⇓ More from ICTworks

USAID Guide: How to Use Machine Learning in International Development

By Wayan Vota on September 24, 2018

Emerging machine learning and artificial intelligence applications promise to reshape healthcare, agriculture, and democracy in the developing world, and show tremendous potential for helping to achieve sustainable development objectives globally.

At the same time, the very nature of these tools — their ability to codify and reproduce patterns they detect — introduces significant concerns alongside promise.

  • In developed countries, machine learning tools have sometimes been found to automate racial profiling, to foster surveillance, and to perpetuate racial stereotypes.
  • Algorithms may be used, either intentionally or unintentionally, in ways that result in disparate or unfair outcomes between minority and majority populations.
  • Complex models can make it difficult to establish accountability or seek redress when models make mistakes.

These shortcomings are not restricted to developed countries. They can manifest in any setting, especially in places with histories of ethnic conflict or inequality. As the development community adopts tools enabled by machine learning and artificial intelligence, we need a clear- eyed understanding of how to ensure their application is effective, inclusive, and fair.

The USAID Guide to Making Artificial Intelligence Work for International Development aims to inform and empower practitioners as they navigate an emerging machine learning and artificial intelligence landscape in developing countries.

What is Machine Learning?

Machine learning is a set of methods for training computers to learn from data, where “learning” generally amounts to the detection of patterns or structures in data. This differs from how statistical analysis has traditionally been done.

Machine learning approaches flips the usual method is to first develop a model based on mathematical rules and then apply this model to data. Machine learning begins by finding patterns in training data and return a model that can make predictions for new, unseen data. Machine learning techniques can be especially effective at finding complex, nonlinear relationships, and for making sense of large amounts of unstructured image, audio, and text data.

What is Artificial Intelligence?

Artificial intelligence is the science and technology of creating intelligent systems. Artificial Intelligence systems are often enabled by machine learning, and apply data-derived predictions to automate decisions.

Machine learning focuses on learning and prediction, while artificial intelligence applications often create, plan, or do something in the real world. For example, a machine learning model might be used to predict driving time between two places. An artificial intelligence application would plan routes and drive the car.

5 Suggestions for Using Machine Learning

You don’t need to be an machine learning or artificial intelligence expert to shape the development and use of these tools. Digital development practitioners already have deep expertise in their respective sectors or regions. We bring necessary experience in engaging local stakeholders, working with complex social systems, and identifying structural inequities that undermine inclusive progress.

Our perspective needs to inform the construction and adoption of these technologies for them to reach their transformative potential in development.

The five suggestions below from the USAID Guide, focus on implementing partners who are exploring or collaborating with government ministries, donor agencies, academic institutions, and technology companies. Development actors can take these concrete steps now to help their organizations make the best use of these new technologies.

1. Advocate for Your Problem

Technology-development partnerships often pair “solution people” with “problem people.” As a development practitioner, you can help others to stay focused on the problem and ensure that solutions don’t become self-justifying.

Effective technology solutions require those familiar with the problem to be outspoken, well-informed, and focused on development challenges rather than exclusively on solutions. Even if you’re not actively managing a project incorporating ML, a deep understanding of your project and where new technologies will (or won’t) help will set you up for future success.

2. Bring Context to the Fore

Technology experts are often new to international development. Even when technology partners are local to the region, development practitioners have a unique and helpful perspective. They can bring much-needed awareness of some of the ways that the development context makes machine learning deployment more challenging than in well-known “textbook” applications.

Data Considerations

Given the foundational role of data in all machine learning tools, it is necessary to understand who or what is represented by available data, and who or what isn’t. Context can influence what is recorded in data sets that may be used for machine learning.

Context can also affect people’s willingness to share data. People may fear (often justifiably) that any collection or use of personal data could link to a government surveillance system. This can exacerbate problems with bias in data, as the most vulnerable populations often avoid participation and are thus excluded from datasets.

Capacity Considerations

The ability to adopt and maintain machine learning tools depends strongly on the capacity of partner organizations. Leveraging machine learning tools requires capacity to use as well as maintain the models from which they are built. Long-term use can be enhanced by aligning the requirements of model use and maintenance with the capacities of the organizations who will ultimately use the tools routinely as part of their work.

3. Invest in Relationships

Building effective ML-backed tools requires listening to many voices and perspectives. Development practitioners can be key advocates for investing in respectful, productive relationships over the course of both the development and use of machine learning models. In an ideal situation, machine learning tools for development projects can be built and maintained by local technology partners. By working with local companies, we can help to grow fledgling technology sectors and leverage the local knowledge and experience of technologists.

4. Critically Assess Machine Learning Tools

Especially when managing a grant or a contract, the development partner fills the role of a customer on whose behalf a technology tool is being developed. Understanding both how machine learning tools are built and how to assess their performance and suitability will help you to be an informed customer.

One of the most important actions you can take is to ask about model errors and potential bias, and make sure you understand how these were evaluated. If you’re not sure what to ask for, then start with a candid discussion about how a model’s errors can be quantified and what types of bias you’re most concerned about given the context in which the tool is likely to be used.

In particular, identify subsets of the population (e.g., male/female, urban/rural) across which error rates can be compared. If there are uneven failure rates, what real-world consequences might these have? For models that will be evaluating “live” data after an initial testing phase, it’s important to ensure that error testing and performance monitoring continues after deployment.

5. Ensure Two-Way Communication

Finally, for machine learning models that inform decisions about individual people, the development partner may need to view the model as part of a two-way communication process.

  • If someone receives a score (e.g., for credit risk) and wants to know what she can do to improve it in the future, is the model interpretable enough to provide her an answer?
  • If someone feels he has been wrongly evaluated, is there a way for him to seek redress?

These feedback loop processes are often missing, even when decisions are made without algorithmic help, and correcting this is likely to be more about institutional processes and priorities than about technology. When it comes to receiving feedback, providing explanations, or correcting mistakes, it is often better to create formal channels than to rely on ad hoc improvisation.

Listening to the people impacted by our programs is always good development practice — no high-tech tool will change that.

Filed Under: Data
More About: , , , , , ,

Written by
Wayan Vota co-founded ICTworks. He also co-founded Technology Salon, MERL Tech, ICTforAg, ICT4Djobs, ICT4Drinks, JadedAid, Kurante, OLPC News and a few other things. Opinions expressed here are his own and do not reflect the position of his employer, any of its entities, or any ICTWorks sponsor.
Stay Current with ICTworksGet Regular Updates via Email

2 Comments to “USAID Guide: How to Use Machine Learning in International Development”

  1. Neil Penman says:

    Good advice and timely as machine learning and AI have a lot to offer International Development.

    I would add another suggestion which is to “embed” the machine learning in more conventional systems. Its usually wasteful to try and process all the data in a large AI system and certainly you don’t want your AI tool managing the 2 way communication! You can often narrow down the problem domain using simple questions such “What is your age?”. And then call the appropriate specialised AI tool to answer difficult questions defined by the context of the answers you have received.

    For this reason I have been embedding machine learning queries into ODK forms. The app can call out to the AI tool based on answers to questions in the ODK form. This can be done offline or online. The following post illustrates this approach using AWS Rekognition. http://blog.smap.com.au/adding-artificial-intelligence-to-your-forms/

    As a rule of thumb the machine learning should I think only be a small part of any solution particularly when you include the non IT aspects of the solution.

    • Wayan Vota says:

      Great point! Artificial intelligence and machine learning are not going to achieve real results all by themselves. They need to be embedded in an ecosystem that can support and extend its computing ability. Here is a post from a Googler running Google Forums on how they incorporated ML to improve their product.