PhD Supervision

I am currently accepting self-funded PhD students for the following topics.

Prospective students must have their own funding and must conform to the University of Sheffield's entrance requirements. Please use my contact page with a note of your interests and background if you're thinking of applying.

1. Max Headroom meets Majel Barrett — how can we build usable AI on top of hallucinating LLMs?

The capabilities of generative large language models (LLMs) have suddenly reached a performance threshold sufficient to enable multiple applications. These new possibilities have projected Artificial Intelligence (AI) into many new areas, but in doing so provided multiple examples of the negative safety and reliability implications of LLMs tendency to hallucinate (or, to be less euphemistic, confidently state bald falsehoods (Zhang et al. 2023)). In addition it is in principle impossible to prevent injection attacks on networked LLMs (Peng et al. 2022; Willison 2023b, Willison 2023a). The work proposed here is to develop mechanisms that can be interposed between LLM and end user in such a way as to mitigate unreliability. If these mechanisms target models capable of running off-network (LLAMA, Vicuna, etc. (Touvron et al. 2023; Zheng et al. 2023)) then injection attacks can also be mitigated.

1.1. Key objectives / outcomes

Work on restricted domain control systems — for example smart homes — is beginning to exhibit both command isolation and model independence (the latter enabled by heavily optimised sub-20 billion parameter LLMs that can run on a single consumer grade GPU) (Hewitt and Cunningham 2022).

The objective of this project is to generalise and extend this type of approach to control problems of multiple types and thus extend the safe application terrain for the new technology.

1.2. Description

Max Headroom was an early TV AI character; Majel Barrett the voice of the Starship Enterprise, which, at least in later generations, constituted an Artficial General Intelligence (AGI). Max was glitchy and off the wall — you might find him likeable but you probably wouldn’t lend him your car. Majel’s Starship was a bit like Spock: rational, factual, calm, but often too literal. She inspires trust (though not necessarily confidence) from the moment she starts speaking.

We have a similar dichotomy now: precise and inflexible automatic control systems on the one hand, and the new world of unpredictable, unreliable and unsafe — but nevertheless almost miraculously human — generative large language models on the other. Control systems in general, and computational user interfaces more specifically, strive to eliminate ambiguity for obvious reasons. I don’t want my car to interpret “turn left” as a comment on how I would like to change the BBC’s radio news coverage at a crucial moment, for example, but without ambiguity we also sacrifice inference. I would like my car’s navigation system to respond to “maybe time for a break” by suggesting a quiet village not far from the current route where I’ve previously stopped for coffee. This research project is to investigate mechanisms at the crossover of these two objectives: usable and safe automated control that nevertheless exploit the inferential capabilities of LLMs.

1.3. References

  • Hewitt, Mary, and Hamish Cunningham. 2022. “Taxonomic Classification of IoT Smart Home Voice Control.” on arXiv [cs.HC].
  • Peng, Xutan, Yipeng Zhang, Jingfeng Yang, and Mark Stevenson. 2022. “On the Security Vulnerabilities of Text-to-SQL Models.” on arXiv [cs.CL].
  • Touvron, Hugo, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, et al. 2023. “LLaMA: Open and Efficient Foundation Language Models.” on arXiv [cs.CL].
  • Willison, S. 2023a. “We Need to Tell People ChatGPT Will Lie to Them, Not Debate Linguistics.” April 7, 2023.
  • ———. 2023b. “The Dual LLM Pattern for Building AI Assistants That Can Resist Prompt Injection.” April 25, 2023.
  • Zhang, Yue, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, et al. 2023. “Siren’s Song in the AI Ocean: A Survey on Hallucination in Large Language Models.” on arXiv [cs.CL].
  • Zheng, Lianmin, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, et al. 2023. “Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena.” on arXiv [cs.CL].

2. Aquaponics Monitoring and Control Systems for Local Sustainable Food Production and Community Resilience.

Increases in domestic and community food production promise to reduce transport-related carbon emissions and promote resilience in face of supply chain interruptions or economic shocks. Aquaponics is a low-impact, high-density agriculture method which has a recognised potential to increase sustainability of food production, according to the UN’s FAO and others (Somerville et al. 2014; Kotzen et al. 2013; Goddek et al. 2015). By combining elements of aquaculture and hydroponics, it offers reduced input requirements and waste disposal load in comparison to each, while still providing high volume fish and vegetable outputs. Year-round growing is possible even in unfavourable climates, and operations can be tailored to many types of environment, including arid, urban and peri-urban areas. The recirculation and biofiltration of water performed in aquaponics systems can reduce load on agricultural water sources (in contrast to both hydroponics and soil-based growing).

The use of aquaponics is increasing, but there are several factors which create barriers to entry, including:

  • Expertise. The method relies on a continuously balanced multi-element ecosystem and requires a combination of technical and agricultural skills that are not widespread. Further, optimal configuration varies radically according to local conditions, and a formulaic statement of how to tailor the method is not yet available for many cases.
  • Setup costs; energy usage. These are of the order of aquaculture and hydroponics systems, i.e. significantly higher than soil-based growing.

We propose to lower these barriers by a combination of technical innovations and social interventions. To reduce the skills barrier we will deploy open source measurement and control electronics to domestic scale aquaponics systems. The data collected will be aggregated and abstracted in the cloud and shared in a community of back garden growers bootstrapped via a crowdfunding campaign. We will engage schools via events, schoolyard systems and lesson elements. The longevity of the technology will be ensured by nurturing a community of open source, open hardware and creative commons developers and contributors.

Domestic and school aquaponics can be a springboard to community scale systems, boosting local food production and making a significant contribution to food sustainability and community resilience.

  • Goddek, Simon, Boris Delaide, Utra Mankasingh, Kristin Ragnarsdottir, Haissam Jijakli, and Ragnheidur Thorarinsdottir. 2015. “Challenges of Sustainable and Commercial Aquaponics.” Sustainability: Science Practice and Policy 7 (4). Multidisciplinary Digital Publishing Institute: 4199–4224.
  • Benz Kotzen. 2013. “Memorandum of Understanding for the Implementation of a European Concerted Research Action Designated as COST Action FA1305: The EU Aquaponics Hub: Realising Sustainable Integrated Fish and Vegetable Production for the EU.” COST Cooperation in Science and Technology.
  • Pearce, Joshua M. 2012. “Open Source Research in Sustainability.” Sustainability: The Journal of Record 5 (4). 238–43.
  • Somerville, Christopher, Moti Cohen, Edoardo Pantanella, Austin Stankus, and Alessandro Lovatelli. 2014. Small-Scale Aquaponic Food Production: Integrated Fish and Plant Farming. FAO Fisheries and Aquaculture Technical Paper No. 589. FAO.

3. Internet of Things

The Internet of Things: mi casa su botnet? I am working on a variety of IoT projects using the ESP32 wifi microcontroller, including:

  • It stinks! An air quality monitor.
  • CrowdEcurity. A campus panic button.
  • Scuttlebot. A web-controlled mini robot.
  • WaterElf. An aquaponics sustainable food controller.
  • Opinionmungers
    • That's rubbish! A lecture quality feedback system.
    • VoteImp. A liquid democracy voter system.
  • Audiscians. Musicians and audience blend into a symbiotic entity.
  • Slow down!. An off grid car speed warning display.

Student-defined research projects are also welcome.

4. Cloud Privacy in the Age of Surveillance

Two modern trends are colliding head-on: 1. the migration of computational tasks onto the cloud; 2. the appetite of governments for surveillance. How can we preserve privacy amidst this collision?

5. Software Architecture for Multilingual Cultures

Diversity is just as big an advantage for social evolution as it is for biological evolution. What computational infrastructure is appropriate to support multilingual diversity, and how can it be efficiently expressed in cloud-centric environments?

6. Collective Awareness for Sustainability and Resilience

It is now widely feared that dependence on high levels of consumption makes our economic systems unsustainable. Socio-economic models that require the natural resources of more than one planet (or the invention of revolutionary new ways of collecting and accumulating energy) continue to drive carbon emissions, block appropriate social responses, and generate instability and crisis. Many citizens now view a shift to sustainable societies as vital and believe that a key challenge of the age is to increase societal resilience in face of shocks from increased costs of basic resources like energy or food, from interruption of resource supplies and from the effects of a changing climate. Unfortunately, this awareness does not inevitably lead to sustainable practices. Awareness alone is not sufficient to stimulate widespread behavioural change. Can we turn the problem on its head and build awareness via grassroots involvement in processes of transition to sustainability that generate immediate positive benefits that then reinforce involvement in a positive feedback loop?

7. Open Technology for Resilience

That light at the end of the tunnel? It's a train :-( Climate chaos is accelerating and our market-driven system has no mechanism to respond. (There's no short-term profit in saving the planet!) What to do? Many things, but one of them is to build resilience: our capacity to withstand shocks and changes. A society is resilient if it can supply its needs (e.g. for food, shelter or health) without relying on systems outside of its control. How can openness help resilience? Open source software is the only possible defence against indiscriminate state surveillance. Open data can help: the more transparency we have in the use of government and corporate funds the more we can understand the implications of policy, deregulation and the like. Open hardware contributes to resilience by making designs repeatable, adaptable and local.