Matthew Wilkinson is a final-year PhD student co-funded by the CSCT and CMAC working with Dr Bernardo Castro Dominguez, Dr Uriel Martinez Hernandez and Prof Chick Wilson. Matthew’s research aims to apply Artificial Intelligence to challenges associated with pharmaceutical manufacturing, leading on an overall more efficient and sustainable manufacturing process.
In the spring of 2021, in partnership with the Alan Turing Institute and their industrial placement network, Matthew had the opportunity to take part in a 4-month placement at TRL Software Ltd. Here, as illustrated by the company tagline, Matthew’s work focused on “delivering transport science through software”.
Why did I want a placement?
Having read the above paragraphs, I’m sure questions will arise about how I went from working on chemical problems to road transport. Sure, it is not an obvious transition, but nevertheless, it happened, and here’s why.
Casting my mind back to where my placement hunt began, it was clear to me that I wanted to do something that would contribute significantly to my personal development. The chance to work in industry was the perfect challenge. I could apply all I had learnt over the previous 3 years, face challenges well beyond those in chemical applications and experience how AI truly impacts the real world. Facing and learning from these new challenges to become a more well-rounded machine learning practitioner was the single most important factor in my decision to take on a placement.
Who are TRL Software and what do they do?
Transport Research Labs (TRL) Software are somewhat of a unique company. To quote their own webpage, “independent from government, industry and academia, TRL helps organisations create global transport systems that are safe, clean, affordable, liveable and efficient”.
In my experience, what this really means is that their primary purpose is to make a difference by effecting change in the problems they face. From saving commuters from being stuck in traffic to ensuring pedestrians have enough time to safely cross the road, the application and impact of the software products TRL provide will be felt by most of us each and every day.
Making AI less scary...
Before I explain my project, I want to take the time to try and make the AI buzzword a little less intimidating. When we apply AI to problems, our goal is to learn from historical data so that we can make informed predictions going forward. To do this, we can choose from a variety of machine learning models, all of which are algorithms designed to find patterns in data. In my project, we used a machine learning (deep learning, technically) model called a neural network.
Neural networks are mathematical models that take inspiration from our brain. As such, we have to train them so that they learn how to predict our target of interest accurately. To put it another way, imagine a child attending school on their first day. If we ask this child to sit a maths test, they will probably do poorly. However, if we ask this child to learn from lots of experience in the form of many maths lessons, then they can do much better on that test.
The neural network works the same way - we give it lots of experience in the form of data, and it learns to predict a particular value from this.
My project at TRL:
The project was a collaborative effort between me and Shreshth Tuli, another PhD student from Imperial College London. During my time on placement, our project aimed to identify incidents on a particular road traffic network in Greater Manchester.
'Incidents' is a fairly loose term, but here I will define it as anything that causes disruption to the road and makes changes to the normal traffic flow.
Currently, road traffic incidents are identified manually, which results in missed incidents and has the potential to reduce response time – if the operator doesn’t see the problem, they can't start trying to solve it. The project focused on using a neural network to identify these incidents, allowing for faster, more accurate detection, which in turn leads to quicker response measures and overall less disruption to road users.
As you can imagine, road traffic data can be very noisy, with things like rush hours, sporting events, bank holidays or even the COVID lockdown all making it very hard to see what normal conditions should look like on a given day.
[If you aren’t interested in the technical detail then you can skip this next paragraph]
Road networks present a perfect example of spacio-temporal data. A road network can be represented as a mathematical graph, with nodes and edges representing the junctions and the roads connecting them. Having reviewed the literature, most previous attempts focused heavily on either the spatial (how the road layout affects traffic flow) or temporal (how day of the week / time of day affects flow) aspect of the problem, with none fully accounting for both. Hence, we designed a novel neural network architecture that used graph attention and transformers to capture both aspects and intelligently combined them to help make the predictions. We used data from Transport for Greater Manchester and open-source traffic data to prove that our network was state-of-the-art in this task.
Having completed the project, we have a few publication goals, but at the time of writing this blog post I’m very excited that we have been accepted to present our work at the IJCAI 2022 - Workshop on AI for Time Series Analysis.
Having completed the placement, I can’t speak highly enough of it.
Working with Shreshth and Chris Kettell (our boss) was fantastic, and I would like to express my sincere gratitude to them both for their support, enthusiasm and sharing their knowledge with me over the 4 months. I must admit, managing a half-time placement and full-time PhD simultaneously did not come without challenges (or its share of late nights), but looking back, it was excellent for developing my ability to manage big projects in parallel. Although yes, there were lots of hours, it never felt like a burden as the project was so exciting that often I would forget it was technically work.
I must admit, I have been pleasantly surprised by how much of my new experience can now carry back to my PhD project as my attention falls back to thesis writing and concluding my PhD.