Milind Tambe is the Principal Scientist and Director of ’AI for Social Good’ at Google DeepMind. He is also Gordon McKay Professor of Computer Science and Director of Center for Research in Computation and Society at Harvard University. Google DeepMind brings together two leading AI labs — Google Brain and DeepMind — into a team which has been responsible for some of the biggest research breakthroughs in AI over the last one decade. Tambe and his team have developed AI systems that deliver real-world impact in public health; maternal and child health, public safety, and wildlife conservation. Tambe spoke to indianexpress.com on the challenges nonprofits face while working on maternal and child health in India, in reaching out to rural and disadvantaged women and communicating pregnancy health regimens, and how tech could be helpful in solving these challenges. Edited excerpts: Venkatesh Kannaiah: Tell us about your work in maternal and child health in India from a Tech for Good perspective? Milind Tambe: We have two goals for our work in India — to take up a health challenge and to use AI and help underserviced communities. We have been successful in both our goals. Our work on using AI has improved and matured and we have been able to help a large number of people. In India, as part of Google DeepMind, I work mostly on maternal and child health issues. The goal is to use AI to increase information access to pregnant women, who are in rural and impoverished regions of the country and provide them with timely information. The journey of pregnancy and childbirth needs timely medical interventions like vaccinations and we work on the messaging part of the journey. We work with a nonprofit, ARMMAN, which has a call centre with health workers calling a cohort of two lakh expectant mothers, in hard-to-access regions of the country. There are lakhs of beneficiaries who are being serviced by this nonprofit network, and we work on increasing the efficiency of the 'calling process'. We can determine through AI, as to who could be at risk of dropping off, and how we need to maximise our calling resources and time to get the best results. One must understand that with a nonprofit, or with a government agency, there are no unlimited resources to follow up on health communication. Over the journey of pregnancy and childbirth, in this programme, there are around 140 automated messages that are sent to expectant mothers. Our work is to find out who the health workers should call for personal follow ups, who needs it the most, who is more vulnerable, the periodicity, and the reminders to keep the maximum number of people within the messaging network for better health outcomes. We have found that after our intervention, the dropout rates for pregnant women in the programme had fallen by approximately 30 percentage points. More mothers now stay in the programme, and call centre efficiency has increased. The nonprofit that we work with has around 2 lakh mothers on its calling programme, and many expectant mothers just drop out due to a variety of reasons. Our AI intervention has helped stem the drop off and has helped more than 3.5 lakh mothers. Venkatesh Kannaiah: You talk about Restless Multi-Armed Bandits. What are they, why should we study them, and what problems does it solve? Milind Tambe: The name comes from imagining a gambler at a row of slot machines (sometimes known as ‘one-armed bandits’), who has to decide which machines to play, how many times and in which order to play them, and whether to continue with the current machine or try a different machine. Restless multi-armed bandits concept is now being used in India in our initiative to raise the efficiency of the calling process. To explain it further, the multi-armed bandit problem simultaneously attempts to acquire new knowledge and optimise its decisions based on existing knowledge. The model has also been used to control allocation of resources to different projects, answering the question of which project to work on, given the uncertainty about the difficulty and the payoff. Health awareness programmes using mobile phones suffer from low engagement over time from participants and health workers have to make live service calls to encourage the participation of beneficiaries. The goal in a restless multi-armed bandit problem is to decide which arms to actively manage. This is a challenging optimisation problem, especially when there are many arms and limited resources to allocate. It is very relevant to the problem that we have on hand. Who among the expectant mothers is likely to drop off, whom should we call, whom to remind, whom to follow up, etc, all with limited resources. Apart from our work with expectant mothers, tuberculosis prevention is another problem where the restless bandit model can be used. TB patients in most cases need to follow a six-month regimen of taking medicines and it is easy to drop out of the programme. We can rank or prioritise the number of people who would be needing the personal intervention or visits from the health worker to remind them to take the medicine. There is another interesting intervention with RMAB in the wildlife sphere. Assume we need to monitor a large national park or forest area, with the threat of poachers killing the wildlife. We can divide the whole forest or park into grids of one square km each, and predict the probability of poaching and how we need to deploy forest officials to prevent poaching. So this can be used in different dynamically evolving situations. Venkatesh Kannaiah: Can you talk about your SAHELI programme, and how is it being deployed in India? Milind Tambe: SAHELI is the name of the AI programme which is used by the nonprofit ARMMAN in India. It uses the Restless Bandit framework to identify beneficiaries for outreach. We have worked on multiple innovations in this model and its development, using real-world data. This is the first application demonstrating the utility of RMABs in real-world public health settings. SAHELI has been developed as a platform with the ability to be scaled to more nonprofits in more domains. Venkatesh Kannaiah: Tell us about how you identify nonprofits in India to work with? Milind Tambe: For us, we look at the problem statement and then come up with the AI models and the solutions. We are problem-centered in our approach and don’t force fit models or solutions on to problems. Our team cuts across various verticals in the company, and based on requirements of the partnerships and the challenges that we face, we can bring people from all over Google into the project. It is a flexible arrangement, the issue is to solve the problem. ARMMAN too emerged from a chance meeting that we had with Aparna Hegde. We are open to working with nonprofits in the maternal health sector in India. It also helps that even before our partnership, ARMMAN was tech-friendly and was digitising most of their operations and their data. Some of the interesting interventions in the health sector could not be followed up as some of the nonprofits were not robust enough in their digitisation process. I assume that it will change in due course and we would like to be involved in many more maternal and child health programmes and projects in India.
