AI sketch

November’s session surrounding the impact of AI on statistical careers was led by Nick Schurch (BioSS) and Brett Drury (Rothamsted).

Why should statisticians care about AI?

Having encountered a lot of scepticism towards AI amongst the statistical community, Nick hopes to encourage constructive engagement. AI is unlikely to go away and is already impacting our work in several ways:

  • How we work and what we work on

  • How the quantitative science community works

  • Distribution of funding - AI-focussed grants and collaborators looking for AI-based solutions

  • Expectations of our collaborators, partners and stakeholders - demand for more sophisticated output, digital twins, integration across large datasets etc; expectations of lower costs and faster turnaround; and a focus on predictions that are more accurate

We looked at some suggested definitions – while opinions can differ a lot, it may be helpful to think of AI as the broad field encompassing methods that enable machines to perceive their environment and learn to maximise their chances of achieving defined goals, ML (machine learning) being the subfield of AI involving algorithms that learn from data, and statistics being the foundation underpinning these.

We briefly looked at timelines (1, 2, 3) summarising developments and trends in AI, noting that the sustained focus on development and improvement of AI over the past ~15 years might indicate that AI technologies are coming into their own. However, different aspects of AI are at different stages of the hype cycle, e.g. computer vision being long-established while excitement surrounding generative AI may decrease as its shortcomings become more apparent.

Introducing AI into an agricultural organisation

During the session we were also joined by Brett Drury, an AI researcher and lecturer on secondment to Rothamsted from Liverpool Hope University. Brett’s role at Rothamsted includes the introduction of AI/ML technologies to maximise the potential of existing data assets, by encouraging voluntary buy-in from researchers. In his experience so far, he has found the following to be helpful:

  • An approach that involves identifying what researchers find difficult with current approaches, identifying their end goals, meeting often, and offering quick demonstrations of what can be done

  • Avoiding jargon when communicating and inviting questions and feedback from researchers

Brett is currently collaborating on a range of projects at Rothamsted where the use of AI/ML allows existing research to be sped up. Some of the approaches used in these projects include computational chemistry, literature mining and natural language processing (NLP), and computer vision.

Small group discussions touched on a few themes:

  • Determining the robustness, reliability, and transparency of AI models developed by others

  • The importance of the coverage, quality, and description of data used to train AI models

  • Using simpler or more traditional models when they are fit for purpose

Nick also covered some of the ways AI methods might be used in statistical work, especially to process large, complex, and high-dimensional datasets, which AI models excel at. Mature programming toolkits like scikit-learn, TensorFlow, and Keras have facilitated easy use of ML/AI methods. However, the current emphasis is on predictive power rather than explanatory power and interpretability, though explainable / interpretable AI are developing fields that will impact this over time.

We looked at the roles statisticians might have in a landscape where AI approaches are more prevalent:

  • Improving the scientific credibility of AI approaches by focussing on rigour, validation, transparency, interpretability, and detecting and quantifying biases

  • Tempering the amount of trust put in algorithms - providing critical and contextual understanding of the methods and data used to train them, and interpreting results appropriately

  • Being ethical champions – making sure AI approaches are used responsibly in fair, sustainable, and open ways

  • Leveraging our collaborative skills to engage with AI experts and provide sceptical positivity – focussing on when AI approaches are useful and emphasising scientific rigour.

Nick suggested a few ways we could be using NABES in relation to AI in statistics:

  • Feedback to stakeholders and funders, including participation in committees or advisory groups, and providing advice on how realistic outputs from AI approaches might be

  • Shaping organisational policies on AI use – when and where AI can or should be used

  • Networking – sharing our experience, knowledge, and resources