General Profile: 

I recently graduated with an MSc in Computer Science from the University of Bristol. I am interested in how AI can be applied to the conservaon of endangered wildlife and improve the welfare of animals that exist in capvity. So far, my research has focussed on animal-computer interacon (with the Gorilla Game Lab, hps://gorillagamelab.com) and developing deep learning AI systems to automate the detecon and idenficaon of great apes in zoo environments (here is a link to our most recent work, hps://arxiv.org/abs/2012.04689). Beyond this, I am looking forward to exploring how deep learning can be augmented with methods from tradional machine learning and informaon theory to answer more complex quesons on individual and group behaviour and populaon dynamics. More widely, I am interested in the role AI systems will play in shaping the future of society and what can be done to stop exisng inequalies from being abusively reinforced. I hope to get involved in any project which has this goal in mind.

Research Project Summary:

The analysis of wildlife presence, abundance, distribution, and behaviour is growing increasingly important. As the climate crisis gathers pace, the existence of many endangered species grows ever more perilous. Consequently, the need for methods allowing this analysis to be carried out quickly, at scale and with greater rigour is growing more urgent. In particular, it is recognised that understanding the behaviour of endangered species will be critical in devising solutions to the current biodiversity crises. There are several databases that have accumulated footage of endangered wildlife species over long periods of time. The analysis of this data will be invaluable in providing insights which can help conservation efforts. At present, the analysis of high volumes of data is limited by the time and resource of human expertise. AI systems on the other hand have the potential to automate and expedite this process, allowing large scale analysis to be performed at speed.

The complex behaviours of great apes (e.g. infant handling, coordinated movement, tool use, teacher learner activities) are easily recognised by domain experts, whereas the ability of machine learning techniques to detect these behaviour has been largely unexplored. This project aimed to investigate the application of deep metric learning techniques for classification of great ape behaviours. Although this has been performed for classification of individual animals, it has not yet been investigated for behaviour recognition. Deep metric learning seeks to train a model to embed inputs onto a metric or latent space. A loss function, such as contrastive or triplet loss, is used to learn a metric space where semantically similar representations are close to one another and dissimilar representations are distant. It is especially attractive for interactive systems because it is possible to visualise what is being learned. That is, embeddings in the learned metric space can be visualised, making the system more interpretable. This provides a good opportunity for experts from other scientific domains, such as ecology or zoology, to engage with the system. In this project we implemented a two-stream model, comprising convolutional and recurrent components, with a triplet loss function. Three triplet mining strategies were also implemented; random, semi-hard and hard negative mining. The model was trained on an annotated subset of the Pan-African dataset; a dataset comprising camera trap footage of apes displaying a range of different behaviours. The best performing model, trained with random triplet mining, achieved approx. 64% and 81.6% top-1 and top-3 accuracy, respectively. Additionally, we performed a visuale analysis of the learned metric space using the high-dimensional data visualisation technique, t-SNE. This revealed interesting patterns into the way in which behaviour embeddings cluster in the metric space.

A vast amount of knowledge has been amassed by humans in the field of conservational biology, primatology and ecology.  Therefore, it is vitally important that this knowledge is continually integrated into AI-based systems. With this mind, the project investigated the capacity of a human-in-the-loop (HITL) system, based on active learning techniques, to draw upon human intelligence for system optimisation. More specifically, a prototype human-in-the-loop system was developed. The system employs active learning to sample informative, unlabelled data for human annotation. However, in this project, the human interaction was simulated by partitioning the existing training set. An additional holdout set functioned as a pool of unlabelled data and the existing annotations allowed the role of a human annotator to be simulated. Two sampling techniques were then implemented; uncertainty and diversity sampling. Both techniques were applied with a range of different labelling budgets. Ultimately, it was found that uncertainty sampling is effective with a small labelling budget. Conversely, diversity sampling improves as labelling budgets increase but random sampling becomes equally effective.