What is ValidateAI?

---

We are championing a voluntary code of standards by facilitating community discussion to signpost existing best practice and develop new standards where needed.

Our primary objectives are to validate AI and promote:

1. Assurance and practitioner-centric standards.
2. Innovation in how we validate AI.
3. Convening of the community to AI validation best practice.

The inaugural Validate AI conference was inspired by the participation of the co-proposers to discussions at the Royal Society in 2018 about the development of data science skills nationally, promoting greater collaborative work between public, private and academic sectors.  These discussions were shared in a Royal Society publication titled Dynamics of Data Science Skills.

The importance of the Validate AI theme was also highlighted through a specialist Masterclass learning program led by Shakeel Khan at HMRC and Dr Tony Bellotti of University of Nottingham Ningbo China to promote machine learning maintenance best practice that is commonly found in the financial services sector that is very relevant to tax authorities.

We are particularly grateful for the immense contribution of the sponsors of the launch event at the Royal Society on the 19thNovember 2019, Imperial College London - Data Science Institute, Ada Lovelace Institute and The Operational Research Society. We were also very pleased that the inaugural Validate AI conference ran concurrently with the Imperial College London - DSI's 5th Anniversary to celebrate their world class contribution to data science.

Algorithms touch every aspect of our modern lives, helping us make millions of decisions every day. The additional monetary benefit of Artificial Intelligence (AI) to UK GDP has been estimated to be up to £232bn by 2030. The use of algorithms is also growing exponentially and is well established across a broad spectrum of sectors, including banking, health, taxation, environment, agriculture, security and defence. Examples of the impact on individuals are numerous and include supporting the diagnosis and prognosis of life-threatening illnesses such as heart disease and cancer, self-driving cars, identifying high risk taxpayers, crop management, identifying banking fraud cases, customer experience enhancement and disease containment. 

There is also a major debate ongoing about public acceptance of AI and concerns that such technology may compromise fairness and quality of decisions when replacing the many roles previously undertaken by humans. Governments, and indeed the private sector, recognise the need to demonstrate that AI is both ethically sound and robust for the task it was developed for, in order to maintain public trust. Democratisation of algorithm development and the recent relative ease with which such solutions are being created and deployed provides great opportunities for more widespread economic and societal benefits. With this also come risks to AI validity and ongoing maintenance. There is an immense responsibility to ensure models are fit for purpose and fair both in the public and private sector, and importantly for this to be demonstrated openly in society.

The assessment of quality and methods for AI maintenance may differ dramatically depending on the nature of the problem and sector to which the AI is being applied. There may also be legislative constraints that impact on the particular algorithm, requiring greater transparency, for example, in how a specific outcome or decision is reached.

Whilst it is now becoming easier to produce AI solutions at lower cost, we need a more rigorous debate to raise the importance of measuring whether the solutions are fit for purpose, safe, reliable, timely and trustworthy. This initiative explores how such systems can depart from this ideal, examining tools and methods for ensuring sound and appropriate behaviour in a variety of different application domains, and looking at open challenges. Issues explored will include accurate and unbiased performance and its evaluation, model testing and formal verification, ensuring resilience against adversarial attacks, and the effective maintenance of systems as their working environment evolves. This last point may include changing populations, increasing data loads, new and unforeseen kinds of data, and policy and other changes.

A critical aspect of our conferences is to bring together representatives from public, private and academic sectors to promote greater collaboration of these issues, sharing experiences, challenges, and solutions. There are for example sectors where AI reliability metrics have a reasonable level of maturity, such as for customer lending in financial services, and lessons need to be learnt from such practices to promote awareness and adoption across sectors where appropriate.