Knowledge Vault 4 /79 - AI For Good 2023
Building a decision-augmentation platform to address global issues
AI FOR GOOD ML Workshop
< Resume Image >
Link to IA4Good VideoView Youtube Video

Concept Graph & Resume using Claude 3 Opus | Chat GPT4o | Llama 3:

graph LR classDef intro fill:#f9d4d4, font-weight:bold, font-size:14px classDef project fill:#d4f9d4, font-weight:bold, font-size:14px classDef lessons fill:#d4d4f9, font-weight:bold, font-size:14px classDef framework fill:#f9f9d4, font-weight:bold, font-size:14px classDef collaboration fill:#f9d4f9, font-weight:bold, font-size:14px classDef technical fill:#d4f9f9, font-weight:bold, font-size:14px A[Building a decision-augmentation
platform to address
global issues] --> B[Workshop: building
decision augmentation platform. 1] A --> C[Risto Miikkulainen:
project history, goals, MVP. 2] A --> D[Breakout sessions: use cases,
requirements, ideas. 3] A --> E[Project started
during pandemic. 4] A --> F[Demo built, expanded
to XPRIZE. 5] A --> G[Project Resilience: generalize
platform for challenges. 6] C --> H[Platform allows team
collaboration. 7] C --> I[Data-based modeling
approach used. 8] C --> J[System found solutions early,
predicted Delta surge. 9] C --> K[Advised Iceland government
on school reopenings. 10] C --> L[Missed Omicron surge
due to data limitations. 11] D --> M[Evolved rules for explainable
prescriptions. 12] D --> N[XPRIZE challenge engaged
global teams. 13] D --> O[AI leverages human insights
for optimization. 14] D --> P[Continuing work with
GPAI, NSF. 15] D --> Q[Oxford tracker: policies
across 180+ countries. 16] E --> R[Key lessons: political
will a barrier. 17] E --> S[Subnational data increases
effort, valuable granularity. 18] E --> T[Huge volunteer effort for
daily data collection. 19] E --> U[GitHub limitations: custom
platform built. 20] E --> V[Consider entire
data lifecycle. 21] F --> W[Harmonize cloud, networks,
IoT, AI, cybersecurity. 22] F --> X[Open architectures needed
for data sharing. 23] F --> Y[Common global standards enable
interoperability. 24] F --> Z[Future architectures: data,
networking, AI. 25] F --> AA[Framework applied to land
use optimization. 26] G --> AB[Models trained on historical
data for carbon capture. 27] G --> AC[Pareto front visualizations
show trade-offs. 28] G --> AD[Optimize at different
geographical granularities. 29] G --> AE[Secure collaborative development
architecture presented. 30] G --> AF[Requirements: security, CI/CD,
data management. 31] H --> AG[Uses AWS services,
avoids lock-in. 32] H --> AH[Designed for re-use
across domains. 33] H --> AI[Breakout 1: use cases like
education, poverty reduction. 34] H --> AJ[Other use cases: weapons,
space debris, fishing. 35] H --> AK[Consider multi-objective optimization,
unintended consequences. 36] I --> AL[Breakout 2: collaboration
objectives, stakeholders. 37] I --> AM[Tailored communication, strategic
partnerships needed. 38] I --> AN[Define resources, milestones,
data sources. 39] I --> AO[Build partnerships before
crises for rapid response. 40] I --> AP[Breakout 3: technical challenges,
data visualization. 41] J --> AQ[Reusable project templates
accelerate applications. 42] J --> AR[Explore conversational interfaces,
be cautious with AI. 43] J --> AS[Continuous model updating, time
lags are challenges. 44] J --> AT[Address compute access, cost
for developing countries. 45] J --> AU[Post-its: sustainable peace,
lifestyle optimization, surgery. 46] K --> AV[Collaboration: easier data collection,
common vocabularies. 47] K --> AW[Technical ideas: data-for-compute
exchange, generative AI. 48] K --> AX[Visualizing uncertainty
a top challenge. 49] K --> AY[Generative AI promising
but risky. 50] K --> AZ[Bias unavoidable,
diversity helps. 51] L --> BA[Decision makers need
tailored UX. 52] L --> BB[Engage decision makers early,
show benefits. 53] L --> BC[Allow interactive tradeoff
exploration. 54] L --> BD[AI can provide cover
for tough decisions. 55] L --> BE[AI improves on human-only
decisions in complexity. 56] M --> BF[Communicate probabilistic
uncertainty intuitively. 57] M --> BG[Project relies on volunteers,
needs help. 58] M --> BH[Bi-weekly community meetings
held. 59] M --> BI[Workshop materials shared,
more planned. 60] class A intro class B,C,D,E,F,G project class H,I,J,K,L lessons class M,N,O,P,Q framework class R,S,T,U,V collaboration class W,X,Y,Z,AA technical class AB,AC,AD,AE,AF technical class AG,AH,AI,AJ,AK technical class AL,AM,AN,AO,AP technical class AQ,AR,AS,AT,AU technical class AV,AW,AX,AY,AZ technical class BA,BB,BC,BD,BE technical class BF,BG,BH,BI technical

Resume:

1.- Workshop on building a decision augmentation platform to address global issues and enable teams worldwide to collaborate on solving local/global problems.

2.- Risto Miikkulainen presented history, goals, and MVP of the project. Keynotes provided overview of this and similar projects, data, and architecture.

3.- Breakout sessions held on use cases, requirements, and new ideas. Groups reported back findings after lunch.

4.- Project started during pandemic to enable global collaboration on prediction and prescription for COVID-19 mitigation using AI.

5.- Built demo in a few months showing what AI can do. Expanded into global XPRIZE competition.

6.- Project Resilience aims to generalize the platform beyond pandemic to other societal challenges using same technology.

7.- Platform allows teams to work on common core, share data. Provides cloud-based decision support to those without access to data scientists.

8.- Data-based modeling approach used, in contrast to traditional epidemiological modeling. Allows prediction and prescription without assumptions.

9.- System discovered interesting solutions early on, like importance of schools/workplaces. Predicted Delta surge in India.

10.- Advised Iceland government on school reopenings in 2021. Showed potential to communicate with decision makers and have positive impact.

11.- Limitations: Missed Omicron surge as it happened everywhere at once with no early data to learn from.

12.- Rules were also evolved to provide explainable prescriptions, not just opaque neural networks, to help decision makers understand recommendations.

13.- XPRIZE pandemic response challenge engaged global teams to develop predictors and prescriptors. Informed some government policies.

14.- Found AI can leverage collective human insights by using human-designed prescriptions as starting point for evolutionary optimization.

15.- Continuing work on multiple fronts - with GPAI on multiple models, NSF on communication. Project Resilience broadening to new domains.

16.- Toby Phillips presented Oxford COVID-19 Government Response Tracker collecting data on policies across 180+ countries to enable modeling work.

17.- Key lessons: Political will a barrier more than tech for global data sharing. Systems established early as hard to add data later.

18.- Subnational data exponentially increases collection effort but national-level loses important granular policy variations. Careful design decisions needed.

19.- Relied on huge volunteer effort to manually collect policy data daily. Motivating long-term participation is hard. Frequency impacts viability.

20.- Used GitHub to share data but has limitations. Lacked standard APIs. Custom platform had to be built to use the data.

21.- Mongchul Lee emphasized considering entire data lifecycle, ensuring common understanding through ontologies/models, building trust, governance, decentralized user-centric approaches.

22.- Important to harmonize technologies like cloud, networks, IoT, big data, AI, and cybersecurity in an overall framework.

23.- Open architectures and data spaces needed to share data and create value. Marketplace and transactions mechanisms for incentives.

24.- Common global standards (e.g. from ITU and others) are key to enable interoperability and reuse of data at scale.

25.- Future architectures must consider end-to-end data, networking and AI - "DNA". Privacy, ethics, policy, regulation are critical elements.

26.- Olivier Francon demoed applying same framework used for COVID-19 to land use optimization for carbon capture.

27.- Models trained on historical land use and emissions data. Prescriptions suggest land use changes to maximize capture and minimize cost.

28.- Pareto front shows tradeoffs between competing objectives. Visualizations help decision makers understand impacts of policy choices.

29.- Can optimize at different geographical granularities. Local optimizations may have negative global impacts that need to be considered.

30.- Prem Natarajan presented an architecture enabling secure collaborative development and deployment of models by distributed teams.

31.- Key requirements: security, tools, CI/CD, data management, cost efficiency, access control, insights sharing, and "bring your own model" support.

32.- Uses AWS services like Studio and Model Registry, but aims to avoid lock-in. Enables end-to-end ML pipelines.

33.- Designed for re-use across multiple problem domains and use cases. Focuses on iterative MVPs to manage costs.

34.- Breakout 1 on use cases: Education to optimize resource allocation. Poverty reduction via policy simulation. Agriculture to optimize land use.

35.- Other use cases: Weapons systems and autonomous conflict. Assisting space debris avoidance maneuvers. Fishing stock management.

36.- Need to consider multi-objective optimization and unintended global consequences of local decisions. Work closely with domain experts.

37.- Breakout 2 on collaboration: Define clear objectives and value proposition. Identify stakeholders, sponsors, domain experts, volunteers.

38.- Plan tailored communication and engagement for each group. Build strategic partnerships for reputation and trust.

39.- Define resources, milestones, data sources, implementation roles. Consider data accuracy. Have regular demos and progress updates.

40.- Build partnerships before crises emerge to enable rapid response when needed. Data aggregation across organizations a key missing role.

41.- Breakout 3 on technical challenges: Visualizing and explaining data and model uncertainty is critical but difficult.

42.- Need reusable project templates to accelerate new applications. Must consider data as a commodity and enable exchange.

43.- Explore conversational interfaces for reports, explanations and Q&A. But beware over-reliance on generative AI.

44.- Continuous model updating, handling decision/effect time lags, and complex objective interdependencies are challenges.

45.- Must address compute access and cost barriers for developing country participation. Transfer learning may help with limited data.

46.- Post-its yielded use case ideas: sustainable peace, lifestyle optimization, surgery, resource management, education, migration, water, fishing, budgets.

47.- On collaboration: Easier data collection, engaging public and domain experts, common vocabularies, tailored communication, know when to end failing efforts.

48.- Technical ideas: Reusable templates, data-for-compute exchange, conversational AI, model uncertainty, generative AI for transparency, engaging startups.

49.- Visualizing uncertainty was top technical challenge. Also latency, complex objective trade-offs, explainability, infrastructure access.

50.- Generative AI holds promise for easier explanations but has risks. No-code tools can accelerate adoption. Model cards aid transparency.

51.- Bias cannot be eliminated as it is fundamental to human and AI reasoning. But diversity of perspectives helps.

52.- Current decision makers often don't directly use analytical/AI tools. Need to tailor UX to their workflows to get adoption.

53.- Best to engage decision makers from the start. Provide relevant outputs in their preferred formats. Show tangible benefits.

54.- Enable them to interactively explore tradeoffs and uncertainties. Support the entire decision lifecycle, not just the decision point.

55.- Adoption may be easier if it provides "cover" for unpopular but necessary decisions. Build trust gradually.

56.- Consider the baseline - AI may not be perfect but can still improve on human-only decision making in complex environments.

57.- Communicate probabilistic uncertainty intuitively. Set appropriate expectations that there will still be some errors.

58.- Project Resilience relies heavily on volunteers and needs help with development, testing, deployment, outreach and more.

59.- Bi-weekly open community meetings held to coordinate work. Recordings available. New contributors actively welcomed.

60.- Workshop materials to be shared via website and email. Ongoing collaboration enabled via mailing list. More workshops planned.

Knowledge Vault built byDavid Vivancos 2024