My notes from Climate Change AI summer school 2024

Agenda Started on June 20th, ended on August 31st.

Day 1 - Introduction to Climate Change and AI - June 20, 2024

Lecture Recordings:

Lecture Materials:

Day 2 - AI for Agriculture, Forestry, and Other Land Use - June 24, 2024

Lecture Recording:
https://youtu.be/99U0ePt9DzY

Lecture Slide Decks:

  • AI for Agriculture pdf
    • direct and feedback effect on agriculture and climate change (food production & water vs. land use change, biodiversity loss and emissions)
    • Who wants to to know what: Farmers: crop yield and perfomance, threats (potential and actual) , env. (soil moisture, rainfall, temp.,..) / Policy makers: same as farmers but in much bigger scale (spatio-temporal) ML methods that is discussed : ML is used to extract Essential Agricultural Variables (EAV) from satellite observations crop type mapping -> Multi-class classification other typical ML problems Binary classification, example of the work can be used by CERISE project is Vits for SITS with the GitHub using vision transformer for satellite image recognition (more in [[AI - Remote sensing]]) segmentation, regression, OOD detection (outlier detection) and some other methods yield estimations, pest and disease detection, precision agriculture , robotic farming
    • We need to increase yield to reduce the land use
  • AI for Forestry pdf
    • super intersting course!
    • sounds of nature (https://xeno-canto.org/) and other sources
    • https://lila.science/datasets
    • https://www.worldweatherattribution.org/about/
    • https://www.the-iea.org/
    • great slide about carbon storage
    • large deforestation in global south while global north is rebounding
    • deforestation is contributing to 18% of global anthropogenic emission / we should differentiate between the deforestation and forest degradation
    • Climate crisis results in biodiversity loss as well
    • Monitoring, Reporting and Verification (MRV) for any incentives and ML can be used
    • Bioacoustic : identify the sounds of biodiversity with sound scape
    • equitable ai: bias

Tutorial:

Day 3 - AI for Biodiversity and Ecosystems - June 26, 2024

Lecture Recording:

https://youtu.be/I9TIRncO9HE

Lecture Slide Decks :

  • AI for Wildlife Conservation (pdf)
    • habitat loss is concentrated in most vulnerable areas , many species are missing enough information
    • sensors are used in different scales to monitor animals
    • ML for animal detection and conversation and pose detection ; it helps to scale up
    • Reconstruction of environment (3d models 7 LiDar)
    • uncertainty should be always provided
  • AI for Conservation Decisions (pdf)
    • 60% of wildlife is lost since 1970
    • 1M out of 8M species are in danger of extinction
    • In causal inference, the key challenge is that we cannot observe what would have happened to the same unit (e.g., person, group) under a different treatment or condition (the counterfactual). This is often referred to as the “fundamental problem of causal inference.”

Tutorial:

Day 4 - AI for Social Sciences, Economics and Policy, Part I - June 28, 2024

Lecture Recording:

https://youtu.be/q-rbAyZuZjo

Lecture Slide Decks:

  • Climate Innovation Policy pdf
    • innovation: for 2030 goals we already have the technology, but we need political power and will / for 2050 goals we still dont have the necessary tech
    • innovation like tech (as hardware) doesnt happened in vacuum , investment, policies, and regulation are software
    • on average we need 4-5 Trillion $ a year until 2050
    • now carrots are much stronger than sticks in innovation policies
    • political economy of climate change is difficult because benefits are in future and for all, while costs are visible and immediate and bares by specific groups
  • Artificial Intelligence for Climate Action under the UNFCCC pdf
    • UNFCC TEC (Technology Executive Committee) : policy
    • UNFCC Climate Technology Centre and Network (CTCN) : implementation and assistance
    • “#ai4climateaction” initiative by UNFCC started on April 2024 : implementation, capacity building and raising awareness
    • AI Innovation Grand Challenge : Deadline is August 12th 2024

Tutorial:

Day 5 - AI for Climate Science - July 2, 2024

Lecture Pre-Readings:

Lecture Recordings:

Lecture Slide Decks:

  • AI for Climate Science (pdf)
  • AI for Weather (pdf)

Tutorial:

Day 6 - AI for Monitoring, Reporting, and Verification - July 5, 2024

Lecture Recording:

https://youtu.be/j3_jE0ZqsMM

Lecture Slide Deck pdf talk consist of multiple subject , most interesting for me was

Tutorial:

Day 7 - AI for Social Sciences, Economics and Policy, Part II - July 16, 2024

Suggested Readings:

Lecture Recording:

  • AI for Social Sciences, Economics and
  • How GHG influence environment (previous talks)
  • How changes in environment influence human wellbeing (focus of this talk)
    • climate variability impacts on social outcomes : linking the climate and the response function as below:
    • we can build this based on historical function , for example yield function based on temperature or cloud cover ; then combining it with climate outcomes we can have a prediction of outcomes on different scenarios
    • Social Cost of Carbon (SCC): the monetized value of all future net damage associated with a 1 metric ton increase in C02 emissions
    • SCC as basis of the cost benefit analysis
    • discounting damage in the future as it assume that money ( negative as damage) has less value in future compared to present
    • How to estimate SCC?
      • linear regression y = f(c) + controls + e
    • Panel Fixed effect Regression
      • Panel : repeated observation of y and C for multiple locations over time
      • Fixed effect: a set of intercepts , often high dimensional and controlling for time and space (like intercept for a county)
      • regression: ordinary least square linear regression
    • implementing of the fixed effect to isolate the variations to be “as good as randomly assigned”
    • differentiate between the causal effect of the weather and longer term effect that is climate
    • impact of temperature and soil moisture on crop yields : as climate variables effecting covarying variables compared with only temperature
    • challenge
      • non linearity : observation are only aggregated, and local potential could be missed by canceling the variability or due to non-linear response function derivation : non-linearity of the sum-up filed response to generate over the county level (by quadratic form assumption ) Example in the slides
      • Heterogeneity : response differs across space, time, and other attributes
      • Model selections : remove and normalised for all the variables that you want to remove

Lecture Slide Deck:

Day 8 - Ethics, Impacts, and Regulation of AI - July 18, 2024

Lecture Recordings:

  • Ethics of AI (very bad quality voice)
  • Regulation of AI (starting at 1:03:53)
    • new regulation in horizon , China has some limited, US has Biden act, but EU AI act is the first broadly defined AI acts
    • Artificial intelligence liability directive
    • AI definition in AI ACT : machine-based system with autonomy , that infers from input to an output that can influence world
      • excluding : inference goes beyond basic data processing / rule based system by natural persons
    • Four risk levels are defined
      • prohibited ai : social scoring , wide surveillance
      • high risk : product safety , medical use case , education , credit system, migration, military, elections (not included in e-com, search engine, digital marketing)
      • Limited risk: Transparency
      • Minimal risk: Ai literacy (all people who work with AI needs to have educations about AI) example of is chatbots: Disclosure of generated media by AI, a text production for public usage , Labeling of AI-generated and water mark
      • ChatGPT / Claude doesnt fit any of them as they are in all of them
    • Environmental aspect is not included
    • Rules for pipeline
  • All foundation models (approximate GPT4 models) has specific obligations
  • GHG effect pf ICT/AI : around 1.8 - 3.9 % almost 2x more than air travel , inference also has big impact (one image generation = charging an iphone fully), their applications also has impact (like drilling oil for cheaper)
  • water consumption of ai is huge, and lot of it is evaporate (which is GHG)
  • Gen AI will not have big impact on mitigation CC
  • sustainable ai : GHG emission and water usage so far AI is not covered by Environmental regulations so far
  • EU AI act is applied to anywhere as long as they are services and served client in EU / there are transparency requirement / provider of high-risk ai system should disclose their computing consumption / assessment and mitigation of the systemic risks for very large foundation models (GPT4) might be required (not fully clear) / an emissions trading regime for ai is not including the large models and most of them are build in places that have no tight regulation
  • Open-source FMs are excluded (creates a loop hole)
  • Future : require standard , rule on training and define carbon and energy budgets
  • recsai.org The International Expert Consortium on the The International Expert Consortium on the Regulation, Economics, and Computer Science of AI (RECSAI) provides a platform to facilitate cross-disciplinary enquiries on key questions and challenges related to artificial intelligence.

Lecture Slide Decks:

Day 9 - AI for Buildings and Cities - July 19, 2024

Lecture Recordings:

Lecture Slide Decks:

Tutorials:

Day 10 - GHG Impact Assessment of AI - July 22, 2024

Lecture Recordings:

Lecture Slide Decks Summary:

  1. GHG Impact Assessment of AI

    • Impact Assessment: Measures greenhouse gas (GHG) emissions in terms of CO2 equivalent, focusing on the Global Warming Potential (GWP) of AI technologies.
    • Life-Cycle Assessment (LCA): Examines emissions from the entire lifecycle of AI systems—from production (cradle), operation (gate), to disposal (grave).
    • Emissions Scope: Covers three categories:
      1. Scope 1: Direct emissions from owned or controlled sources.
      2. Scope 2: Indirect emissions from the generation of purchased electricity.
      3. Scope 3: All other indirect emissions, including those from customer use.
    • GHG Emissions in Machine Learning (ML):
      • Compute-Related Impact: Includes emissions from ML training and inference, with tools like CodeCarbon used to measure energy consumption. The energy demand for training foundation models (FMs) such as GPT-4 is rapidly growing, particularly for inference, which can account for 60-70% of energy use due to the high volume of inferences.
      • Local vs. Global Energy Impact: Future scenarios of energy consumption vary globally and locally, depending on the energy mix of local grids used by data centers.
      • Google’s ML Energy Use: Reports that 10-15% of its total energy consumption is attributed to ML.
    • Application-Related Impact: ML can reduce costs and emissions in some sectors (e.g., improved cooling in data centers), but it may also encourage increased usage, known as the rebound effect.
    • System-Level Impact: ML can help reduce emissions across the supply chain by optimizing various processes.
    • Shift to On-Device Inference: There is a trend towards performing AI inference directly on devices, which could alter the emission landscape.
    • Importance of Impact Assessment: Understanding the cost and environmental impact of ML is crucial for developing sustainable AI practices.

  2. Responsible AI and Sustainability

    • AI as a Socio-Technical System: AI models use data extensively and have significant impacts on society.
    • AI Ethics and Environmental Concerns: AI ethics often overlook environmental impacts, while sustainability discussions might neglect issues of justice and equity. The Sustainable Development Goals (SDGs) cover both aspects.
    • AI Guidelines: Current guidelines for ethical and sustainable AI have limited convergence, are often vague, and open to interpretation.
    • Efficiency and Usage Paradox: As AI systems become more efficient, their use increases, leading to more significant overall impacts—a concept known as Jevons Paradox or the rebound effect.
    • NeurIPS Code of Conduct (2023): Encourages ethical and sustainable AI research practices.
    • BLOOM Model Governance: Focuses on governance and regulations for large language models like BLOOM with 176 billion parameters.
    • Key Ethical and Sustainability Issues:
      • Representativeness: Assumptions made by AI models can lead to mislabeling, as seen with ImageNet’s biodiversity categories.
      • Evaluation Metrics: AI evaluation should not rely on a single metric; multiple aspects should be measured to understand trade-offs and impacts better.
      • Transparency: Due to the complexity of models like transformers, transparency is limited. Efforts like model cards and datasheets aim to improve understanding of models and datasets.
      • Equity: As language models grow larger, they risk increasing the digital divide and creating disparities in justice and power.
    • Resource for AI Imagery: Better Images of AI provides high-quality visuals for AI.

Additional Lecture Materials:

Tutorials:

Day 11 - AI for Energy Systems - July 29, 2024

Lecture Pre-Readings:

Required readings

Supplemental readings

Prerequisites

  • No special prerequisites, beyond engaging with the required readings above ahead of time.

Lecture Recording:

AI for Power and Energy Systems

Lecture Slide Deck:

Tutorial:

Day 12 - AI for Transportation - July 31, 2024

Lecture Pre-Readings:

Lecture Recording:

Lecture Slide Deck:

Tutorial:

Lecture Recording:

Planetary Health: Impact of Human-Induced Planetary Changes on Human Health

  • Overview:
    • A quarter of deaths worldwide are linked to environmental factors.
  • Health Impacts:
    • Direct: Heatwaves.
    • Indirect: Various environmental changes leading to health issues.
    • Long-term: Climate-induced migration.
  • Vulnerable Populations:
    • Increased heat affects certain groups more severely, including the elderly, pregnant individuals, outdoor workers, those with lower socioeconomic status, and people with specific lifestyles or genetic backgrounds.
  • Air Pollution and Health:
    • The link between air pollution and mortality is well-documented, along with other health impacts.
  • The Role of Informatics:
    • Real-Time Data: Provides immediate insights into environmental changes.
    • Improved Knowledge: Enhances understanding of environmental health impacts.
    • Adaptation Strategies: Includes early warning systems for heatwaves and floods, predicting disease outbreaks, tracking and monitoring air quality, and responding to pandemics.
  • Challenges and Opportunities:
    • Patient-Level Health Analytics: Utilizing electronic health records to generate raw data for patient phenotyping.
    • Climate-Informed Health Data: Integrating climate information (such as temperature, air pollution, land use, and vegetation) with health data.
    • Geospatial Linkage: Combining patient-level data with geographic information to improve outcomes.
    • Data Classification Issues: The International Classification of Diseases (ICD) may not easily align with climate data, complicating the assessment of excess deaths.
    • Healthcare Data Issues: Challenges include biases, privacy concerns, safety, and the multi-modality of data.
    • Biases in Data and Algorithms: Data and algorithmic biases can lead to incorrect health assessments.
    • Need for Awareness: It’s crucial to recognize inequalities and the lack of accurate representation of populations in data.