David Rolnick on The Robot Brains Season 2 Episode 3
Transcript edited for clarity.
Pieter Abbeel: While the world's temperature rises, there are scores of scientists working around the globe to study causes and solutions. One scientist in particular has stood out for his passionate lobbying for the application of machine learning to better understand climate change and potential solutions. This scientist has been successful in building a broader movement, including others like Andrew Yang and Yoshua Bengio. to champion the amazing possibilities that exist at the intersection of AI and climate. He organized the first ever AI event at the United Nations Climate Change Conference. He's been named a top innovator by MIT Technology Review, all before the age of 30. I am, of course, talking about today's guest, David Rolnick. David got his PhD at MIT and is currently a professor at McGill University.
David, welcome to the show. We're so excited to have you here with us.
David Rolnick: Thank you so much for having me. It's a pleasure to be here. And thank you so much for the kind introduction.
Pieter Abbeel: Well, I’m so happy to have you on and get to chat with you about one of the most important things that people are working on in AI - solving climate change. Now, I imagine, you didn't from your very earliest days worry about climate change and AI. I'm curious, where did you grow up? And how did that lead to your interest in AI and climate change?
David Rolnick: Yeah, growing up, I was interested in a lot of different fields. That hasn't really changed. And I spent a lot of time on math. I also loved natural history, entomology and ornithology. I spent a lot of time looking at birds and insects. And then in college and grad school, I worked on math that gradually became more and more connected to deep learning, deep learning theory, connections of deep learning with computational neuroscience. But I really felt like I should be working directly on climate change. And in some sense, what I am doing now has come full circle with some of the projects that my group works on. We’re actually focused on applications of machine learning and AI, to other things like biodiversity, birds, and insects, some of the things that I was first interested in, as well as applications and electricity, grid optimization, materials discovery for green chemistry, and a host of other applications associated with climate change mitigation and adaptation.
Pieter Abbeel: Quite the range of exciting AI directions we could do research on to help with fighting climate change. I think it's fair to say most people weren't really aware of all those opportunities. In 2019, you published a paper tackling climate change with machine learning and your co-authors included Yoshua Bengio, Demis Hassabis, as well as many others. I'm curious, you got interested in AI and climate change, and then you bring together this large group of highly established researchers for a single machine learning paper. How did you make that happen?
David Rolnick: I was surprised a few years ago to see that climate change wasn't being talked about as a flagship opportunity for machine learning. The initial goal of the group that I put together was to bring together expertise from many different fields in machine learning, but also in complementary areas like energy policy and land use to provide a call to arms for the machine learning community to describe the opportunities that existed along with showcasing the terrific work that had already been done in this space, including within movements like climate informatics and computational sustainability. So that initial group became the climate change and AI initiative which I lead together with Priya Donti and Lynn H. Kaack in addition to my group's own research in the area of AI and climate change. So the mission of Climate Change AI is to catalyze impactful work at the intersection of climate change and machine learning, which involves a global network of experts and stakeholders interested in this area.
We also run the Climate Change workshop series at NeurIPS, ICML and ICLR. We launched that a couple of years ago and also resources like reports and tutorials and webinars. We also recently launched a grants program to find work in this area, and catalyze the creation of new datasets.
Pieter Abbeel: Oh, wow. Now, when you say you launched a new grants program that means you have to get some funding agency really excited about this idea.
David Rolnick: Yes, it’s huge thanks to Schmidt Futures and Quadrature Climate Foundation (QCF) for providing funding for this round.
Pieter Abbeel: Now in your paper, you actually highlight several directions that people could work on that definitely surprised me. Could you maybe describe some of those directions?
David Rolnick: Absolutely. So the paper was intended to provide an overview of many different areas and opportunities for machine learning within all the different fields that touched climate change because there's so many aspects of climate change. Mitigation means reducing greenhouse gas emissions while adaptation means responding to the effects of climate change. And then also climate science is studying the climate itself. So we looked at all of those along with cross-cutting tools that could be relevant in lots of different lots of different areas.
The overall themes for impact for machine learning and AI, were in distilling large unstructured datasets into useful information. For example, to guide policy and contexts like pinpointing deforestation or understanding built infrastructure and energy use associated with buildings or highlighting where in corporate financial disclosures (which are large corpora of text) there is climate relevant information. So really taking some massive amount of unstructured data and creating useful insights that could be used to guide policy or further shape decisions.
We also saw a lot of applications of machine learning in optimizing complicated systems. So everything from reducing the energy needed to heat and cool buildings, optimizing industrial processes, freight transport – there are a lot of different applications there. In forecasting, we saw applications in forecasting supply and demand for electricity minute scale level, which is really important if one's going to integrate further renewables into the grid and reduce the amount of surplus power that's being produced. Also, forecasting agricultural yields in the event of climate-induced agricultural disruption, and many other applications.
And then we finally saw a number of applications in scientific modeling and discovery like accelerated science where one can often speed up the process of experimentation by suggesting new experiments that should be tried or new experimental parameters like the discovery of better battery materials or materials for solar perovskites and photovoltaics, for example. But then also in speeding up simulations in contexts like atmospheric physics or aerodynamics of vehicles and other settings where there are very complex physical equations, which we understand extremely well, we're not going to come up with a better model of the Earth's climate using machine learning, but we can speed up pieces of it. And it's not that our current understanding of the climate needs machine learning but it does need some speeding up sometimes to run some of these very time intensive simulations. So that's a good application of machine learning as well.
Pieter Abbeel: So we have these good computational models but machine learning could speed up the rate at which we run, those models get results back. What's the intuition behind that? How does that happen?
David Rolnick: Yeah, so some of the biggest computers are being used right now to run physics and simulations of things. For example, climate and weather, they require just massive amounts of computational resources to solve these differential equations that come in lots and lots of different variables. For example, something that my group works on is fast approximations to radiative transfer, which is a particular physics primitive, which is being used in global climate and weather models. And if we can solve that faster, then we can run these models at much higher resolution, both temporal and spatial resolution, which means that they're much more useful to people, right?
You know, if you're a small municipality or region trying to plan for climate change, you want to understand the effects of climate change on you right there right now, rather than some enormous grid cell that’s hundreds of kilometers on a side, which is typically what happens with climate models. Climate change is fundamentally a local issue as well as a global issue. It affects different places very, very differently. So that kind of simulation, speeding that up and making it more feasible to run the physics but with certain approximations being made, is really an opportunity for machine learning.
Pieter Abbeel: So should I imagine it in a way where you first run the very expensive calculation several times, and then you train a neural network that somehow is more compact than running the full simulation, but ends up with the same prediction in maybe a day. I don't know what timescales this works out. But the same prediction, while using a lot less computation.
David Rolnick: That's absolutely the goal. One can imagine supervised learning for this one can also imagine building in various domain constraints from the start before knowing the physical equations that govern the governance system. And that is obviously an area of cutting-edge research in machine learning – building in physics constraints. And overall, we've just seen a lot of areas where machine learning boundaries and methodological boundaries can be pushed by climate relevant problems. So whether that's in hybrid physical models, incorporating physics based constraints, or transfer learning and mental learning in uncertainty quantification, there are a lot of areas where there really is a need in these domain relevant problems for AI and machine learning to push the envelope.
Pieter Abbeel: That's interesting. I'm imagining here, something where, and I think I saw a paper yesterday that incorporates constraints, hard constraints in neural network predictions where you can put a constraint on properties that prediction should satisfy, as I'm imagining, maybe the production has to satisfy conservation of energy. Is that a good way to think of it?
David Rolnick: Absolutely. It will depend upon the particular situation, what constraints there are that are relevant. So conservation laws might be relevant in an atmospheric physics situation. But in the setting of the electrical grid, maybe the constraints are that power has to actually obey the laws of electricity. So that was another paper that we put out that was focused on AC optimal power flow, where you have constraints that are a big, non convex quadratic constraint, which you need to satisfy in order to abide by the laws of electricity. And if you are trying to route power on the electrical grid, and you don't do that, then the lights go out. So obviously, you need to satisfy that constraint, you need to guarantee that you can satisfy that constraint. So that particular paper was focused on guaranteeing the satisfaction of hard constraints, even in the context of deep learning based approximations.
Pieter Abbeel: And deep learning approximations are usually very, fairly blocky. It’s very hard to ensure something is actually true. So that's great.
Now, when I think about climate change, I imagine many other people will be similar. It's a big problem. But it feels like it's hard to actually make change. It's like, I'm just one person, how am I going to make change? And I'm kind of curious, what is your opinion on that? Because clearly you're dedicating your career to this and how you can also maybe encourage others to be more optimistic about the ability to to make change on their own and in a kind of big, impactful way.
David Rolnick: So there are lots of ways to have an impact on climate change and AI is not used in isolation when it is being used. One of the most important things to remember is that it's a powerful tool. But obviously, it's not a silver bullet. And we're not going to magically solve climate change with AI, machine learning or anything else. Machine learning is useful, and it's well matched to existing bottlenecks in policy, energy, land use, or other areas.
And change happens when different people come together. Collaboration is essential. Experts in machine learning and relevant application areas and also stakeholders who will be using are affected by technologies. This is really essential to avoid pitfalls and ensure a pathway to meaningful impact. As with other areas of applied machine learning, domain specific knowledge is generally really essential. There are always constraints and contextual information that aren't just captured in the data. And it's really important to consider also how something's going to be used to build in any deployment considerations right from the start. More broadly, one doesn't have to be working in AI or machine learning to make a difference in climate change. It's not the most powerful tool that we have. I'm working on climate change as an AI practitioner, because that is the tool that I have, and it's a powerful one. But I'm not here to say that people should be working, if they aren't computer scientists on AI, go work on policy, go work on energy go come on any of any number of areas, we need all hands on deck for climate change.
Pieter Abbeel: Yeah. Now, of all things you're currently working on, one of your projects is about identifying butterflies. Why is that?
David Rolnick: So I told you I liked insects, right? It turns out that ecological monitoring is a massive opportunity for AI and machine learning. And something where there are enormous implications both for sustainability more broadly and specifically in climate change. So gathering data on ecosystems and how ecosystems are changing is something that is currently done by people which in some sense has to be done by individuals who are out in the field.
But there are all kinds of ways to help that system and contribute to scientific knowledge. And that particular paper that you referenced is focused on providing tools for citizen scientists, who are not always experts to actually be out in the field collecting information about individual species. We also work on automated systems, so sensors that you can put out in the field that will attract insects, in this case, moths. Fun fact, 1/10 of all species of anything are moths, including plants, including bacteria. So if you're studying moths, you're getting a very good idea of how healthy the ecosystem is. So this system attracts moths and automatically identifies them using computer vision. It's a very hard computer vision problem, because it's really, really fine grained, all those species of moths. So this is one of the sets of projects that we work on in biodiversity monitoring, to really have a sense of how ecosystems are changing. Also, there are a lot of CO benefits, like understanding pollinator distributions and understanding how particular species that people may especially care about are being affected by climate change and other human interventions.
Pieter Abbeel: That's so interesting. So you're getting an early signal there. If something's happening in a certain region, you would know likely before others by seeing a change in diversity of the moth populations, is that right?
David Rolnick: Yeah, there really isn't nearly enough data and there aren't enough people out there with relevant expertise to gather all the information that we need on how the planets are changing, so augmenting scientific experts and amateur scientists with technology, having all these pieces working together is really important.
Pieter Abbeel: So David, you recently organized an event on AI at the United Nations Climate Change Conference. What's the global perspective that you experienced there on the intersection of AI and climate change?
David Rolnick: So, for climate change, I organized events at the last UN Climate Change Conference, which is called the COP, the Conference of Parties. That was two years ago. And we're also organizing several events this year along with working on a number of different global policy related initiatives. So for example, we authored a report that will hopefully be released soon focused on shaping international and national policies related to the intersection of AI and climate change, and how governments and policymakers can better facilitate meaningful work at this intersection. Now, there are a lot of challenges associated with global action at this intersection. And there are obviously innumerable priorities for global policy on climate change, which do not all relate to AI. Some of the opportunities and challenges for policymakers are in ensuring meaningful pathways to deployment in terms of capacity building within existing institutions, in terms of facilitating integration into slow moving industries that often are resistant to change. And there are a lot of considerations to bear in mind with respect to global adoption of different technologies, including a huge number of equity considerations.
So who is empowered to build solutions? What public problems are being prioritized and how these problems are being worked on? And obviously empowering a global set of stakeholders to shape the intersection of machine learning and climate change is essential to ensuring that the technologies are owned by the people that are affected by them? And don't reinforce existing power imbalances across countries and institutions. Related to the question of who is the question of what's being worked on? Since often problem priorities are going to reflect inequities that already exist within technology.
And those can also involve geographic boundaries and considerations. So for example, I'm seeing a lot of interest in ML for fighting wildfires, which is obviously a problem that's particularly relevant here in North America, in Europe, and Australia. And that often receives more attention and fundings and machine learning for fighting locusts, which is a problem that's particularly relevant in East Africa and the Middle East and in India, and both of these problems are extremely important and both are being exacerbated by climate change. And I would like to see more attention given to both of these problems.
But definitely there are some of these inequities and whose framing problems and what occurs to the people who are framing the problems. And then also finally, how projects are being worked on is important and women's considering global implementation of, of and climate strategies, data and balances between regions or between communities and within a single region. can mean that machine learning isn't applicable to the entire population, it's only applicable to some subset, or that the algorithms are most effective within the data rich regions. So really, ideally, AI for climate would serve to improve equity. But this active work was set at a high level, including policy and at a low level, including project management.
Pieter Abbeel: Now, I'm kind of curious as you think about these problems, you're a machine learning researcher, originally, at least by training, and from there look at AI. To what extent do you feel like you want to stay in machine learning and really focus on that versus spending, you know, more and more of your time possibly on the non machine learning aspects to effect change?
David Rolnick: I think that when one's working on a problem that is driven by societal impact, one needs to start out with the simplest possible methodology, the simplest possible technologies. But one can pick problems where one thinks that those won't be enough, and that one will need more sophisticated tools. So that's what we try to do in my group, I try to pick problems, where I think that we will need significant innovation. But we still start out with the most basic possible techniques, because the most basic techniques are always the best if you can use them, but in some cases, you will need really groundbreaking technologies.
Pieter Abbeel: Now, as we're reflecting here on the research a bit more, I'm kind of curious. What do you think are some of the maybe easiest starting points for a machine learning researcher? What are some great data sets or simulators for people to look at it just kind of dive in and try to make something happen?
David Rolnick: We try to provide a lot of resources for people who are getting started at the Climate Change AI websites, I encourage you to check that out. I would say that I would caution against just diving into a data set, except as a sort of way to get your feet wet. Because most data sets that are really well structured, are ones that have in some sense, been picked clean. Most problems that are most meaningful, require interfacing with people in a relevant community, understanding what the data means, potentially working to structure new data or work with stakeholders who haven't necessarily made the data available yet. So in many situations, the most impactful areas of exploration don't involve sort of a single percent improvement on an existing data set. But they would involve integrating existing data with domain knowledge and with other considerations, which are sort of not nicely packaged in the way that we're used to in the context of pure machine learning. But I highly recommend getting your feet wet exploring some of the many, many wonderful datasets out there. And there are a lot of things to do just by looking at public data that has just not been explored. well enough.
Pieter Abbeel: Now that really resonates. You want to have an impact. You want to be engaged with the stakeholders. I couldn't agree more. On the flip side, I'm thinking for most, let's say, PhD students who are, you know, looking to get their feet wet and working on something that probably feels like an almost insurmountable bear? I'm not saying it has to be, but it probably feels like that. And I'm curious. Now, what do you recommend?
David Rolnick: Just reading a lot and trying things. But ultimately, I think that my advice for students is to learn about a specific area and become really well-versed in that area, become an expert in machine learning and power sims or ml and infrastructure. We really need people at the intersection of fields, and there aren't nearly enough who can innovate in machine learning, but really understand the technical language in another area. That's holding back both machine learning and those other areas. People who can speak different languages and cross boundaries, can understand both the real opportunities for impact in an area and the potential mechanisms for solving them. There are so many different particular problems in areas that one can focus on. And I encourage you to consider delving into one because that will yield really significant benefits for society and also for machine learning.
Pieter Abbeel: So David, of course, at the risk of leaving some important ones out, could you list maybe some of those areas that come to your mind?
David Rolnick: Absolutely. So at a very high level, there are fields like electricity systems, and transportation and heavy industry, buildings and cities and urban planning, disaster response, all kinds of applications and societal adaptation like understanding biodiversity and ecosystem change. Climate Science, which is again, the science of the climate itself, and related earth systems. So these are overall headings. But within those, there are so many different avenues for impact. I encourage you to check out the tackling climate change machine learning paper, because we really did try to provide a taxonomy of particularly high leverage areas for work, but there are dozens of them. So if you're interested in that, go check out the paper or the climate change AI website where we have interactive summaries where you can actually search by particular machine learning techniques. You can say, I am an expert in computer vision, or I am an expert in NLP. And I'd like to see where my areas of expertise can be most relevant.
Pieter Abbeel: That sounds absolutely wonderful. I think a lot of people love to leverage their existing expertise and find a way to help out. Now, there's another angle to climate and machine learning, which is also often talked about which is in the somewhat other direction. Machine learning, in some sense, pushes the amount of power that we want to generate, and hence actually, possibly accelerates climate change inadvertently. I'm curious about your thoughts on that. Is that something that's already happening that we should be worried about? What are the concerns there?
David Rolnick: Yeah, some machine learning models definitely use a lot of energy to train and run. In particular, very large models that are used, for example, and LP and in various corporate applications. This is definitely something the field should consider as one factor when deciding what models to use and also how important small incremental improvements over state of the art really are. So it's in some sense, a cultural question that we can ask ourselves. However, I also feel that the discussion of energy use can divert attention from another problem, which is the way that many applications of machine learning can have negative consequences for climate change. It's not just in energy use. For example, machine learning is being used extensively to accelerate oil and gas exploration.
The World Economic Forum has estimated that the oil and gas industry is going to make $425 billion more by 2025, thanks to machine learning and advanced analytics. So that is definitively making climate change worse to be accelerating oil and gas discovery and extraction. Machine learning is also being used very widely in advertising systems, which almost definitely increased society's consumption of resources. And these kinds of applications of machine learning are definitely having a significant effect on climate change. It's just harder to measure.
So In these cases, the energy use of the algorithms themselves is clearly a drop in the bucket. We don't have the equivalent of humanely manufactured machine guns, right? The thing you're doing is hurting the climate, but you're using a little bit less energy is not really relevant in those cases. On that note, it's really important to remember that the implicit choices that we are making as technologists have the potential to change the impact of a new technology. For example, self-driving cars may make climate change worse but they have other benefits like safety, which are extremely significant. But from a climate change perspective, if we design autonomous vehicles with a focus on personal cars, then driving will become easier, people will probably drive more and global carbon emissions may increase from that even if each mile driven becomes a little bit more efficient.
On the other hand, developing AV technology with a view to vehicle sharing and public transportation could decrease carbon emissions. Now, I'm not here to bash AVs, this is just an illustration that all of us are doing work that is affecting climate change. And the choices that we're making, both implicitly and explicitly, are meaningful. And AI for good isn't about adding on new societal beneficial applications until the business as usual. It really needs a lot of subtle choices, to better align our work with societal goals in lots of different ways.
Pieter Abbeel: Now, David, I'm kind of curious when you take a step back, and you think about the future of our climate? How optimistic are you?
David Rolnick: It's a great question, and one that I get asked a lot. Um, I guess two things. First of all, since I started working more directly on climate change, I am not as stressed about climate change anymore, I'm stressed about other things. I'm stressed about other societal problems.
So something to bear in mind, for anybody who's really stressed about something, if you work on that thing, you start being stressed about the other things, but not about that thing. Because you're doing your bit if you're working more directly on it. But also climate change isn't an on/off switch. If I had to pick, are we going to solve climate change? Are we? Are we doomed or not? Then yes, I would say yes, climate change is already killing many people. And it's already going to kill many more, regardless of what we do. Now we are locked into some of the effects. But we have the decision to make how bad it's going to be. It's not an on/off switch. It depends on the choices that we make now, just how terrible the effects are. And everything that we can do now is going to have an impact. So in that sense, I'm optimistic that we can, and every one of us can have an impact on climate change.
Pieter Abbeel: It's interesting, because what I'm hearing is you're saying, yes, things are gonna be bad. But actually, we can make it less bad.
David Rolnick: We can make it less bad. We can't make it good. We can make it less bad.
Pieter Abbeel: And you're calling that being optimistic. That's optimistic about our ability to do something, I guess, to have some influence but not optimistic about being able to, you know, truly solve it.
David Rolnick: Yeah, I mean, humanity has already seriously impacted the world in many ways that are not irreversible. Most tropical coral reefs are, for example, doomed to disappear regardless of what we do now. Because we've already affected the climate in various irreversible ways, various kinds of ecosystem destruction. Similarly, that's not necessarily a climate effect. But in all the ecosystem services that depend upon global fisheries, for example, we've done a lot of things that are ultimately irreversible, at least in a short ish timespan, but we get to choose how bad things become or you can say also how good things are, because a lot of a lot of we've done has also made made life better for many people.
Pieter Abbeel: Now, when you take it even further, step back and take first principles, look at things. Do you think green energies can cover everything? Maybe with the right machine learning tools and so forth to make that all work?
David Rolnick: Yes, we have the technology to use renewable and zero or low carbon energy for our energy needs across society. But we also have to reduce the amount of energy that is being required by society. So it takes both. There are a lot of parallel efforts that have to go on. And it's not as simple as saying that we have the technology to produce energy in a low carbon way. There are also various new ones. One of the reasons why machine learning can be useful in this kind of situation is because it can help boost adoption or or mitigate bottlenecks to adoption of technologies. Like, for example, you don't have a sense of exactly how much power is going to be generated by a solar panel or a wind turbine since these are variable sources of electricity. Then how do you know that you're going to match supply? What happens currently is electrical grids have fossil fuel turbines running as backup generators in case the sun goes by the cloud or the wind stops blowing. If we have better forecasts, then about supply and demand, then we don't have to do that as much. That kind of situation is really going to try to move us towards full adoption of green energy. There are a lot of smaller things that are needed. It's not just building a better solar panel, though, that is definitely useful.
Pieter Abbeel: Well thank you for coming on the show. Those are all my questions today.
David Rolnick: I really enjoyed it, too. So thank you. That was a really interesting conversation.