How long have you worked for OS?
I started back in 2001, so, 18 years. After I finished my PhD and a couple of post-docs, I chose to spend some time cycling around The Netherlands and eating far too many stroopwaffels. Coming back and starting at OS with its core hours and line management was a bit of a shock, but I clearly got used to it.
How long have you been in your current role?
To be honest, I’m not actually sure. In some ways I’ve been doing this job for over decade, in others I’ve only been doing it a year. The title, responsibilities and department change often enough to keep me on my toes.
Can you describe your working day?
Daily meetings with my team are something I look forward to as they usually wow me with all the amazing stuff they’ve achieved against a whole host of technological odds. From here the days can vary drastically. I’d like to say that I spin up a multi-GPU device and train a deep model for finding subterranean power lines but, typically, I read a paper or report, deal with line management responsibilities and turn up to meetings. We’re trying to develop a DevOps approach to research, so I’ve been spending a fair amount of time recently ‘refining backlogs’ and trying to work out where we need to devise metrics. I’m sure it’ll be really powerful when we’re fully operational, but at the moment it bends my brain a fair bit. I also try and work a couple of days a month at University of Southampton, where I’m a Visiting Researcher. Although I feel completely out of my depth, I enjoy the fact that, every time I visit, I learn something very useful and unexpected.
What are you working on right now?
We’ve set an extremely ambitious goal of using deep learning (a type of machine learning) to process our sensor data in a way that will achieve the dual outcomes of (a) rapidly responding to customer demands and (b) being able to create novel new ways of representing a landscape. When I describe what we’re trying to do the feeling is a bit like vertigo, it’s disorientating trying to achieve so much! However, there’s little advantage in taking the well-trodden path and by working towards this goal, we’ve developed strong knowledge and skills in both the theory and practise of training deep networks. This month, we spent a 2-week sprint investigating approaches to porting our learned models into a scalable production pipeline. This means that our models which have been tested on only a small region of data will be scaled up to the entire country.
What is your favourite part of your job?
I’m only happy if I’m learning new things and, in this role, I have a license to step away from the desk to investigate new topics. The role prevents my agitation at the prospect of getting in to a routine. I also love that I work with people who teach me things and challenge me, and I enjoy the moment when two very separate concepts meet in a new idea. However, my colleagues are definitely the best part of my job – they’re all funny, a bit nutty and very bright!
What is your OS highlight?
A while ago after a year moonlighting on Massive Open Online Courses (MOOCs), I proposed a deep learning research collaboration at the University of Southampton. Over the course of a year I spent half the week based at the University working alongside the postdoc that we had hired to undertake the project. It was like a practical training course and while it had a very steep learning curve, it provided me with a solid foundation for our current work. As a result, I gained Chartered Scientist status and thereafter filed a patent for this idea.
What are you excited to work on (or continue working on) in the future?
We filed another patent last year relating to methods of classifying our topographic data using local context (understanding things in the real world using what is near them). A simplistic example is that a region of small buildings surrounded by small parcels of land could be labelled as a specific type region of housing estate. If you consider the hierarchies (e.g. how single buildings form the context of a housing estates which form the context of suburbia, then cities etc.), it could be possible to extract extremely fine detail about the real world that is already implicitly contained in our data. A 3-day study indicated that different eras of housing estate can be identified using very simple methods all from the data that we already have. I want to spend more time developing automatic approaches to extracting these details and I’m particularly excited by graph-based approaches.
Additionally, I’m keen to develop methods that bring our surveyors’ expertise into the automatic approaches that we are working on. There shouldn’t be two separate streams of activity (manual and automatic), because we’ll always need expert humans to interpret the world in different ways. This is about creating systems rather than just algorithms and is a long-term ambition for me.
Find out more about career opportunities at OS on our website.