of the fee award before treating any total as settled. The application deadline is 8 January 2027 at 12:00 noon UK time; UCL recommends submitting the application and ensuring both references arrive by 16 December 2026. A useful first step is to compare the research question with prior work in sequential decision-making, optimization and machine learning, then ask the supervisor how the four-year registration and MediaTek placement will be organized.

Research Problem and Doctoral Work

What is unsolved and why it matters

AI systems increasingly run on smartphones, connected devices and communication systems. In these settings, the system cannot assume unlimited computation, memory, energy or access to new data. A model may need to decide not only what to predict or do, but also whether another observation or computation is worth its cost and whether the model or decision should be updated. The UELA catalogue frames this as a resource-allocation problem and aims for algorithms with mathematical foundations that expose trade-offs between resource use and learning performance. The application areas named are on-device AI, wireless communication and chip design; they are possible contexts, not promised product deliverables. The official project entry

Work, methods and plausible research angles

The advertised research sequence is broad rather than a fixed work-package list. The researcher would help decide which data to acquire, which computations to perform and when to update a model or change a decision. The catalogue names efficient generative modelling, reinforcement learning and adaptive data acquisition as potential methods; it specifies mathematical modelling, theoretical analysis and computational experiments, and says the precise direction will be shaped jointly around the student’s interests and strengths within a focused four-year PhD.

Recent research by the supervisor and close coauthors provides methodological starting points. In Delayed Feedback in Kernel Bandits, the authors study decision-making when observations arrive late and establish regret guarantees supported by simulations; that is relevant to delayed or costly feedback, but it does not measure device memory or energy. The 2023 paper and the 2024 reward-free kernel RL paper examine adaptive exploration and planning, including adaptive domain partitioning. A 2025 follow-up studies sample complexity under broader data-access assumptions. A 2024 Bayesian-optimization study shows how random exploration can avoid an expensive non-convex acquisition-optimization step under its assumptions. These works connect to efficient learning and decision allocation, but they are largely theory and algorithm studies, not evaluations of a particular mobile chip or wireless device.

A related NeurIPS 2023 paper analyzes regret for kernelized reinforcement learning in large state-action settings. Separately, a 2023 IEEE paper on diffusion models for radio-channel modelling reports work on simulated macro-cell channels and fine-tuning to a smaller out-of-distribution micro-cell dataset. It shows a nearby generative/wireless research direction, not a commitment that this PhD will use diffusion models or produce a communication product.

A useful thesis design would need to translate the broad resource problem into measurable constraints: for example, a stated compute or memory budget, energy or latency measure, data-acquisition cost, and baseline methods. The project description does not specify the device class, benchmark, dataset, model family or hardware access. Those are important scoping questions, not assumptions that can be filled from the papers.

School, Programme and Field Context

The degree host is University College London, specifically the Department of Electronic and Electrical Engineering (EEE), within the Faculty of Engineering Sciences; MediaTek Research UK Ltd is the industry partner, not a separately named degree-awarding institution. UCL’s course page confirms a UCL PhD route in London and describes EEE research areas including machine learning, communications, control and networks. The department reports more than 100 PhD students and regular primary-supervisor meetings with subsidiary-supervisor contact. UCL also identifies wider Statistical Machine Learning and reinforcement-learning groups in its AI research ecosystem, but does not state that this project will be placed in either group. EEE MPhil/PhD course information, department overview, and UCL AI research groups.

UCL reports that EEE ranked #2 in the UK for Electronic and Electrical Engineering in The Times Good University Guide 2026. This is a subject-table signal, not an evaluation of this specific project, supervisor or postgraduate research experience. UCL’s subject-area summary.

UELA says this studentship is a four-year registration and funded period, with the thesis expected within those four years. The general EEE course page currently describes a normal three-year full-time route and says students initially register for an MPhil, then transfer after a year subject to an upgrade examination. The programme page may be describing the standard course rather than the UELA award; the project’s exact registration calendar and upgrade milestones need confirmation. UCL lists the 2027/28 generic full-time international EEE fee at £36,100 per year; the Home fee is marked “to be confirmed.” UELA’s general funding page says four years of fees at the Home rate and that eligible International awardees receive a UCL award to cover the Home/International difference. Ask for the project-specific award letter rather than assuming the generic course fee is payable personally.

Supervisor and Lab Assessment

Supervisor research directions

The official project catalogue names Sattar Vakili as principal supervisor. His own CV lists him as an Associate Professor at UCL from 2026 and records prior research leadership at MediaTek Research from 2020 to 2026, including generative models, reinforcement learning, optimization, communication systems and chip design. The CV is a supervisor-maintained source; the project catalogue independently confirms his role on this opportunity. Supervisor CV and publication list.

The paper set is coherent with the advertised interest in efficient learning and sequential resource allocation: kernel bandits, delayed observations, reward-free reinforcement learning, sample complexity, computationally cheaper exploration and wireless generative modelling. Its strongest evidence is for the supervisor’s methodological research experience. It does not verify a ready-made hardware platform, an industrial dataset, expected publication count, or guaranteed application-specific results.

Collaboration, resources and unknowns

MediaTek Research UK Ltd is named as partner, with an industrial co-supervisor and at least three months of industrial placement. The individual co-supervisor, placement location, timing, duties, access to devices or proprietary data, and publication/IP arrangements are not specified in the project entry. UCL’s departmental pages describe broad research activity and industrial links, but should not be read as confirmation that this project has a specific chip, testbed, compute allocation or dataset. Clarify what the partner contributes in addition to the listed stipend enhancement and placement.

Funding, Application and Living Cost

Item	Verified information	Decision implication
Scheme and duration	UELA 2027/28; expected start 1 October 2027; four years full-time funding and registration.	Thesis is expected within the funded four-year period.
Base funding	UELA lists four years of fees at Home rate, an enhanced maintenance stipend and a Research Training Support Grant (RTSG).	RTSG supports eligible training/project costs; it is not extra living income. The project’s exact RTSG allocation is not shown.
Partner enhancement	Project 2531cc1057 lists an additional £2,500 stipend enhancement per year from MediaTek Research UK Ltd.	Confirm payment timing and terms in the written offer.
Stipend-year conflict	UELA’s key summary calls £26,022 the 2027/28 estimate; its detailed funding paragraph labels £26,022 as 2026/27. UCL’s separate rate page projects a £24,522 UKRI baseline for 2027/28.	Do not calculate a definitive total or monthly budget until UCL confirms the applicable UELA rate and partner top-up for each year.
International fees	The project is collaborative and the catalogue accepts Home and International fee candidates. UELA caps scheme-wide International offers at 15 and says eligible International awardees receive a UCL award covering the Home/International fee difference.	Competition is especially constrained for International places. Fee status is self-assessed initially and formally assessed by UCL Admissions.
Generic EEE tuition	UCL’s 2027/28 course page lists £36,100/year International tuition; Home tuition is not yet confirmed.	This is the generic course fee. Ask for the exact project award and fee-difference coverage in writing.
Deadline and references	Full application due 8 January 2027 at 12:00 noon UK time. UCL recommends that applications and references be ready by 16 December 2026; two references must be submitted by the deadline.	Late or incomplete applications with fewer than two references are ineligible for review.
Application limit and form	Up to three projects may be selected, with a separate application per project. Q1–Q3 are up to 300 words each; Q4a/Q4b up to 500 words each; optional Q5 up to 200 words. Initial review is anonymized.	Avoid personal details in answers; tailor evidence to this project and keep each response within its limit. A successful UELA selection is followed by a formal UCL admissions application and official transcripts.

UCL’s indicative postgraduate budget lists halls at £153–529 per week, or £7,956–27,508 on a 52-week basis; postgraduate contracts are usually 51 weeks. Shared private housing is listed at £150–300 per week, excluding household bills. UCL’s £426/month estimate covers selected essentials such as food, course materials, mobile, wellbeing and a Zones 1–2 student Oyster travelcard, but excludes rent and discretionary costs. These are planning ranges, not personal forecasts. UCL living-cost guide.

For International applicants, the current GOV.UK Student visa financial requirement for London is £1,529 per month for up to nine months, subject to exemptions and acceptable financial-sponsorship evidence. This is a visa documentation rule, not the expected actual monthly spend, and should be checked again before a 2027 application. The UELA page also notes that some candidates may need ATAS, English-language evidence, a Student visa and the Immigration Health Surcharge. Current visa funds guidance and UCL’s successful International candidate guidance.

Applicant Fit and Preparation

Required background and skills

The project welcomes applicants with a strong background in mathematics, computer science, engineering or a related discipline, good programming skills and an interest in foundational machine learning. UCL’s generic EEE course entry requirement is a UK Master’s degree, or at least an upper-second-class UK bachelor’s degree, or an equivalent overseas qualification in a relevant subject. UCL also lists English Language Level 1; depending on nationality and research area, an International candidate may need ATAS clearance. The project-specific catalogue does not name a programming language, framework, prior publication or previous edge-AI project as a formal requirement.

Preferred experience and suggested preparation

Probability, optimization and algorithmic analysis are explicitly described as valuable. For an application, useful evidence would include a research or dissertation example that shows mathematical reasoning, well-structured code, experimental design and clear reporting of limitations. This is preparation advice derived from the advertised work, not an additional admissions rule. For the interview, candidates can prepare a short discussion of one resource constraint, how it could be measured, what trade-off matters and how a theoretical method might be tested computationally.

The UK market evidence is broad rather than project-specific. DSIT estimated 130,403 AI-related roles in 2025, up 51% on 2024, alongside £33.6bn in related revenues; the study combined secondary data, a 376-company survey and 50 interviews. A separate 2025 employer survey found that 35% of respondents with AI vacancies had at least one hard-to-fill vacancy, with work experience and technical knowledge among cited challenges. The survey had varying respondent counts and a self-selection limitation. These sources suggest an active but changing market for technical AI skills; they do not forecast an individual PhD graduate’s role, pay or placement outcome. Plausible directions include AI/ML research, research engineering, applied machine learning, embedded/edge systems and wireless communications R&D, depending on the thesis and candidate’s subsequent experience. DSIT sector study, AI Labour Market Survey 2025, and DSIT AI skills summary.

Next Actions and Questions to Verify

Before applying

1. Self-assess Home/International fee status and confirm the official classification process; only collaborative projects can accept International candidates, and UELA makes at most 15 International offers scheme-wide.
2. Ask UELA/EEE for the year-by-year stipend schedule, confirm which year £26,022 applies to, and obtain written terms for the £2,500/year MediaTek enhancement, the Home/International fee-difference award and RTSG.
3. Request clarification of UELA’s four-year registration versus the generic EEE course page’s three-year route, including the MPhil-to-PhD upgrade milestone.
4. Submit through the official UELA portal early enough to have both references recorded by 16 December 2026 if possible; the hard deadline is 8 January 2027 at noon UK time.

At interview

1. Which exact question and method family will be prioritized, and which are alternatives: efficient generative modelling, RL, adaptive data acquisition or a combination?
2. What resource constraints and evaluation metrics will define “efficient” (for example, compute, memory, energy, latency, data access or an explicit trade-off), and what baselines will be used?
3. Who is the named MediaTek co-supervisor, where and when is the three-month placement, and what project data, hardware or communication-system access is actually expected?
4. How will research outputs, confidential partner information, code/data access and publication review be governed?

Before accepting an offer

Obtain the formal award letter showing each year’s stipend, partner top-up, fee coverage, start/end dates, registration and any conditions on progress or placement. International applicants should verify that the CAS and sponsorship documents meet current visa evidence rules and budget for any visa, IHS, arrival and accommodation costs not covered by the award.

Evidence and Sources

Core Project, Programme and Funding

• UELA 2027–28 full project catalogue
• UELA 2027–28 collaborative-project catalogue
• UCL UELA studentships: funding, eligibility and timeline
• UELA applicant guidance (updated 23 October 2026)
• UCL Electronic and Electrical Engineering MPhil/PhD
• UCL Electronic & Electrical Engineering department
• UCL EEE subject-area ranking summary
• UCL Centre for Artificial Intelligence research areas
• UCL postal address

Supervisor and Recent Papers

• Sattar Vakili CV
• Sattar Vakili publications
• Delayed Feedback in Kernel Bandits (ICML 2023)
• Reward-Free Kernel-Based Reinforcement Learning (ICML 2024)
• Near-Optimal Sample Complexity in Reward-Free Kernel-based Reinforcement Learning (AISTATS 2025)
• Random Exploration in Bayesian Optimization (ICML 2024)
• Kernelized Reinforcement Learning with Order Optimal Regret Bounds (NeurIPS 2023)
• Generative Diffusion Models for Radio Wireless Channel Modelling and Sampling (IEEE GLOBECOM 2023)
• Adversarial Contextual Bandits Go Kernelized (ALT 2024)

Costs, Visa and Career Evidence

• UCL cost-of-study guide
• UCL research studentship stipend rates and administration
• GOV.UK Student visa financial requirement
• DSIT Artificial Intelligence Sector Study 2025
• DSIT AI Labour Market Survey 2025
• DSIT AI Skills for Life and Work summary