How we work in artificial intelligence

Contents

Examples from UKRI’s AI investment portfolio

Fairer algorithm-led decisions

Big data and artificial intelligence (AI) are increasingly used to make decisions about humans.

These decisions range from assessing creditworthiness, hiring and firing decisions, to sentencing criminals.

Due to the ‘black-box’ nature of these systems, we are often unable to understand their assessments and decisions, which can lead to privacy-invasive and discriminatory outcomes.

Researchers at The Alan Turing Institute have explored what European courts consider ‘fair’. The researchers are:

  • Sandra Wachter
  • Brent Mittelstadt
  • Chris Russell.

The team proposed a statistical test in their paper ‘why fairness cannot be automated’ (conditional demographic parity) which aligns with this legal notion of fairness and can be embedded in AI systems to detect and prevent discriminatory decision-making.

This method is especially successful in detecting discrimination that other methods have previously overlooked, namely:

  • minority-based, for example: religion, ethnicity, sexual identity
  • intersectional, for example black women.

The same team has also proposed counterfactual explanations. This method was among the first concrete and technically feasible solutions to compute ‘good everyday explanations’ of decisions made by black box models that are often thought to be fundamentally incomprehensible to humans.

Companies that have implemented counter-factual explanations in their products include:

  • Google
  • Vodaphone
  • Flock
  • IBM
  • Accenture
  • among others.

Using AI to understand COVID-19

Health records are commonly stored in multiple incompatible formats, with much of the useful data held as text. This means that the information can be hard to access and utilise.

Scientists at the national institute for health data science, Health Data Research UK (HDR UK), are developing novel natural language processing methods to extract information from these unstructured data sources.

With these powerful data mining techniques, researchers can search through medical records quickly and efficiently, helping to understand diseases and develop more effective treatments.

The team analysed medical notes of COVID-19 patients. Their preliminary data suggests that patients taking medicines known as ACE-inhibitors to manage high blood pressure and diabetes, are no more susceptible to a severe form of COVID-19 infection.

ACE-inhibitors were initially thought to exacerbate COVID-19 symptoms. These findings indicate COVID-19 patients can continue receiving treatment to manage their underlying conditions without further detrimental effects to their health.

Deep machine learning to understand plant performance

Cultivating higher-yield plants can contribute to addressing the challenge of global food security. Scientists use a variety of techniques. For example, crop phenotyping, to identify types of plants that are:

  • more likely to produce higher yields
  • more resistant to weather conditions or diseases.

This information can be used by crop breeders to make more efficient and reliable decisions about what plants to cultivate.

However, it can be time-consuming for a human to assess how well a candidate plant performs in real life and it can be limited by the amount of information available for the analysis.

With funds from the Biotechnology and Biological Sciences Research Council, researchers explored the use of convolutional neural nets (CNN) to speed up evaluating plant performance. In collaboration with Syngenta, the team included:

  • Professor Tony Pridmore at the University of Nottingham
  • Dr Andrew French
  • Dr Michael Pound
  • Dr Aaron Jackson.

CNNs are a type of artificial deep neural networks used in image analysis and computer vision. The exploratory projects focused on using this technology to detect and count biological objects, such as seeds or insects on leaves, which provide information about the performance of the plant.

The researchers showed CNNs to be highly effective at analysing the available material. For example, through using CNNs, the time needed to count the number of insects on a single leaf is reduced by 20 times, compared to the time a human would need to perform the same operation. This leads to significant time and cost savings for researchers in plant sciences, ultimately supporting crop breeders at producing food as efficiently as possible.

Next generation of autonomy inspired by insects

Opteran is a spin-out company from the Engineering and Physical Sciences Research Council funded grant ‘brains on board: neuromorphic control of flying robots’, at the University of Sheffield. The spin-out is building on eight years of research led by Professor James Marshall and Dr Alex Cope which aimed to model the brain of a honeybee.

Insect brains are capable of perceiving depth and distance to aid their navigation and decision-making process.

By reverse engineering the processes that occur in the insects’ brains, Opteran has been able to produce algorithms that do not require extensive pre-training or huge amounts of data to make decisions.

This advance is a key block to the autonomous technologies of the future and could be used in a range of systems such as drones, cars and robots.

Opteran has secured significant £2.1 million seed funding to continue expanding their pioneering work on lightweight, low-cost silicon brains to enable robots and autonomous vehicles to see, sense, navigate and make decisions.

Last updated: 17 September 2021

NOTE Council web content is being transitioned to this website – let us know if you have feedback or would like to help us test new developments.