Recipient organisationUniversity of EssexSource-published name: University of Essex
Funding£570K
PeriodSept 2025 — Sept 2028
In plain English
AI plain-English summary
Machine learning models routinely fail when faced with data that differs from what they were trained on—a single unusually warm growing season can break a crop-yield predictor trained only on cooler years. This brittleness, known as out-of-distribution (OOD) generalisation failure, is invisible in high-dimensional data but pervasive across industries that rely on automated predictions. The problem is that current function approximators—the mathematical engines behind classifiers and regressors—cannot extrapolate beyond their training distribution. This project proposes a new type of approximator called an "element-compositor" (EC), which learns a set of low-dimensional functions ("elements") from data and combines them using simple, pre-defined "compositors." The goal is to build models that remain reliable when conditions shift. If successful, EC methods could make machine learning trustworthy in settings where data distributions naturally drift: energy grids facing unprecedented weather, medical diagnostics encountering new patient populations, or supply chains disrupted by unexpected events. The project also aims to establish a standard evaluation framework for OOD tabular data, quantify how much shift causes mainstream algorithms to fail, and develop refutation methods to check whether a learned composition is sensible. This is fundamental science—it does not promise an immediate product, but it addresses a core weakness that currently limits the safe deployment of machine learning in high-stakes systems.
View original technical description
As it stands, machine learning methods are brittle. If there is a data distribution shift between the training and test set, most methods tend to fail, often unpredictably, as they cannot extrapolate, For example, if one has only collected crops during relative cold periods, one abnormally warm period tends to confuse most models. These "distribution shifts" are not often visible, but omnipresent, as in higher dimensions it is not easy to discern where the data gaps lie. The ability to extrapolate from data successfully is often called "out-of-distribution" (OOD) generalisation. The aim of this proposal is to develop and new form of function approximator (i.e. a regressor or classifier), which we term "element-compositor" (EC), that is able to generalise OOD by assuming that the function under examination can be approximated using a set of low-dimensional functions learnt from data (that we term "elements"), which are than composed together by an a-priori simple set of functions we call "compositors" . While our overarching aim is to create robust OOD methods, OOD generalisation is still in its infancy, and we aim to develop a comprehensive research programme around it. Our individual objectives are as follows: (1) Establish a complete evaluation framework for OOD tabular data, which includes standard supervised learning, causal learning, and off-policy reinforcement learning. (2) Improve our understanding of how great the distributional shift should be before generalisation becomes impossible for mainstream algorithms. (3) Explore the number of compositions that are meaningful in real-world data through the development of new methods around the ``element-compositor'' paradigm we have introduced and measure their performance. (4) Develop refutation methods to help identify whether a composition achieved is sensible. Machine learning is used widely for prediction across all industries, and the development of new methods that generalise OOD will allow practitioners to trust that their models will not fail when conditions change.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know