Machine Learning for Changing Environments: Using covariate shift to detect and address concept shift in continual learning settings
Interview date: Tuesday 27th May 2025
Background
Most existing machine learning methods are static; while the models they generate can generalise to deal with instances that they have not seen before, those instances must be drawn from the same population as the ones the model has been trained on [1, 2, 3]. So, a model trained to distinguish weeds from crops will work on images it hasn’t seen, but they must be images of the same crop and weeds at the same stage of growth [4]. This is a limitation for applications in agriculture and environmental monitoring where plants will change as they grow and as the environment changes with the climate. At an abstract level we can think of the issue as dealing with changes in the distribution of instances that make up the data.
The project
This ambitious EPSRC-funded PhD project aims to mitigate the negative effects of the data distribution changes, specifically covariate and concept shifts on the performance of AI models during long-term and continuous deployment. Covariate shift refers to changes in the distribution of input features, while concept shift pertains to changes in the relationship between input features and the target task. Notably, covariate shifts often precede concept shifts in real-world scenarios. By effectively detecting covariate shifts, we aim to proactively address subsequent concept shifts, thereby maintaining AI model robustness in continual learning environments. The project will investigate methodologies for detecting and quantifying covariate and concept shift in the input data and AI models and explore selective retraining and domain adaptation techniques in the continual learning setting.
The project will closely collaborate with industrial stakeholders at Crop Intellect (https://cropintellect.co.uk/), and the successful candidate will have access to field data trials pertaining to agronomic optimisation, and benefit from the in-house company expertise on plant physiology, environmental sciences and analytical chemistry. The successful candidate will develop approaches to address the performance degradation of environmental monitoring AI systems over time and improve their robustness to domain shift, especially targeting the changes introduced by climate change.
Candidate requirements
We are looking for a curiosity-driven candidates with strong analytical skills, with a degree (minimum 2:1 bachelor’s degree, or international equivalent) in Computer Science or a closely related discipline. Previous experience with machine learning or deep learning is desirable. We encourage candidates passionate about designing advanced AI techniques, and particularly those with an interest in interdisciplinary research and addressing sustainability challenges, to apply for this position.
Research environment
Dr Mamatha Thota is a Senior Lecturer in Computer Science, while Dr Petra Bosilj is a Senior Lecturer in Computer Science (Imaging/Vision), both in the School of Engineering and Physical Sciences. Prof Simon Parsons is a Global Professor and the Director of Research in the Lincoln Institute for Agri-Food Technology. Their respective schools are currently hosting two separate UKRI and EPSRC-funded Centres for Doctoral Training, AgriForwards and Sustain, and the successful candidate will be joining a vibrant research environment made up of many PhD students and postdoctoral researchers. The candidate will also have close connections to a wide range of interdisciplinary expertise by working alongside research teams focusing on plant and soil science as well as machine learning, computer vision, and agri-robotics.
To apply for this project, visit our Apply Now page.