Bio
Rachael Laidlaw is a PhD student on the Interactive AI CDT programme at University of Bristol. Her current research focuses on ecological applications of computer vision – in particular, transfer learning in order to fine-tune classification models for the automatic detection of data-deficient animal species from camera-trap images taken in the wild. She graduated from University of Leeds with an MSc Statistics in 2022, not long after successfully concluding her time on the Mathematics and German course at Lancaster University to obtain a joint-honours undergraduate degree in 2021. During the latter, she gained a wealth of international experience whilst living, working and studying abroad for a year. More recently, Rachael undertook an internship in the field of machine learning for earth observation at the European Space Agency's Φ-Lab in Italy, and prior to that, operated briefly as a UK government statistician for the Department of Work and Pensions. Alongside her PhD so far, Rachael has been involved with the Jean Golding Institute as a data science consultant, and was also both an event manager and host for her local Pint of Science festival last academic year. In terms of outreach, she attended Cheltenham Science Festival as a STEM ambassador and school competition judge, and was an invited speaker at an online careers event delivered by The Brilliant Club. She has a keen interest in science communication, too, demonstrated by her previous endeavours manning stalls or giving talks at various large-scale events such as AI UK, Festival of Nature and FUTURES.
Research interests
Training a computer-vision model for an ecosystem that contains species for which there are little to no labelled images can be difficult and time-consuming. As a result, Rachael’s research interest lies in trying to utilise the abundance of publicly available data within online repositories to leverage machine-learned knowledge from one context by transferring it into another, in order to improve the performance of animal classification models on scarcely photographed species. The main aims are (i) to boost the efficiency of the pipeline for training these tools by removing the need for unnecessary additional labour-intensive image collection or labelling, as well as (ii) to enable the models to achieve higher accuracies for classes that would otherwise be frequently misidentified due to the sparse training datasets associated with the actual target species. This is particularly relevant for conservation efforts such as the monitoring of rare or endangered species using camera-trap data.