What is Statistical AI?

The world's most successful companies have revolutionized their operations by collecting and utilising data like never before. Companies such as Amazon, Google, and Tesco have harnessed the power of data to enhance their services and offerings. For instance, Amazon's personalized product suggestions cater to the unique preferences of each individual customer. Google's advertising campaigns are targeted at specific individuals, ensuring that the right ads reach the right audience. Tesco's Clubcard program tailors its offers based on a shopper's particular purchases. These remarkable capabilities are made possible through the application of learning algorithms, which have the ability to "learn" from data about the environment and user behaviour.

Statistical learning is a field that analyses and advances these algorithms by leveraging statistical theory and various techniques from the wider mathematics literature. It draws upon disciplines such as functional analysis, probability theory, and combinatorics to deepen our understanding of learning algorithms and develop new innovations. These mathematical foundations have profoundly influenced the analysis and evolution of learning algorithms, allowing researchers and practitioners to uncover patterns, make predictions, and gain insights from data.

As the availability of data and computational power continues to grow exponentially, the demand for statistical learning in industry is skyrocketing. Within Lancaster’s Statistical Learning group, we work with a number of leading industries to advance the applications of our research, with partners including Amazon, Microsoft, Shell and Tesco.

Current areas of research

• Bandit algorithms - companies face the challenge of selecting the most effective advertisement to display from a pool of several possibilities. Each display opportunity presents an opportunity to exploit an advert that is currently believed to be the best or explore a new option for which limited information is available. The management of this exploration-exploitation trade-off lies at the core of bandit research, and the researchers at Lancaster have made fundamental contributions to this field.
• Inference at scale - Bayesian inference is a statistical framework that allows us to update our beliefs about uncertain quantities based on observed data. It provides a principled way to incorporate prior knowledge and update it with new evidence. However, traditional Bayesian inference methods can become computationally intractable when dealing with massive datasets or complex models. Researchers at Lancaster have made significant contributions to the area of stochastic gradient Markov chain Monte Carlo algorithms for machine learning, developing the underpinning theoretical convergence for these algorithms and creating open-source software to support the use of these algorithms.

Academics in the Statistical Learning group supervise a number of PhD students in the department and within the STOR-i CDT. Through the MSc Statistics and MSc Data Science, dissertation projects on topics of statistical learning are offered to PGT students. At the undergraduate level, Statistical Learning techniques are taught in the 3rd Year Machine Learning course.