Lead Supervisor – Prof Kate Jones (University College London)
Co-supervisors – Dr Oisin Mac Aodha (University of Edinburgh), Prof Gabriel Brostow (University College London)
Wildlife and wild places face a variety of environmental and human-induced challenges, with catastrophic declines documented in both invertebrate and vertebrate populations across the world. Despite a number of global international commitments to reduce biodiversity loss by 2020, current assessments show that these are unlikely to be achieved. To meet post-2020 targets, actions that both aim to ‘bend the curve’ of biodiversity loss, and robust metrics tracking the status of wild nature and places are urgently needed. To address these questions, UCL with WWF UK is deploying terrestrial and marine sensors at a landscape-level across four different biomes to understand different species-biome-threat responses (Biome Health Research Project – https://www.biomehealthproject.com
). Using cutting-edge sensor technologies such as camera traps, acoustic sensors, diver-operated stereo-video systems, and tools for 3D coral-mapping, the goal is to develop automated methods for global scale biodiversity monitoring.
We are offering a PhD studentship as part of the Biome Health Research Project, related to the development of machine learning methods needed to analyse large image and acoustic datasets. Current state-of-the-art methods in machine learning for image and acoustic classification are based on supervised deep learning neural networks (DNNs), and require large amounts of carefully annotated training data (e.g. 1, 2). However, these annotations often need expert knowledge to recognise and label different animal species from images (3). As such these methods are both time consuming and expensive to collect. The goal of this project is to investigate and develop next generation methods for training DNNs that are more efficient in how they utilise supervision. This will include methods for self-supervised representation learning (4), utilising additional available metadata (5), and multi-modal learning from paired audio and visual data. The student will be based at UCL’s Centre for Biodiversity and Environmental Research (CBER) and the Department of Computer Science. In addition, the student will work with researchers at the School of Informatics at University of Edinburgh, and will have the opportunity to work collaboratively with other partners in the Biome Health Team which include WWF UK, Imperial College London and the Zoological Society of London.
Candidates should have a first or upper-second class BSc degree in an appropriate subject and a relevant MSc/MSci/MRes qualification, with a background in computer science, machine learning, computer vision, statistics, mathematics or a related field. The candidate will work on state-of-the-art research in machine learning with the intention of publishing in venues such as NeurlPS, ICML, CVPR, ICCV amongst others. Candidates should have strong analytical skills with knowledge of programming languages such as Python, familiarity with deep learning frameworks such as PyTorch or TensorFlow, and should be willing to learn:
• Deep learning
• Machine learning
• Management of large datasets
• Effective communication with scientific and non-scientific audiences.
Start Date: January 2020 (this is flexible)
How to apply
Please send a CV and a cover letter describing your fit to the position and what your research ideas for your PhD as part of the Biome Health Research Project to Prof Kate Jones [email protected]
by 22nd November 2019. Informal enquiries can also be directed to Dr Oisin Mac Aodha [email protected]
 Mac Aodha, Brostow, Jones, et al., Bat Detective - Deep Learning Tools for Bat Acoustic Signal Detection, PLOS Computational Biology 2018
 Van Horn, Mac Aodha, et al., The iNaturalist Species Classification and Detection Dataset, CVPR 2018
 Beery, Van Horn, Mac Aodha, Perona, The iWildCam 2018 Challenge Dataset, FGVC Workshop 2019
 Godard, Clement, Brostow, Unsupervised Monocular Depth Estimation with Left-Right Consistency, CVPR 2017
 Mac Aodha et. al, Presence-Only Geographical Priors for Fine-Grained Image Classification, ICCV 2019