About

NUOMI Lab is an artificial intelligence (AI) research group at London South Bank University (LSBU). Our name is short for Normative and User-Oriented Machine Intelligence: we study what AI systems should do, not only what they can do; we build technology around the people and communities who use it and who are affected by it.

Normative. Every AI system encodes choices about whose goals count, whose values drive outcomes, and who bears the risks of AI system failures. Our research makes the values implied by the technical design of AI systems explicit choices; we ground our choices in human values and social context, and develop machine learning (ML) methods that hold AI systems accountable to these values.

User-Oriented. An AI system is only useful if the people relying on it can understand it and decide when and how to use it, and if the system complements human expertise and capabilities. We design AI systems personalised to the needs of diverse real users making decisions in specific real-life settings.

In our work, we collaborate closely with communities (mental health advocacy groups, clinicians, educators), AI research groups worldwide (across academia and industry), as well as policy makers interested in governing the impact of frontier AI technologies.

Research

We combine technical research on ML methodology and AI systems design with human-centred evaluation, so that the AI systems we develop are judged by whether they help real people make better decisions, not only by benchmark performance.

1. Making Values Explicit: We develop ML frameworks for specifying what a “good” AI model means in human contexts, including fairness, accountability, and trade-offs between the goals of different stakeholders (quantified as training objectives and metrics). We build and train AI models that optimise for capturing human values.

2. Models People Can Understand and Question: We develop ML methods for interpreting complex AI models and for estimating and communicating uncertainty in the outputs and the internal reasoning of these models, so that users can tell when a model’s output should, and should not, be relied on.

3. Evaluating AI With Its Users: We rigorously study how people actually use AI systems in high-stakes settings (e.g. disease diagnosis, mental health support), and use what we learn to rework how those systems are designed, evaluated, and deployed.

Technical Research Themes

The themes below come from prior work by our PI, Weiwei Pan, and collaborators. For each, we describe the main machine learning approach and the central insight behind it, with a few papers as examples.

• Behaviourally grounded reinforcement learning: We model the people an AI system assists as reinforcement learning agents, with interpretable parameters for human traits (such as how much they discount the future, or how costly they find effort), and frame AI assistance as an intervention on those parameters. Inverse reinforcement learning lets us infer these traits from behaviour.
  • Key insight: Helping someone does not require a complete model of human cognition, only a minimal, task-specific model of the few traits that drive their behaviour on the task at hand.
  • Example work: Nofshin et al., 2024; Moore et al., 2025; Yao et al., 2024.
• Sequential decision-making under real-world constraints: We develop bandit and reinforcement learning methods for high-stakes settings such as healthcare, where an algorithm must personalise while also producing valid scientific evidence, acting safely, and learning from scarce and uneven data.
  • Key insight: Standard algorithmic choices carry hidden assumptions that break under deployment conditions. Building the real constraints into the learning problem itself, rather than handling them afterwards, yields methods whose guarantees hold where they matter.
  • Example work: Yao et al., 2021; Rathnam et al., 2023; Brown et al., 2025.
• Interpretable machine learning: We treat explanations as measurement instruments. The properties an explanation should have (such as faithfulness, robustness, and simplicity) are formally defined, quantitatively validated, and directly optimised, rather than encouraged by heuristics.
  • Key insight: An explanation is useful only if we know which property of the model it measures, and how reliably. Different tasks need different properties, so explanation design is a multi-objective problem with task-dependent trade-offs.
  • Example work: Chen et al., 2022; Mahinpei et al., 2021; Tadesse et al., 2025; Nofshin et al., 2024.
• Bayesian deep learning and approximate inference: We study probabilistic models such as Bayesian neural networks and deep latent-variable models, combining theoretical analysis with controlled experiments to characterise when approximate inference methods (such as variational inference) yield reliable uncertainty.
  • Key insight: Approximate inference can fail silently, producing uncertainty that looks reasonable on standard metrics but does not reflect the data. These failures often have structural causes, such as non-identifiability or the inference objective itself, that can be diagnosed and corrected.
  • Example work: Yao et al., 2019; Coker et al., 2022; Yacoby et al., 2022.
• Representation learning and model steering: We design models whose internal structure supports targeted intervention, such as latent-variable models with interpretable factors and ensembles with deliberately diverse members, and develop methods for analysing and steering the internal representations of large language models.
  • Key insight: When representations are entangled, an intervention aimed at one behaviour can change many others. Control methods must therefore be evaluated for fidelity (whether they change only what was intended), not just for aggregate success.
  • Example work: Ross et al., 2020; Subhash et al., 2023; Raedler et al., 2025.
• Evaluation and auditing of generative models: We evaluate generative models through the distribution of their outputs across prompts, tasks, and populations, described in human-interpretable concepts, and draw on measurement theory from the social sciences to design probes.
  • Key insight: Properties such as bias are not fixed attributes that a single benchmark can certify; they shift with the task and the prompt. Evaluation must be task-specific and grounded in a theory of when a measurement generalises to real use.
  • Example work: Boussard et al., 2024; Abdel Magid et al., 2025; Morehouse et al., 2025.

Designing With Communities

Our HCI research embeds us in communities of practice over sustained collaborations. Through interviews, iterative co-design, and user studies, we learn how practitioners work, what expertise they bring, and what success means to them, and then design AI tools around those needs. We also teach workshops on AI fundamentals, so that practitioners can take part as informed co-designers and judge AI tools for themselves.

• Key insight: Community needs often redefine the technical problem itself. Humanitarian frontline negotiators did not want AI that tells them what to do; they wanted support for analysing context, exploring alternatives, and assessing risks. That shifted the design goal from replacing expert judgement to strengthening it.
• Example work: Ma et al., 2025, “ChatGPT, Don’t Tell Me What to Do”: Designing AI for Context Analysis in Humanitarian Frontline Negotiations.

How to join?

Master’s Students and Visiting Students

NUOMI Lab welcomes LSBU master’s students who are interested in exploring AI/ML-related research, and we also welcome self-funded visiting research students. If your interests overlap with the lab’s research, please reach out to the PI with:

1. a short explanation of your interest in the lab
2. a short summary of your background
3. a CV

PhD Students

NUOMI Lab may have PhD openings in 2027 (details to follow). Applicants should have:

1. a master’s degree in Computer Science, Mathematics, Statistics, or a related field (or an equivalent degree with substantial mathematics and programming components)
2. courses in Machine Learning and in Probability or Statistics

In addition, it’d be desirable for potential candidates to have:

1. coursework or experience that engages with how technology affects people, for example in human-computer interaction, ethics, the social sciences, or an application domain such as healthcare
2. hands-on experience training and evaluating ML models

Note: Don’t be discouraged if you feel like you don’t have all the listed qualifications! These skills can be learnt with dedication and drive. If you are not sure your background is a good fit for NUOMI Lab but are keen to join us, please reach out to the PI.

Research Collaborations

NUOMI Lab welcomes collaborations with researchers and domain experts in academia, industry, and the public sector. If you are interested in working with us, please reach out to the PI with a statement of interest, including:

1. the problem you’re interested in solving
2. how you think NUOMI Lab can help
3. the kinds of deliverables you’d be interested in and the types of collaborations you’re open to

Diversity and Inclusion

Our Philosophy: Inclusion requires more than offering everyone the same opportunities. People arrive at research through social and institutional systems that have already shaped their preparation, their resources, their sense of safety, and whether they have been recognised as capable. Identity and lived experience are therefore not “personal” matters separate from research training: they affect who is heard, who receives informal guidance, who feels safe admitting confusion, and who has time to take on research or unpaid opportunities. Ignoring identity does not create a neutral environment; it leaves existing inequalities in place. When people know their experiences will be heard rather than minimised, they ask for help earlier, take intellectual risks, and can give more of their attention to research.

Over time, differences in access turn into seeming differences in achievement. Students from well-resourced backgrounds reach advanced coursework, research, and internships early, while equally capable students from underserved backgrounds may spend years completing prerequisites, often while managing jobs and long commutes. We work towards changing processes that decide who gains early access to advanced work, professional knowledge, and scientific networks.

Our Practice: In NUOMI Lab, we enact our philosophy by:

• making expectations and routes to support explicit, and offering many ways to ask for help;
• tailoring mentorship to each person’s preparation and goals rather than following a rigid template, and treating disagreement about research choices as part of training;
• recognising that unfamiliarity with academic conventions, financial pressure, health, or doubts about belonging can affect anyone’s work, and helping people through these barriers without lowering expectations;
• taking people’s accounts of their own experiences seriously, and stepping in when group norms silence or marginalise someone.

Equity is also part of the intellectual content of our work. How a technical problem is formulated encodes social assumptions, and the same forces that shape who participates in technology also shape who bears its effects when it is deployed. We treat identifying and acting on these inequities as a core skill for anyone who builds AI, not only as a matter of interpersonal kindness.

Contact

Please reach out to the PI, Weiwei Pan, with questions: nuomi.lab at gmail dot com.

Resources

Resources for prospective and current students.

For Prospective Students

• How to join NUOMI Lab: current openings, what we look for in applicants, and how to get in touch.
• What research is (and isn’t): common misconceptions about research, and reasons you might want to do it. A good place to start if you are considering a PhD.

For Current Students

We have put together a set of materials to guide you in developing concrete research skills, collected on our Research Skills page. Good places to start:

• Four Fundamental Skills of Research: defining and scoping a research problem, surveying related work, sustaining progress, and communicating results, with what we expect for each.
• Research Best Practices: how to make the most of your research meetings, your research week, and your code.
• How to Write a Machine Learning Research Paper: how to go from a one-sentence-per-point introduction to a polished paper.
• Skills for a Research Career: a map of the skills researchers keep developing, from managing research to managing yourself.

For the technical foundations of our work:

• Introduction to Machine Learning: a set of lecture notes by Weiwei Pan for an introduction to machine learning course that they taught at Harvard. These notes cover regression and classification, neural networks, Bayesian modelling and inference, and latent-variable models, and treat uncertainty, interpretation and responsible use (not only prediction) as primary goals.
• Probabilistic Inference & Optimization: a textbook by Weiwei Pan based on the course on foundations of probabilistic machine learning that they taught at Harvard for five years. It covers Bayesian modelling, Markov chain Monte Carlo, variational inference and stochastic optimisation, with an emphasis on implementation.

Further Reading

The following are materials for research students that we co-developed with MOGU Lab:

• How to communicate about research progress: how good communication with collaborators gets you unstuck and leads to clearer, more structured feedback on your work.
• How to read a research paper: strategies for reading dense papers that assume a lot of background knowledge.
• How to make the most of your PhD: a guide from current to incoming PhD students on the “hidden curriculum” of doctoral programmes, useful for graduate students in general.