General Profile:

My background lies in pure mathematics and computational statistics, both of which equipped me with very valuable knowledge ahead of joining the CDT. Despite briefly working as a government statistician and spending time involved in various maths outreach projects, my MSc year helped me decide that I wanted to pursue a research route that would expose me to a wide range of new ideas in computer science, whilst keeping my options open in regard to staying in academia or moving into industry. During this time, my summer project was centred around Markov chain analysis of text to aid in determining the authorship of tweets, with the overall aim of combatting fake news. Although my enjoyment of learning modern foreign languages contributed to an interest in natural language processing, environmental applications of AI also really appealed to me, and the potential for making a difference to the animal world felt like it would give me the greatest sense of fulfilment. As a result, I settled on my current area of research: ecological computer vision. Outside of the office, I like swing dancing to jazz music, roller skating and swimming, and I’m currently also learning British Sign Language.

Research Project Summary:

My research springboards off the back of the success of Microsoft’s MegaDetector and its ability to accurately pick out living beings (such as mammals, birds and reptiles) within camera-trap images, by flagging them up as belonging to the general class of “animal”, regardless of their species. This speeds up the pre-processing pipeline of ecological computer vision tasks by cutting out a significant amount of human effort that would otherwise be required to sift empty images out of the dataset before use.

Although beneficial as an initial step, in order to draw meaningful conclusions from data for specialist tasks in practice, it is typically undesirable for all animals to fall under a single category. Instead, it might be necessary to focus solely on endangered or invasive species in a given ecosystem, in which case, a classifier is needed rather than just a detector. This, however, is where problems may arise: what is the best way to proceed if we cannot directly fine-tune a model for our animal of interest due to a lack of applicable training data?

The aim of my project is to demonstrate that what a model has learned from image data depicting animals can be transferred between visually similar species, such that one can stand in as a proxy for the other when there is little data available for the animal of interest. The premise is that utilising a substitute species from a geographically distant ecosystem will boost the performance of an animal classifier on datasets that feature, for example, endangered or invasive species in unfamiliar habitats that we care about, without getting confused between the animal of interest and the proxy, thanks to camera-trap location information.

This method takes advantage of cryptic biodiversity – specifically, pairs of species that are hard to distinguish between in the context of poor-quality (i.e., blurry and dark) images through the lens of a machine – and makes use of an abundant species (of which, images should already be readily available in large, pre-labelled repositories like the Labelled Information Library of Alexandria: Biology and Conservation) to support another for which we only have limited training data.