Artificial Intelligence for Economic Development Conference: Roundup of 27 presentations

Is artificial intelligence the future for economic development? Earlier this month, a group of World Bank staff, academic researchers, and technology company representatives convened at a conference in San Francisco to discuss new advances in artificial intelligence. One of the takeaways for Bank staff was how AI technologies might be useful for Bank operations and clients. Below you’ll find a full round-up of all the papers and research-in-progress that was presented. All slides that were shared publicly are linked here, as well as papers or other relevant sites.

Keynotes

• Tom Kalil: the Bank will need staff who are familiar with both development and computer science, who understand the technology developments coming from computer science and can turn them into development knowledge and action. He (and many other speakers) emphasized the need for more high quality micro datasets, to train AI algorithms.
• Joshua Blumenstock: machine learning (ML) is particularly promising in data scarce environments (e.g. many of the Bank’s client countries). ML opens the possibility of using new data sources to measure poverty and vulnerability, and could be trained to predict changes in welfare, which could improve project targeting, facilitate crisis response, and create new approaches to impact evaluation (slides).
• Susan Athey spoke on the potential uses of AI for causal inference. ML has primarily been used for prediction and there is little literature on the use of ML for causality. While ML won’t solve identification problems, it certainly helps when data is scarce, and can help systemize the search for optimal model specification. All materials (slides, videos, r scripts and a github data link) from Susan’s AEA course with Guido Imbens on the same topic are available publicly here.

AI for Governance

• Big data to increase trust? In Brazil, researchers used a Mixed Initiative Social Media Analysis (MIXMA) to study the relationship between social protest and citizen trust based on sentiment analysis of twitter activity during the 2014 World Cup (N Calderon et al paper, slides)
• In a study of policy options to reduce criminality among ex-combatants in Colombia, a ML ensemble was used to improve the credibility of propensity score matching by allowing for inclusion of 100+ control variables (best technical explanation we heard: “unleash the statistical zoo”) (C Samii, L Paler, & SZ Daly paper, slides)
• Big data to improve government performance? In India, ML algorithms on tax data can more systematically identify ‘suspicious’ firms to target for physical audits (A Mahajan link)
• To quantitatively analyze qualitative deliberation in Indian village assemblies, the Social Observatory used text-as-data through unsupervised natural language processing. Data shows that women are less likely to speak and the issues they raise are less likely to be picked up by others (Parthasarathy et al paper)

AI for Transport & Urbanization

• Use nighttime lights data to ‘train’ for better classification of urban areas in daytime satellite images (Landsat). Better data on the pace and extent of urbanization could improve infrastructure development, industrial policy, environmental planning, and land management (Goldblatt et al paper, slides)
• The HumNet lab is generating urban demand models using mobile phone data. One of many promising applications is to better manage demand and reduce pressure on infrastructure during large events like the Olympics (Gonzalez & Xu, slides)
• By crowdsourcing location updates through smart phones, Mapbox created models to predict how and when people evacuate during disasters (Farley details, slides)

AI Lightning Talks

• Most data on violence is self-reported, under-reported and not available at disaggregated levels. Computer vision algorithms help identify and label destroyed areas in Syria (Hersh, contact)
• ML to predict the likelihood of farmers adopting the use of lime (On, contact)
• Using AI provide personalized agricultural extension information at low cost through interactive voice response, improved by rapid a/b testing. (Reich, Precision Agriculture for Development)
• ML to predict food security in southern Malawi (Knippenberg, paper)
• Sentiment analysis of news has real signal for economic forecasting (Fraiberger, paper)
• Facebook using satellite-based data and government census information to map global population (Nayak, Facebook)
• Improve Educate Girl’s targeting of out-of-school girls in India, by using random forest regressions to predict likely out-of-school girls based on government data (Brockman, IDinsight)
• Use big data to reduce gender gaps in access to credit, use ML algorithms to predict credit worthiness rather than credit history (which women are less likely to have) and allow for different predictors for men and women (Higgins, details)
• Natural language interfaces (e.g. chatbots) to help low-income people navigate complex bureaucracies (Kendall, Digital Financial Services Lab)
• Use ML on satellite imagery to detect changes in horizontal and vertical growth of cities (Clough, Planet Labs)
• Machine learning should complement, not substitute for, development practitioners. Practitioners are critical for identifying the problem to solve, providing local expertise, and contextualizing analysis. (Paul, USAID Global Development Lab) 
• Introduction to an AI-based social robot scientist for automating impact evaluations, tailored to Ghanaian policy makers (Opoku-Agyemang, contact)

AI for Tracking Poverty, Targeting Development

• Satellite imagery data provides a less expensive and scalable method to estimate consumption expenditure and assets (a convolutional neural network trained to identify image features explains up to 75% of variation in economic outcomes). In 5 African countries, performance of satellite models is tested against ground-truth survey data, for outcomes such as maize yields, poverty and access to electricity. (Burke, paper)
• Man vs. machine? A test of 3 methods to predict likelihood of being a successful entrepreneur in Nigeria – expert judges, models developed by economists, and ML algorithms – showed that ML does not obviously out-perform the other methods (Mckenzie & Sansone, paper, slides)
• Current models of predicting credit worthiness have had limited success. A ML algorithm using data on how people use their mobile phones (5500 indicators such as top-up patterns, usage and mobility) is predictive of loan repayment and can be used as an alternative to credit scores, and therefore could be used to extend credit to the unbanked (Bjorkegren, paper)
• Researchers at Berkeley are working on a new initiative to leverage data-rich environments, such as the US, to build ML models and then systematically degrade the data to mimic data-poor contexts. (Hsiang, paper)
• Deep learning models are used to recognize objects from Google street view; the objects are then used to infer socioeconomic outcomes. In this case, cars (separated by make and model) were identified and used to create measures of safety, segregation, and “greenness” (Gebru, slides, paper)
• Descartes Labs is assessing food security in the Middle East and North Africa, through an automated capability to analyze, monitor and forecast the wheat crop, which will allow for real-time alerts and insights (Moody, details).  

The ethics of AI for development
A panel composed of Florence Kondylis (World Bank), Moorea Brega (Premise), Stefano Ermon (Stanford), and Aubra Anthony (USAID) discussed ethical issues and raised the following concerns:

1. Informed consent? Social science worries a lot about making sure human subjects know what they’re being ‘subjected’ to, but in the case of AI participation is not voluntary.
2. Transparency? Incentives for transparency are well-established in academic computer science. However, for tech start-ups, ‘algorithms are the new oil’ and are typically proprietary. Yet knowing what’s in an algorithm is important for establishing trust. If transparency won’t happen on its own, what type of intervention is needed?
3. Privacy? AI presents new twists to concerns over personally identified data, such as ‘demographically identified data’, which need to be carefully considered.
4. Is the ‘ground-truth’ true? Survey data has its own sources of error and bias, which need to be considered when it is used to train ML models.