AI, Sustainability, and Climate Change

Kris Sankaran

March 12, 2020

This talk: https://tinyurl.com/vtrsjf9

Contact: kris.sankaran@umontreal.ca

• Historically: Tools help people transcend physical limits

​

• Today: Limits in ability to share and process information

The Manchester-Liverpool line, opened in September 1830.

The most retweeted tweet of all time (as of March 2020).

• Nothing is inevitable: people and institutions are constantly reorganizing society

​

• What directions are taken can be deeply influenced by tools

​

• How might our tools help us organize into a more sustainable civilization?

Map of new (orange), operating (yellow), and closing (grey) coal power plants, from CarbonBrief.

Roman road system during the reign (117–138AD) of Hadrian.

Outline

• Useful conceptual frameworks
• Amplification
• Augmentation

​

• Examples from computational sustainability

​

• Climate change and machine learning

Amplification

Packaged Interventions

• These are technologies that help vulnerable populations, which can be packaged up and sent over

​

• Canonical example: vaccines

​

• Attitude can keeps promising projects from maturing

One Laptop per Child

• $100 laptop

​

• Greeted with great fanfare…

​

• … but failed to produce results (grades, attendance) in randomized controlled trials [1]

What happened?

• Children used the laptops for entertainment, games (and who are we to judge?)

​

• Packaged interventions are implementation dependent

​

• Access alone is not enough

Guess X

What if the full power and vividness of X teaching were to be used to help the schools develop a country's new educational pattern? What if the full persuasive and instructional power of X were to be used in support of community development and the modernization of farming?

​

​

​

​

​

​

Guess X

What if the full power and vividness of X teaching were to be used to help the schools develop a country's new educational pattern? What if the full persuasive and instructional power of X were to be used in support of community development and the modernization of farming?

​

​

​

​

​

​

​

Quote from Wilbur Schramm, 1964 (founder of Stanford University Communications Dept) [2].

Theory of Amplification

• « what people get out of technology depends on what they can do and want to do even without technology » [Geek Heresy, K. Toyama]

​

• Technology alone does not do much
• Talent needs to be nurtured, not just given laptops

​

• Technology can dramatically change costs, but context + intent are critical

Corollaries

• Inequalities can be exacerbated
• Differential access, capacity, and motivation
• Those who need the most benefit the least

​

• Explains failures of packaged interventions
• But also successes of projects that augment competent partners

​

• Local successes need not scale

Automation & Augmentation

Du Prony’s Tables

• Division of labor in the calculation of logarithmic tables

​

• [Hired hairdressers unemployed after the French Revolution…]

​

• Inspired Babbage’s Difference Engine No. 1

Gaspard Du Prony and some entries from his tables.

Automation

• Automation tends to follow division of labor
• Tasks are automated, not jobs

​

• For AI & social good, seek out work that has already been decomposed into small tasks
• AI applications don’t materialize out of thin air

Babbage’s difference engine: the automation of arithmetic.

The Memex

• Thought experiment from As We May Think
• Vannevar Bush, early computer pioneer, 1945

​

• A device that let you quickly search through and tie together pieces of information

​

• Augmentation is qualitatively different from automation

Augmentation

• AI can augment sensing and acting

​

• Sensing: Compress data to human-consumable form

​

• Acting: Inform the decisions people will make

Connections

Augmentation and amplification are dual to each other

• Ability vs. motivation to do a task

​

AI is most useful when it’s a piece of larger systems

• AI for Good —> Public Health (using AI), Education (using AI), Renewable Energy (using AI), …
• See also: Hooker, « Data For Good » lacks precision

Computational Sustainability & the Developing World

Developing world contexts [3]

Challenges

​

• Planning is relatively chaotic (rather than data driven)
• Poor infrastructure (including in computing)

​

Opportunities

​

• Lots of people online, data generated rapidly
• ML hacks are useful, in absence of more comprehensive solutions

Themes

• Intelligence gathering
• Story: How many doctors in Nepal?

​

​

​

• Amplify the limited # of experts

​

​

​

• Allocating scarce resources

Point-of-Care Diagnostics

• Pathology experts are often in high demand

​

• Have to prepare blood smear slide and recognize malaria

Need to keep in mind,

​

• Is it actually the bottleneck? (TB + slide prep)
• Artifacts will be present
• Needs to work with low-power devices

Example plasmodium detection on a real-world slide [4].

Detecting Asphyxia

• Diagnostics for newborn asphyxia require experts

​

• Phone app can detect asphyxiated cries to make sure parents seek medical treatment

Another example of point-of-care diagnostics, using speech signals from crying children [9].

Crop Disease Monitoring

• Governments have agricultural extension workers

​

• Promote farming practices
• Check for pests
• Give these extension workers smartphones

​

• Images processed to see presence of crop diseases
• Can alert farmer, and build maps for policymakers

Counting the whiteflies responsible for cassava mosaic disease in the mCrops app [5].

Commuting & Traffic

• Traditional commuter behavior surveys are expensive
• Use Call-Detail-Record data as a cheap proxy

​

• Road-based congestion sensors also expensive
• Use existing CCTV images as a cheap proxy

Using CCTVs to monitor traffic in Jakarta, from [6]. Left is vehicle type detection, right is optical flow.

Humanitarian OpenStreetMap

• Accurate maps help relief workers know what resources are available where

​

• Humans are good at generalization, while machines are good at scaling

A humanitarian OpenStreetmap campaign (pure crowdsourcing) helped to map features in Nepal after the 2015 earthquake [7].

Land Use Mapping

• Land use maps are useful for conservation, urban planning, and food security

​

• On-the-ground surveys are slow and expensive

​

• Satellite imagery provides an easily accessible proxy

A landcover mapping tool, allowing rapid human feedback and model retraining [8].

Climate Change &

Artificial Intelligence

Climate Change Basics [10]

Carbon from ground to atmosphere —> Warming planet

​

Shifts in climate patterns

• Rising sea levels (instability in sea shores)
• Greater extremes in water cycle (drought, floods)
• Species are not adapted (loss in biodiversity)

​

Definitions

• Mitigation: Reduce amount of carbon in atmosphere
• Adaptation: Increase resilience to coming changes

Community & Climate Change [11]

The discourse on climate change has changed over the last decade or so.

​

• Grassroots activism
• Going beyond green consumerism
• Lost faith in leadership of politicians and experts

Role of Machine Learning [12]

• Not just data from climate models (though that's important too)

​

• ML can support decarbonization efforts across domains

​

• ML is not a silver bullet
• … but it can be a potent amplifier / augmenter
• Given context and intent, can have real CO2 Impact

Case Studies

Forecasting Energy Supply & Demand

• Renewables have variable supply, depending on weather

​

• Need to manage the grid to minimize use of backup (fossil fuel-dependent) plants

​

• Energy scheduling improves with better forecasts

Example approach that couples energy demand forecasting with grid scheduling strategies [13].

Detecting GHG Emissions

• Where are GHG emissions the worst?

​

• Better data can drive high-level priorities

​

• Sensors can collect huge amounts of raw data, ML is needed to guide action

​

Methane detection from hyper spectral cameras [14].

Reducing Transportation Demand

• Turn sensor (inc. image) data into usable traffic estimates

​

• These are used as inputs to traffic demand models
• Can inform infrastructure and policy efforts

​

• Estimate / forecast public transportation ridership

​

• Many types of data: smartcards, CDRs,
• Helps plan public projects, reduce urban sprawl

Labeled trucks from satellite images, to estimate truck traffic along routes with fewer sensors [15].

Estimated traffic flow in Yangoon, using cell tower networks [16].

Efficiency in Buildings & Cities

• Control systems can make buildings more responsive
• Increase energy efficiency, e.g. HVAC need forecasting
• Can respond to energy grid signals

​

• ML can generate useful infrastructure data
• E.g., from LiDAR, satellites, or street view images
• Informs policy (# solar panels, or needed retrofitting)

Example HVAC control system using both data and physical priors [17].

Estimates of built infrastructure using LiDAR, from ongoing work of Milojevic-Dupont and Creutzig.

Public Health

• Climate change is a health hazard, through heat waves and decreasing air quality

​

• As habitats for mosquitos expands, ML can support disease surveillance and outbreak forecasting

​

• Passive sensors can measure exposures experienced by vulnerable populations

Estimates of air pollution mortality burden in China, using satellite imagery linked with public health statistics [18].

Personal energy use

• ML can help estimate personal carbon footprints
• And carbon associated with consumer choices

​

• Can help evaluate interventions for behavior change

Two apps (North and Sense) for personal carbon footprint tracking.

Agriculture and Forestry

• Agriculture creates emissions, when it could be a carbon sink

​

• ML can allow adaptation under heterogeneous conditions

​

• Carbon stock can be estimated from aerial imagery

Estimating carbon stock lost due to deforestation [19].

Distinguishing weeds from crops, in precision agriculture.

Supply Chain Management

• Modern supply chains are complex, but optimize profit, not decarbonization

​

• Waste prediction and reduction another possibility (especially food waste)

ML Examples

Computer Vision

• Creating useful annotations from « found » imagery
• Satellite / aerial imagery, street view

​

• Deliberately placed sensors
• Methane-detecting bots, biodiversity camera traps

Labeling and counting birds, using an active learning pipeline [21].

Estimating the number of solar panels using satellite images, since there are no official databases [20].

Text Data & NLP

• Summarizing and navigating huge literature areas (materials science, climate change)

​

• Generating useful annotation of found data
• Processing receipts for personal footprint, using social media for disaster response

Example classification of hurricane-related tweets, from the database compiled in [22].

A map of the climate change literature, from [23].

Time-Series

• Forecasting for real-time resource allocation
• Predictive maintenance, energy supply + demand, drought risk

​

• Forecasting + attribution for long-term decision making
• Transportation demand, climate system features

Forecasting, for a predictive maintenance system built for the New York City power grid [24].

Predictions of cryosphere properties from [25].

RL & Control

• Reducing emissions from complex human-made systems

​

• Designing responses in complex natural environments

Deciding whether to allow controlled burns in national forests, using RL [27].

« Power usage effectiveness » in data centers increases when using an RL control mechanism [26].

Unsupervised & Transfer Learning

• Sometimes have large unlabeled proxy, coupled with small, labeled counterpart

​

• Encoder models useful for navigating complex spaces

Features trained on plentiful proxy tasks transfer well in limited data settings [28].

An approach to searching for concrete formulas with lower carbon impact, using generative modeling [29].

Mila Projects

Motivation: Cloud Emulation

• How will clouds change in response to climate change?
• Source of uncertainty in warming
• High clouds reflect sunlight
• Low clouds trap heat

​

• Existing physical models are computationally costly

​

• We build cheap approximations in [Yuan et. al. 2019]

Simulation Approach

• Simulate cloud samples, conditional on weather context
• Modis satellite: Reflectance, describing cloud optical depth
• MERRA 2: Colocated meteorological measurements
• Apply a computer vision method (Conditional Generative Adversarial Network)

Simulation Approach

• Simulate cloud samples, conditional on weather context
• Modis satellite: Reflectance, describing cloud optical depth
• MERRA 2: Colocated meteorological measurements
• Apply a computer vision method (Conditional Generative Adversarial Network)

Example real (left) and simulated clouds (right) based on previously unseen meteorological measurements.

Simulation Approach

• Simulate cloud samples, conditional on weather context
• Modis satellite: Reflectance, describing cloud optical depth
• MERRA 2: Colocated meteorological measurements
• Apply a computer vision method (Conditional Generative Adversarial Network)

We’re currently adapting the strategy to work with cloud types measured at more local scales.

Motivation: Glacier Mapping

• Mapping glaciers using satellite images and standard GIS tools is very time consuming
• Good sign when partners find problem important enough to do manually

​

• HKH glaciers are important because
• Can be used to monitor impact of climate change
• Are water stores for 250 million people

Preliminary Results

Underlying image, predictions, and supplied labels (ground truth?) from a current model.

Conclusion

Recurring Themes

• Accelerated experimentation (batteries, nuclear fusion, new materials)
• Remote sensing (emissions, infrastructure data, deforestation)
• Improving efficiency (freight consolidation, food waste)
• Approximating time-intensive simulations (climate, energy, policy)
• Need for interpretable, causal, and gray-box models �(solar forecasting, disaster planning, informing policy-makers)

Avoid pitfalls

• Flashy != impactful problem
• We probably shouldn’t all work on fusion

​

• Jevon’s Paradox: Making something more efficient won’t necessarily lead to decarbonization
• e.g., ridesharing may put more cars on the road

​

• Work with domain-expert collaborators
• Communities have been working to reduce carbon emissions for decades now

Roadmap for action

• Learn. Identify how your skills may be useful (our paper is one place to start)

​

• Collaborate. Find collaborators (e.g. researchers, entrepreneurs, established companies, or policy-makers).

​

• Listen. Listen to what your collaborators say is needed.

​

• Deploy. Ensure that your work is deployed where its impact can be realized.

Climate Change AI (www.climatechange.ai)

Discussion forum (forum.climatechange.ai)

Past Workshops

Recordings and papers all online.

Thank you!

• Resources
• Slides: https://tinyurl.com/vtrsjf9
• climatechange.ai
• forum.climatechange.ai

​

• Talk to me about,
• Mila humanitarian AI
• Climate Change AI initiative
• Data Science for Social Good
• DataKind
• Teaching at DL India, Nepal AI School
• …

[1] Fairlie, Robert W., and Jonathan Robinson. "Experimental evidence on the effects of home computers on academic achievement among schoolchildren." American Economic Journal: Applied Economics 5.3 (2013): 211-40.

[2] Toyama, Kentaro. "Ten myths of ICT4D." Summer School on Computing for Socio-Economic Development (2010).

[3] Quinn, John, Vanessa Frias-Martinez, and Lakshminarayan Subramanian. "Computational sustainability and artificial intelligence in the developing world." AI Magazine 35.3 (2014): 36

[4] Quinn, John A., et al. "Deep convolutional neural networks for microscopy-based point of care diagnostics." Machine Learning for Healthcare Conference. 2016.

[5] Aduwo, Jennifer R., Ernest Mwebaze, and John A. Quinn. "Automated Vision-Based Diagnosis of Cassava Mosaic Disease." Industrial Conference on Data Mining-Workshops. 2010.

[6] Caldeira, João, et al. "Improving Traffic Safety Through Video Analysis in Jakarta, Indonesia." Proceedings of SAI Intelligent Systems Conference. Springer, Cham, 2019.

[7] Anhorn, Johannes, Benjamin Herfort, and João Porto de Albuquerque. "Crowdsourced Validation and Updating of Dynamic Features in OpenStreetMap-An analysis of Shelter Mapping after the 2015 Nepal Earthquake." ISCRAM. 2016.

[8] Robinson, Caleb, et al. "Human-Machine Collaboration for Fast Land Cover Mapping." arXiv preprint arXiv:1906.04176 (2019).

[9] Onu, Charles C., et al. "Ubenwa: Cry-based diagnosis of birth asphyxia." arXiv preprint arXiv:1711.06405 (2017).

[10] IPCC, 2014: Climate Change 2014: Synthesis Report. Contribution of Working Groups I, II and III to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change [Core Writing Team, R.K. Pachauri and L.A. Meyer (eds.)]. IPCC, Geneva, Switzerland, 151 pp.

[11] McKibben, Bill. The global warming reader: A century of writing about climate change. Penguin, 2012.Rolnick, David, et al. "Tackling climate change with machine learning." arXiv preprint arXiv:1906.05433 (2019).

​

[11] McKibben, Bill. The global warming reader: A century of writing about climate change. Penguin, 2012.

​

​

​

​

[12] Rolnick, David, et al. "Tackling climate change with machine learning." arXiv preprint arXiv:1906.05433 (2019).

[13] Donti, Priya, Brandon Amos, and J. Zico Kolter. "Task-based end-to-end model learning in stochastic optimization." Advances in Neural Information Processing Systems. 2017.

[14] A camera that sees methane. https://www.discovermagazine.com/planet-earth/a-camera-that-sees-methane

[15] Kaack, Lynn H., George H. Chen, and M. Granger Morgan. "Truck traffic monitoring with satellite images." Proceedings of the 2nd ACM SIGCAS Conference on Computing and Sustainable Societies. 2019.

[16] Lwin, Ko Ko, Yoshihide Sekimoto, and Wataru Takeuchi. "Estimation of Hourly Link Population and Flow Directions from Mobile CDR." ISPRS International Journal of Geo-Information 7.11 (2018): 449.

[17]

[18] Liu, Miaomiao, et al. "Spatial and temporal trends in the mortality burden of air pollution in China: 2004–2012." Environment international 98 (2017): 75-81.

[19] Asner, Gregory P., et al. "High-resolution forest carbon stocks and emissions in the Amazon." Proceedings of the National Academy of Sciences 107.38 (2010): 16738-16742.

[20] Yu, Jiafan, et al. "DeepSolar: A machine learning framework to efficiently construct a solar deployment database in the United States." Joule 2.12 (2018): 2605-2617.

[21] https://agentmorris.github.io/camera-trap-ml-survey/

[22] Muhammad Imran, Shady Elbassuoni, Carlos Castillo, Fernando Diaz, and Patrick Meier. Practical Extraction of Disaster-Relevant Information from Social Media. In Proceedings of the 22nd international conference on World Wide Web companion, May 2013, Rio de Janeiro, Brazil.

​

​

[23] Callaghan, Max W., Jan C. Minx, and Piers M. Forster. "A topography of climate change research." Nature Climate Change 10.2 (2020): 118-123.

[24] Rudin, Cynthia, et al. "Machine learning for the New York City power grid." IEEE transactions on pattern analysis and machine intelligence 34.2 (2011): 328-345.

[25] Shepherd, Andrew, et al. "Trends in Antarctic Ice Sheet elevation and mass." Geophysical Research Letters 46.14 (2019): 8174-8183.

[26] Evans, Richard, and Jim Gao. "Deepmind AI reduces Google data centre cooling bill by 40%." DeepMind blog 20 (2016): 158.

[27] Houtman, Rachel M., et al. "Allowing a wildfire to burn: estimating the effect on future fire suppression costs." International Journal of Wildland Fire 22.7 (2013): 871-882.

[28] Machine Learning and Decision Making for Sustainability

[29] Ge, Xiou, et al. "Accelerated discovery of sustainable building materials." arXiv preprint arXiv:1905.08222 (2019).

​

​