View profile

#30: Climate Change AI at NeurIPS 2019 (Special Edition)

Hey everyone, welcome to Dynamically Typed #30, the first issue of 2020! Since the holidays are usual
Dynamically Typed
#30: Climate Change AI at NeurIPS 2019 (Special Edition)
Hey everyone, welcome to Dynamically Typed #30, the first issue of 2020! Since the holidays are usually a bit quiet on productized AI and ML research, I’m dedicating today’s edition of the newsletter to the day-long Climate Change AI (CCAI) workshop at the 33rd annual Neural Information Processing Systems (NeurIPS) conference. The workshop included poster presentations, many spotlight talks, four invited talks, and two panels. I wasn’t able to attend in person, but I’ve watched the recording (on SlidesLive in four parts: 1, 2, 3, 4) and summarized the first spotlight talk and the two panels in this newsletter.
You can find the full program on the CCAI website and see my more detailed notes on Notion. Let me know what you think of this special edition! The next issue of DT will be back in the regular format.

Jeff Dean: Computation + Systems vs Climate Change 🖥
Jeff Dean’s spotlight talk featured work that Google AI is doing on climate-related problems. He first mentioned the work Google is doing to get its own carbon footprint down: in 2012, the company started ramping up its renewable energy purchases, to cover 100% of its use starting in 2017. (They’re purchasing as much renewable power as total power they’re using, but they’re not yet always running on renewables.)
Dean describes Google AI’s general approach to using computation and machine learning to help fight climate change as a three-phase process: Algorithm → System → Climate Application. Specific instantiations of this process include:
  • Bayesian inference → TensorFlow Probability → Fusion energy: Google is working with TAE to make a vibrational model for their Norman machine to estimate plasma conditions that are needed to produce fusion energy.
  • Simulation → TPUs → Flood forecasting: Google has hydraulic model maps that are 90x the resolution of public maps, but that’d be 700,000x as much compute to simulate. Google’s Tensor Processing Units (TPUs) can do these computations efficiently enough to make them realistic. They’ve used these simulations to alert two billion people about hundred of thousands of such events (including floods) through Maps and Android.
  • (Simulation → Neural Network Proxies) → TensorFlow → Weather Prediction: Current weather models use partial differential equations (PDEs) that require a large amount of computation especially as compute increases when you make your grid finer. (European weather simulation just switched to a 9-kilometer grid, while 1-kilometer grid is desired for cloud simulation.) Using a neural network directly from sensor predictions, which is more efficient than calculating all the PDEs, makes a more fine-grained (in time and space) and accurate prediction at the short term (< 12 hours). In the longer term, PDEs are still more efficient.
Another set of projects is much closer to Google’s core mission: bringing information to the masses. They think about this in terms of leverage for technology and people: “What information do people need to make good decisions? What are the information leverage points? Who should we build tools for?” Some tools they’ve built include the Environmental Insights Explorer which helps municipalities see where building and transport emissions come from, and Project Solar which shows people the rooftop solar potential of their home.
Overall this was a very inspiring talk. You can watch it on SlidesLive or see my more detailed notes on Notion.
Panel 1: Climate Change: A Grand Challenge for ML
The goal of the first panel was to explore the role of the machine learning community in tackling climate problems. Below I’ve summarized (and in some cases transcribed) each question and answer from the panelists. The text is a bit less polished than usual newsletter text, mostly because there was a lot of content during the hour-long discussion. These are the panelists:
  • Yoshua Bengio (University of Montréal, Mila, Turning Award): happy with the interest from the field but now we have to get our hands dirty. Working on: visualization of impact for education; materials science (longer term).
  • Andrew Ng (Stanford): working on: methane prediction; understanding geospatial imagery.
  • Carla Gomes (Cornell): founder of computational sustainability field; solar fuel materials discovery; takeaway: we can have both social impact and fundamental research.
  • Lester Mackey (Microsoft Research, Stanford): sub-seasonal forecasting at the border of climate and weather w/ predictions 6ish weeks out to help water management etc. deal with climate impact.
  • Jeff Dean (Google AI)
Now let’s dive into the panel.
What recent ML advances excite you for climate problems?
  • Ng: improvements in analysing small datasets (common in satellite imagery e.g. only 100s of examples for things like wind farm detection) using things like unsupervised learning
  • Gomes: combination of ML + simulations + physics-based models (e.g. thermodynamic formulas). Big challenge for ML in science: how to incorporate these things.
  • Bengio: unsupervised learning (e.g. optimizing processes like energy consumption)
  • Dean: excited about huge-scale heavily multi-scale models if we can share data and training between tasks.
  • Mackey: also applying old techniques.
Question RE: CCAI’s 97-page paper; I’ve read the paper, now what’s next? How can I get started?
  • Mackey: get involved in (e.g. Kaggle) challenges in e.g. sub-seasonal climate prediction. Great way to get to know the data and challenges.
  • Bengio: you have to reach out to experts in the fields you’re interested in; you need to read more than you expect; need to engage in discussions. Also connect to other interested people. We also need some humility so that we don’t reinvent the wheel badly.
  • Dean: collaborate! Best projects I’ve worked on had a small group of people with very different expertise.
  • Gomes: connect with experts.
  • Ng: “our job is not done when we publish a paper. Our job is done when we have an impact.” Sometimes we’re very good at the ML bit, but the practical deployment bits are also very hard (change management). One problem: our models are often not robust to the real world. We need to get better at going beyond a result in a Jupyter notebook to a deployment.
How can we encourage work on climate if the metrics we’re measuring scientists against are e.g. papers published instead of tons of CO2 sequestered?
  • Bengio: change your objective function. Look at global impact beyond paper publication. As academics in ML, it seems we’ve lost track of contributing to science and instead keep pushing to get a paper published at the next conference two months from now. We have to step back.
  • Ng: academic system incentivises papers published; but we should push things like tenure applications more towards real-world impact.
How can we prevent ML from having climate-negative impacts (e.g. cost of training)?
  • Bengio: public shaming. Also: when reviewing papers: we should integrate societal impact as an evaluation metric.
  • Ng: ethical question. Good ethical statements are going around (e.g. OECD, Google Microsoft), but two things are missing: (1) not actionable; medical field for example has right checks and balances (doctors have a responsibility to the patient, not to their hospital); (2) due process for making decisions. Wants an ethics code written by and for the AI community.
  • Gomes: we usually optimise for a single objective; we should also consider other criteria and look at impact.
Question from an architect RE Gomes’ point on combining scientific/prior knowledge with ML models: how can we address this challenge in climate change adaptation?
  • Gomes: this is a good question and big challenge. We need to develop new technology that is better at incorporating prior knowledge.
How do you make sure that AI solutions match the actual problems? If you have a hammer, everything looks like a nail—if you want to have a climate impact, should you really be working on AI?
  • Bengio: each of us should ask the question: what can I do? If I know about AI, I should use that. Just AI won’t solve climate, but that’s OK: each of us should contribute their share of what we’re best at.
Should tech unionise?
  • Bengio: some general principles seem to be relevant and useful.
  • Ng: Tech-lash—in tech bubbles, we don’t feel this, but outside of them there is a distrust in tech that we need to solve. Not sure if unions is the best route. Linking back to codes of ethics.
How do we remain inclusive of people in areas disproportionately affected by climate change in climate discussions? (Own addition: especially because all of us working on it will relatively not be affected by it as much.)
  • Dean: question for ML in general—people building AI systems should represent the people who will be affected by AI.
What would you say to the room/community?
  • Bengio: we should question the systems around us.
  • Ng: importance of community: genuinely celebrate the successes of others; when someone asks for a favour, just give it to them; don’t get salty about someone not citing you; having allies (each other) at our backs will help us. “If you can wake up in the morning really happy that someone else had a great result, that’s a really good place to be in.”
  • Gomes: agrees; has a lot of respect for people working on problems in this field; she now worries more about the world than when she was working on abstract problem (but that’s good).
  • Mackey: it’s everybody’s problem and the only way we’ll get through this is together. Take some time to think about how you can help others and the planet and not just your career.
  • Dean: we’re bringing a lot of disciplines and regions of the world together here; we need to work together and celebrate successes; also don’t get discouraged if something doesn’t work.
Panel 2: Practical Challenges in Applying ML to Climate Change
The goal of the second panel was to dig into the long road from an AI demo to a deployment involving many stakeholders and legacy systems, and what the challenges and opportunities there are. I’ve summarized the panel similarly to the previous one. Here are the panelists:
  • Jennifer Chayes (MSFT Research → Berkeley): wants to bring together her EECS school w/ other schools relevant to climate.
  • James Kelloway (National Grid UK): leads team developing ML solutions for the grid operator; thinks prosumers are important in shifting industries (e.g. plastic shopping bags in the UK; now electric cars and solar panels); diversity on teams is important.
  • John Platt (Applied science Google research): partnerships are important (e.g. TAE from Dean’s talk) and you have to spend time understanding each other.
  • Marta Gonzalez (city & regional planning UC Berkeley): develops models for urban infra; computation is just one part.
  • Matt Horne (City of Vancouver climate policy): opportunity to understand e.g. which buildings are causing the most emissions for targeting city resources.
And here’s my summary of the panel.
What are the biggest bottlenecks in using ML for climate problems?
  • Chayes: one problem is that researchers want academic jobs and to e.g. get a postdoc or publications, which is much harder when you’re interdisciplinary. We need better support structures for people pursuing an academic career at e.g. the intersection of climate and ML.
  • Kelloway: bringing theoretical and practical solutions together; e.g. he was presented with a very good optimization for an energy system, which was so fine-grained it couldn’t be implemented (can’t ask a human operator to flip a switch five times a second). Put these people together from the start.
  • Platt: project definition in the sweet spot between risk, impact, and skills fits.
  • Gonzalez: fragmentation and integration at different scales: just a link to where your data comes from doesn’t mean it’s directly ready to be used.
  • Horne: privacy concerns with detailed climate modelling.
How do we prioritise projects under constrained funding? How do we avoid ML becoming a distraction?
  • Gonzalez: bring in communication at the community level of people who are most directly affected.
  • Horne: need to think through mapping to impact.
When and how should decision makers trust ML within established industry like electricity generation?
  • Kelloway: we’re dealing with lots of legacy technology which were not built in mind with responding to ML outputs. One big issue is getting data in and out of legacy systems.
  • Platt: you spend 10% of your time doing the actual ML. Be prepared for the other 90% of getting it deployed.
Dataset maintenance in the long run: you need consistency to keep a model running successfully in production.
  • Platt: some models are one-time discoveries or insights and don’t updating; but yes, in-production models require maintenance and someone who knows what they’re doing.
  • Horne: Vancouver would probably find a partner to do this maintenance instead of doing it in-house.
Data sharing: how can we ensure that private and public sector data sharing happens enough?
  • Kelloway: we’re lucky because we’re regulated to by default make data publicly available; but understandably this is different for companies. Example: they publish regional carbon intensity for their electricity, which companies and consumers can use to e.g. schedule when they’re using it.
  • Chayes: granting agencies could possibly require data openness; much harder for completely private industry.
Challenges and opportunities of cross-disciplinary collaboration?
  • Chayes: it’s not just learning the language of another field, but it’s often much more: understanding the values of another field and the questions that they ask. There’s a math / computer science / stats / ML arrogance in thinking you can just come in and create a model that’ll be helpful to anyone. You really need to learn before you can deeply contribute.
  • Gonzalez: different industries are not always intellectually interested in the same things. We have to be humble: we can make an elegant time series prediction, but then the field says “so what?” where they may be more interested in causes than predictions. More causal inference.
Outside of ML, what can citizens do to encourage councils to start tackling climate change?
  • Horne: there are high levels of concern among the public, but also a lack of information about what the main causes of emissions are and how to change them. There are lots of calls for recycling (low impact) but not for making buildings more efficient (high impact).
  • Platt: local municipalities are much more flexible than e.g. governments and easier to push to change.
  • Kelloway: in the UK we’ve set policies to be carbon neutral by 2050, so we can also work on a more national level.
And that’s it for panel two! I hope you enjoyed this off-format version of Dynamically Typed.
Thanks for reading! As usual, you can let me know what you thought of today’s issue using the buttons below or by replying to this email. If you’re new here, check out the Dynamically Typed archives or subscribe below to get a new issues in your inbox every second Sunday.
If you enjoyed this issue of Dynamically Typed, why not forward it to a friend? It’s by far the best thing you can do to help me grow this newsletter. 🌳
Did you enjoy this issue?
Leon Overweel (Dynamically Typed)

My thoughts and links on productized artificial intelligence, machine learning technology, and AI projects for the climate crisis. Delivered to your inbox every second Sunday.

If you don't want these updates anymore, please unsubscribe here.
If you were forwarded this newsletter and you like it, you can subscribe here.
Powered by Revue