In the early aughts, author William Gibson wrote that the future was here, just not evenly distributed.
But while the US, Canada, China, and other world powers disproportionately enjoy the benefits of technological advancements, a shift is starting to occur. Researchers in the developing world, like Africa and the Middle East, are beginning to use artificial intelligence to address systemic problems holding back growth. At this week’s Neural Information Processing Systems conference, the AI industry’s biggest, experts see opportunity to use the technology to fill gaps in healthcare, agriculture, and finance.
Ernest Mwebaze, a lecturer at Makerere University in Uganada and speaker at the NIPS workshop on AI in the developing world, is using machine learning to help farmers detect disease in crops using smartphones. The project, called Mcrops, shows farmers whether their cassava plants are infected with hard-to-spot diseases after analyzing a photo taken with a smartphone. Since cassava is often locally grown by independent farmers, the project aims to increase food security and reduce poverty.
“We try to impart farmers with the tools to diagnose disease,” Mwebaze says, “so the farmer can do immediate interventions.”
While the software has the potential to help farmers better understand the health of their crops, Mwebaze says the biggest issues around the project aren’t technical, but social. Some farmers are illiterate, so the app has to be intuitive and include pictures. Access to technology is also limited, and while cassava is often farmed by women (pdf), more men own phones.
These are problems not often considered in the developed world. Coupled with the technological constraints of unreliable power and internet in some parts of Africa, as well as a lack of data that Silicon Valley tech companies have amassed on their customers, AI researchers in Africa need to think differently about realistic solutions to problems.
William Herland, a co-organizer of the conference workshop and Ph.D student at Carnegie Mellon University, says that working on problems in the developing world, where tasks are harder due to those limitations, could force new innovations in Silicon Valley as well.
“How can [the developing world] hint at new things that we’re totally missing because we’re not thinking about a whole class of problems?” Herland said. “The distribution from which people at NIPS come are generally kind of well-educated, middle to upper class, developed world countries. Part of our bias is simply what we know and what we’re exposed to.”
For example, Herland points to the use of satellite imagery to identify economic activity. Since the United States has troves of data on where wealthy and poor people live, using satellite imagery to do so would be redundant. But in Africa, that data isn’t available. So researchers have turned to analyzing the number of lights over time seen in satellite images to infer economic growth. In solving that problem, researchers from Stanford were able to show new ways of teaching image recognition algorithms to learn from smaller datasets.
While datasets are typically smaller in Africa, according to experts who spoke to Quartz, they are more accessible. Mwebaze says, compared to the US, it’s easier to obtain patient data from medical facilities in parts of Africa to train machine-learning models for healthcare. This is a stark difference to the norm in developed countries like the US, UK, and Canada, where data protection laws installed to protect individual privacy complicate the process of accessing that sensitive data.
IBM Kenya is working with mobile payments from the massively popular mobile money transfer service M-Pesa. Skyler Speakman, a scientist at IBM Kenya, is working to build a credit scoring system that works exclusively off a person’s M-Pesa transactions. The technology would give banks more confidence to give small loans (about $50) to people, Speakman says.
While the data, just a list of a person’s transactions, can clearly show how much money a person has and whether they have to capability to pay back the loan, Speakman is really interested in finding a way to use this limited transaction data to determine whether the person is trustworthy enough to loan money. The credit scoring system is currently in use, though IBM won’t yet say which banks are using the service.
“The more philsophical question is can we really capture someone’s character with this type of information?” he says. ” We either need to be a little more creative about the way we go about collecting data, or trying to do machine learning with less. Machine learning in the developing world is a great example of doing more with less.”