Probabilistic programming
Probabilistic programming is an emerging research area that has attracted interest from the Artificial Intelligence (AI) and Machine Learning (ML) communities, as well as statistics and theoretical computer science/semantics. The basic concept of probabilistic programming is to use ideas from programming languages to structure complex statistical models.
Following an open call for submission, the Turing announced a new portfolio of projects that promote the development, adoption, and awareness of probabilistic programming within the UK:
- Adjoint-accelerated Programmable Inference for Large PDEs
- A UK Spatial, Climate & Health Probabilistic Programming Language Community
- Development of composable, parallelisable and user-friendly inference, and growing the community of the Turing.jl probabilistic programming language
- Probabilistic Programming for Aquatic Ecosystem Models
- Robust Inference with Probabilistic Answer Set Programs Scaffolds for Large Language Models
Understanding the capabilities of foundation models
Foundation models are machine learning models trained on vast, broad datasets so that they can be adapted for a wide range of tasks. A well-known type of foundation model is large language models, which power chatbots such as ChatGPT and have demonstrated remarkable proficiency in generating realistic language, as well as some ability in solving problems and performing commonsense reasoning. However, these models can also fail on apparently simple tasks, in unpredictable ways.
This project aims to develop new methods for robustly evaluating the capabilities and shortcomings of foundation models, to allow researchers and policy makers to make informed decisions about models’ safety and utility. Our benchmarking techniques will focus on three key attributes:
- Learned values. To what extent can the model be said to learn and consistently apply human values?
- Commonsense reasoning. To what extent does the model understand the physical, causal, spatial, temporal and social rules that govern the world?
- Theory of mind. To what extent is the model able to acquire social reasoning, taking into account a range of beliefs and aspirations?
Learn more about Understanding the capabilities of foundation models.
Automated analysis of strategic interactions
Game theory is a central part of contemporary AI. It is concerned with strategic decision-making, i.e. situations where the decisions of participants are influenced by the decisions of others. In this project, we will develop new tools to support strategic interactions between automated and human agents in complex and changing environments. This work could have applications across a broad range of areas, including AI-powered auction and commerce platforms, cyber defence, health service management and sports modelling.
Our project has two main workstreams:
- We will extend Gambit (the leading software package for doing computation in game theory) by developing and implementing state-of-the-art algorithms that support automated strategic reasoning at the scales required for practical application.
- We will develop computational methods for analysing large games that have too many possible choices/outcomes to be written down explicitly. This will be an important step towards automated analysis of the complex strategic interactions found in AI applications.
Learn more about Automated analysis of strategic interactions.
Participatory budgeting for chores
Citizens can be given a say in how public money is spent through a process called participatory budgeting. By making decisions about how a council’s budget is allocated, for example, residents can help direct resources to where in the community they are most needed.
The challenge of satisfying as many citizen decision-makers as possible while staying within budget can be formulated as an algorithm that seeks to achieve the fairest outcome. In this project, we will develop methods for fair decision-making in situations where residents have to choose between a number of potentially unpopular options, such as when deciding how to reduce energy consumption to combat climate change or how to cut services to deal with a budget deficit. This algorithmic budgeting for ‘chores’ is a largely unexplored research area.
Ultimately, we aim to work with local authorities to help them implement more efficient, better justified and less biased decision-making processes.
Learn more about Participatory budgeting for chores.