Konstantin Klemmer is a PhD student at the University of Warwick working at the intersection of machine learning and geographic data. He also serves as the Communications Chair for Climate Change AI. We talked about his research and the Climate Change AI organisation.
Climate Change AI (CCAI) is a volunteer run organisation that catalyses impactful work at the intersection of climate change and machine learning by providing education and infrastructure, building a community, and advancing discourse.
We regularly host workshops at big machine learning conferences, providing our community with an academic platform for this discussion. We also run a forum and regular community events like our fortnightly happy hour.
We use our platform and the reach we have, which is quite large by now, to promote ongoing efforts and to facilitate discussion within and across different scientific and professional communities. This also includes reflecting critically on trends in the machine learning community, such as the energy consumption of training large neural network models or excessive tech-solutionism.
To stay in the loop, it is best to join the discussion in our forum, to subscribe to the CCAI newsletter and of course to join our webinars and workshops. We are also excited to help people organize events and initiatives that are aligned with CCAI’s mission, so please reach out if you believe we can help you out! The easiest way to reach us is via email at firstname.lastname@example.org, on Twitter @ClimateChangeAI, or join one of our fortnightly happy hours!
I met a couple of CCAI team members at a conference in Lausanne, in January last year. I then helped with organising a CCAI workshop at ICLR 2020 and joined the organisation soon after. My research is aligned with many of CCAI’s topics and goals. I work on machine learning for geographic data, which is, for example, important for monitoring earth systems or ecological processes. That is a natural research fit. CCAI is an incredibly important initiative and I wanted to help out.
There are over 30 of us in the core team now so most work is split up into smaller committees. For example, I lead the Communications Committee which is in charge of our website, our social media, and any communications efforts we undertake. There’s a Programmes Committee that runs all of our workshops and community events. Then we have a Content Committee that is tasked with creating educational resources and Community Leads, connecting us to relevant communities like economists or the energy sector. These committees meet once every two weeks. We have a general, all-hands meeting once a month. There are also lots of one-to-one meetings to work on particular projects.
CCAI has an advisory board consisting of leaders in machine learning and climate change areas. We meet them twice a year and use their expertise and feedback to help inform our strategic decisions and plans for growing and evolving the CCAI organisation.
It is really important to us to be present at the big machine learning conferences; the most recent one was NeurIPS 2020. At these workshops we have extensive programmes with invited speakers and contributed workshop papers. This way, we seek to provide researchers with an avenue to publish academic work at the intersection of climate change and machine learning, bridging different communities. These workshops are also an opportunity to make the machine learning community aware of how they can contribute their expertise to the fight against climate change.
With our regular webinars, we want to offer a stage to important issues and exciting speakers. The democratisation of knowledge is key to us. Obviously climate change is not just a scientific or technical issue, there are many social dimensions to it, as we tried to highlight with the most recent webinar. We see it as our task to provide free educational resources for anyone who wants to learn more and get involved.
The original CCAI team was mostly based in North America and Europe, consisting of the authors of our foundational paper. When we thought about expanding the organisation, extending our geographic representations was one of our highest priorities. With our latest recruiting, we could welcome more new members joining from Asia and Africa. We want to focus particular efforts on issues that are present in the parts of the world that are often neglected in discussions about climate change. The people who are most affected by climate change are not those in the West, but people in coastal areas and low-income countries. It is essential that we as a climate change community acknowledge that.
Yes, there have been projects where people who met through CCAI have published together. We catalyse that kind of collaboration inside the organisation too, for example through an internal reading group. However, CCAI as an organisation does not conduct research and rather facilitates collaboration. Collaborations and research ideas are also often discussed in the CCAI forum!
We are growing a lot and we have some exciting projects in the pipeline that will expand the scope of what we want to do as an organisation. We want to become an even stronger community builder and a bigger source of support for people who want to do work in this space.
One of the things we are working on is a wiki – like wikipedia but for climate change and machine learning, where everyone in the community can contribute. We are also currently planning a summer school and further educational activities. If you want to stay up-to-date on the latest, check the CCAI newsletter!
I do research on machine learning with geographic data – all data that can be mapped to the Earth in some way, for example, satellite imagery or human location data extracted from cell phones. One particular feature of geographic data is that there is often some dependency based on the geographic context of the data. If you think about the price of your house, the best estimate for the price of it is to look at the price of the house next door. This geographic context can be very complex and noisy. Some machine learning models are well-equipped for modelling these kinds of dependencies, but neural networks can sometimes struggle with that. My research focuses on taking ideas from quantitative geography and plugging them into existing neural network models to see if it improves how they work with geographic data.
I originally come from urban science, studying particularly mobility and transportation. There are many ways to make urban transportation more sustainable which can help to tackle climate change. Machine learning plays a big role here: These applications and the challenges they bring are always a motivation for doing more methodological work. When I started my PhD I focused much more on applications, but I have drifted more and more in the methodological direction over the last years.
My urban science doctoral training centre at the University of Warwick is very interdisciplinary: some of my colleagues study cities from the computational side, and some take a more philosophical or sociological angle. It is also a very international program: I have a long-standing collaboration with New York University (NYU). I love interdisciplinary work, and learning from people with different backgrounds than my own. I was actually due to spend all of 2020 in New York but unfortunately it got cut short by the COVID situation.
I research ways to best run shared mobility systems, such as shared bikes, cars and scooters. Machine learning plays an integral part in making these schemes profitable for the companies or public sector agencies that run them. There is an elaborate optimization process behind where vehicles are located. With electric vehicles you have to add another dimension to the optimization problem, which is the charging of these vehicles. You can’t just start one trip immediately after the previous one because you might have to recharge the battery. There are also imbalances in trip destinations: In the morning, for example, people all want to pick up a bike at Waterloo station and do their last-mile commute to work. Then in the afternoon they all want to ride back from work to Waterloo. So, there is a very uneven distribution of these vehicles throughout the day. These are all problems that machine learning can help with – for example, predicting this kind of demand beforehand, or coming up with good relocation strategies. I’ve been working on this for six or seven years now. and the field is still evolving very fast with lots of interesting ideas and projects going around. Interestingly, traditional approaches in optimisation and statistical modelling are very close in performance to the more “hyped” machine learning methods. It’s humbling for the field to see that these “old-fashioned” techniques often perform equally well. Not everything is improved by a large neural network.
Another piece of research I have been involved in concerned the economic effects of bike-sharing infrastructure. We found that putting bike-sharing infrastructure into areas in London actually increased the number of local businesses. Controlling for many different socio-economic factors, we found that the expansion of the local cycle network made it more attractive to open a local business. This finding is very important, as it gives a strong, science-backed argument to cycling advocates, showing that we can be sustainable and still support business. This particular research was enabled by the London Datastore – lots of cities now have these vast open data portals that help us to gain insight into how cities work and the underlying processes. Put together with the advances in scalable machine learning methods, this makes an ideal playground for applied machine learning researchers and practitioners.
Finishing my PhD is the main priority. I would like to continue working in this area, at the intersection of machine learning and geographic data. I want to do research, whether that’s in industry or academia. Luckily, many tech companies have started initiatives to do research in machine learning and sustainability in-house, so this would be another option.
I think the best piece of advice is that they should be proactive with seeking collaborators. If you like someone’s research just drop them an email and pitch them your ideas. The worst thing that’s going to happen is that they don’t reply, or they say no. That happened to me a couple of times. But, on other occasions it worked out and I ended up with some really cool collaborations. I think being proactive early on in a PhD is the best thing you can do.
Konstantin is a PhD student in Computer Science at the University of Warwick and New York University. He was an Enrichment student at the Alan Turing Institute. Konstantin’s research focuses on the representation of geographic data and phenomena in machine learning methods. Beyond that, he is interested in the application of these methods in urban environments, tackling issues in crime, transportation or climate change. Konstantin holds a Bachelors degree in Economics from the University of Freiburg (Germany) and a Masters in Transportation from Imperial College and University College London. Konstantin is also a volunteer member of Climate Change AI, a global organization dedicated to helping to tackle the challenges of climate change using machine learning.