Ray Witlin / World Bank https://creativecommons.org/licenses/by-nc-nd/2.0/legalcode
show image

Sooraj Shah

Contributing Editor

Sooraj Shah is Contributing Editor of New Statesman Tech with a focus on C-level IT leader interviews. He is also a freelance technology journalist.

How machine learning is helping the World Bank to tackle famine

In 2017, more than 20 million people across north-eastern Nigeria, Somalia, South Sudan and Yemen faced famine or famine-like conditions. Today, 124 million people live in crisis levels of food insecurity, meaning they rely on urgent humanitarian assistance just to survive. Over half of them live in areas affected by conflict.

It’s for this reason that the United Nations (UN), World Bank, International Committee of the Red Cross, Amazon Web Services, Google and Microsoft have teamed up to launch the Famine Action Mechanism (FAM), an initiative dedicated to preventing future famines.

“The whole point of the FAM is really to have a coalition of key partners that bring different value propositions to tackle the biggest challenge of all time, which is ending famine,” says Franck Bousquet, senior director of fragility, conflict, violence and forced displacement at the World Bank Group.

The reality, he says, is that at the moment the humanitarian help comes too late – when a crisis has already occurred and when people are already dying.

The World Bank and its partners have looked to AWS and other global tech firms to better leverage data so that it can anticipate and have improved early warning systems that would trigger action and financing much earlier on.

However, Bousquet emphasises that this initiative is not about replacing existing early warning systems, and data provided by IPC and Fusenet, but complementing them.

For example, AWS has focused on providing FAM with conflict-related data.

“90 per cent of famine in the past few decades has happened in countries affected by conflict directly or indirectly, so AWS is focusing on providing this data. AWS is also helping us with forecasting; this means instead of having reports every six months at a country level about the amount of food and security and what percentage is at crisis level, we are aiming for better forecasting on a six to 12 month period of what percentage of the population will be at risk. This will enable us and our partners to provide financing much earlier than is the case today,” he says.

For FAM, it’s not just about getting better information but linking financing with those early warning systems so that it can shift the focus from responding to preventing and taking early action. Bousquet believes this will not only save lives but could also save up to 30 per cent of humanitarian costs.

Currently, all of this work is still in a proof-of-concept (POC) stage which could last between six months and one year, within which time FAM hopes to launch pilots in a number of countries such as Afghanistan.

Bousquet explains that the project has three big pillars – one of which is about predicting famine with better use of data. The second is a focus on leveraging funding and the final one is on the implementation arrangement.

“On this pillar, we’re focusing on if you know famine is going to happen and you know you have capacity to provide funding, you need action on the ground; the ability to provide support, social safety nets and other programmes for those people suffering. It is quite a complicated programme, otherwise we would have stopped famine decades ago,” he says.

What makes it even more complicated is that each country has very different situations.

“When you look at famine situations in Yemen, Somalia and Mali – those are not the same. Some will be more drug-related, others more conflict related, and there won’t be one way of functioning for all countries, but we are working on being able to adapt to those different types of situations,” Bousquet states.

Machine learning as a differentiator

The main theme is that ability to link every part of the chain; the data, early warning systems, financing and action. This is where AI and machine learning could make an impact. AWS has built a machine learning pipeline from data ingestion and storage in Amazon S3 to model deployment using SageMaker.

The idea is to take into account all the different causes of famine, and so AWS uses a dataset processed by the World Bank and UN partners including features from satellite imagery, conflict data, weather forecasts, local food prices, and agricultural production. It uses this to train multiple machine learning models, and AWS claims that these models have been able to identify significant and distinctive patterns across certain regions and countries. AWS also claims that its tools have improved existing models with an 11 per cent increase in accuracy for Somalia and South Sudan.

Bousquet believes that machine learning will not only improve the frequency and accuracy of information but also enable FAM to get access to data that it is currently not tapping into.

“That’s especially important when you’re working in conflict situations where famine happens most of the time and we don’t have the associated data. This is what we could get from AI and machine learning,” he says.

He concludes that FAM requires so many different parties because the epidemic of famine cannot be solved with just one type of organisation, subset of people, or sector.