AI for Social Good – ADM+S Tech Talks
19 May 2021
Prof Flora Salim, UNSW
Prof Milind Tambe, Gordon McKay Professor of Computer Science at Harvard University and Director AI for Social Good at Google India
Prof Lawrence Cavedon, RMIT University
Watch the recording
Prof Flora Salim:
So thank you so much Milnd, for this opportunity that we can actually host you and for your time. I know it’s late there. And this lecture is organized as a part of the course in RMIT AI and Data Science professional course for Masters of AI and Master of Data Science. And thank you to Lars for organizing this with Milind, and also this talk is cross-promoted at the school of computing technologies at RMIT and also cross-promoted to the Centre of Excellence of Automated Decision Making and Society. It’s actually the first talk in our Tech Talk series. So thank you for kicking this off Milind. So I’d like to hand this off over to Lawrence to introduce Milind.
Prof Lawrence Cavedon:
It’s an absolute honour, pleasure, um delight, to introduce Millind. You know it’s very generous of him to make this time to come and present to us. We know it is late over in his side of the world, but I’ve known Milind for you know, more than 20 years. I had the pleasure of visiting his lab and his students and research staff over 20 years ago, which as Flora says- starts to date me, but also dates Milind. But you know, Milind is a superstar in the multi-agent systems and the general AI field. I mean he’s a world leading researcher. His previous earlier known research won many awards for modelling teamwork in multi-agent systems and leading to much more robust multi-agent systems. And he, like I said, he’s won awards for that. He was the IJCAI John McCarthy award winner just a few years ago. So, he’s a leader in the whole AI field. He was previously at the university of Southern California and its associated institute ISI, which I think is information sciences institute.
He was a lab leader there but he more recently a few years ago moved to Harvard which was of course just a recognition of his leadership, so he’s a professor there of computer science and he’s also a director of AI for Social Good at Google research India, which is again, just an acknowledgement of his standing in his field and this new area which he’s been a leader in developing of applying AI to social problems. And you know, problems of benefit to the whole society. So focusing on issues like conservation, public health, social work and public safety, and security. So it’s like I said, it’s a delight and an honour to have Milind come and talk to us. He’s not going to, I guess, go into the depths of some of the sophistication of the technology he’s going to build, but he’s going to certainly talk about the applications of AI to addressing some of the great problems in society in general. So again, welcome Milind, and sorry we can’t pair this with a cricket game for you to enjoy as you usually like to do when you come to Australia, but hopefully we’ll be able to do that in not too distant a future and you’re always welcome here physically, and virtually like now.
Prof Milind Tambe:
Thanks. Hey thank you Lawrence for what a generous- I mean really terrific, generous, warm introduction. I’m very grateful for that and I visited RMIT indeed in 2015 and it was an awesome visit. It just happened to coincide with the cricket world cup at the time, so a fantastic visit.
So I’m going to talk to you today about the AI for Social Impact. For the past 15 years me and my research team have been focused on applying AI and multi-gen systems research towards public health conservation, and public safety and security, with the key challenge of how to optimize our limited intervention resources. And I’ll just get down right to some of the lessons that we have learned.
The first lesson is that achieving social impact and AI innovation go hand in hand. So a concrete example is work we’ve done which I’ll cover today in reducing HIV amongst youth experiencing homelessness in Los Angeles. We are living through one pandemic which is the Covid pandemic, but there are other pandemics, for example HIV.
Harnessing the social networks of these youth, we are able to show that our algorithms are far more effective in reducing HIV risk behaviours among these youth, compared to traditional approaches. What this work required was innovation in the area of social network algorithms and so-called bandit algorithms. With respect to conservation, we have large conservation areas to protect limited range of resources. A concrete example of work I’ll cover here, work we’ve done in Uganda and Cambodia, harnessing past poaching data, we are able to predict where poachers may set traps and for the past several years have been able to remove thousands of these traps. This has required advances in combining machine learning and game theory in so-called green security games. In the past we’ve contributed newer algorithms in game theory in the model called Stackelberg Security Games. This is for counterterrorism, and this was work done in collaboration with the US Coast guard for example, which will be using these algorithms we developed or with the Federal Air Marshals Service and other security agencies in the United States. S second lesson is all of the work I will talk about today, is in partnerships with non-governmental organizations, non-profits, and some government organizations. I’m often asked you know; how do you start this work on social impact? My answer is that the work we do is inspired by the work that’s tremendous work, that the non-profits are doing around the world, and our goal is to empower these non-profits to use AI tools and avoid being gatekeepers to AI technologies for social impact.
A third lesson is in doing this work on AI for social impact, we have to pay attention to the entire data to deployment pipeline. It’s not just about improving algorithms. So our work starts by immersing ourselves in the domain with the non-profit, trying to understand the kind of problems they face, the kind of data they have. Following that, a predictive model – a machine learning model – makes predictions of which of the cases the non-profits has are high risk versus low risk. If we had enough resources we would have intervened on all of the high-risk cases but we don’t, and so which cases to actually intervene on given limited resources? That’s the job of our multi-agent reasoning algorithms. And then finally, field testing and deployment is crucial because social impact itself is a key objective of our work. If we are not achieving social impact then it is not AI for social impact. So in the rest of the presentation I’ll start with public health then go into conservation. I’ll cover papers from 2017 to now, I’ll highlight the role of the key PhD students and postdocs, by putting up their picture in the top right-hand corner of the slides on their work where their work is shown.
So start with this social work on social networks. This work is motivated by trying to prevent HIV amongst youth experiencing homelessness in Los Angeles. There’s 6000 youth who sleep on the streets of Los Angeles every night. The rates of HIV amongst this population is 10 times the rate of the normal house population. So homeless shelters will try to- they cannot obviously communicate with all six thousand youth- so they try to bring in some key peer leaders, educate them about HIV prevention and expect these peer leaders to talk to their friends, and their friends to talk to their friends, and in this way information to spread in their social networks. This is real face-to-face interaction. This is not happening for example, over Facebook now some of you may recognize that this is the classical problem of influence maximization in computer science. The input is a social network graph. We have to choose some K peer leader nodes in this graph, say 5 peer leader nodes or 50 peer leader nodes so as to maximize the expected number of influence nodes maximize the expected number of youth who know about HIV prevention. We assume here that information spreads in so-called independent cascade model and I’ll explain that to you next. So given some social networks remember, this is not over Facebook, this is the actual friendship network, this is an actual network of the youth in one of the shelters in Los Angeles friendship networks. And each number here- we’ve in order to protect the identity of these youth we’ve just given them a number that represents a single youth in this network. So if we pick one youth, let’s call them, let’s call this youth A. Then educate them about HIV prevention. Then in this independent cascade model we assume that there is friend B who will know about HIV prevention with a probability of 0.4 and then B will talk to C, their friend with a probability of 0.4. So C will also come to know about hiv prevention with a probability of 0.4. And so information is cascading in the network in this fashion. So if we are given a budget of five nodes, now we have to select which are the best five nodes, which are the best five youth to select in the network to maximize the spread of HIV information in this network. So is it this five, or is it some other five? That’s the key challenge. Once we select these youth then our social work colleagues are shown here in this picture will you know, there’s a day-long session they go through about HIV prevention and how to spread this message in the network, so that’s a picture of this education session in progress. So that’s the challenge that we face now.
There are standard techniques in computer science for doing influence maximization, but if we try to apply them directly in our work there are three key challenges that come together, and they arise because of lack of data and uncertainty. Now often in the domains we face in AI for social good, lack of data is actually a feature, it’s just the way it is. I’ve seen students who work with me who will complain that you know, they want to do data science but there’s not real clean data sets available in these kinds of domains that we work with, working with marginalized communities and so on. And that’s just the way it is. We have to work with the fact that we have limited data. And so one of the challenges as a result, is that we do not know the exact propagation probability like 0.4. We don’t have immersion in this domain which clearly shows it is hard to get those exact probabilities. Peer leaders themselves sometimes don’t show up, so we may say we want to select these five youth but these are youth under difficult circumstances, so some may- we may have selected a youth but that youth may not have a bus fare to come to the shelter to go through the session. They may send their friend for example, so we have to handle this peer leader no-shows. And thirdly, normally in computer science when this problem is studied, they assume that the whole social network is given as input, but in our domain we do not know the social network ahead of time. All we can do is ask a few of the youth who their friends are. We cannot ask it of the entire set of youth, so we may say you know, we can go to a shelter and in one day ask 20 youth about who their friends are and in this way uncover a small fraction of the social network, but that’s all we get and based on that we have to figure out who are the key influencers. So again, these are challenges that we just have to address in our domain.
I’m going to just sketch some ways we solve these problems. I’m not going to get into some of the details if anybody’s interested, I’m happy to have a conversation later, but hopefully at least it gives you some intuition about how we are, how we solve these problems. So I’ll refer to a couple of these. So, for example, uncertainty over social propagation probability normally in influence maximization as we discussed. It is assumed that if we talk to a youth – C, then an adjacent youth, their friend D will get informed with the probability of 0.4 which is known ahead of time about HIV prevention in our domain. This probability is not known so we can say that we are sampling this from some distribution, but we may not even know the mean of this distribution. So therefore we can say the mean of the distribution lies within some interval. So now we are faced with the problem of influence maximization. When there’s uncertainty in the propagation probability and this problem we handle by this, basically this is a problem of robust influence maximization. We solve this problem as a game. So on the one hand our algorithm is trying to choose peer leaders to maximize the spread of influence, but we assume that nature is trying to choose parameter settings, is trying to choose those propagation probabilities to cause our algorithm to perform as worse as possible. And so for those who are familiar we’re basically just trying to solve this as a zero-sum game against nature, where we are trying to maximize the payoff, nature is trying to minimize the payoff, and the payoff is a measure of regret. A ratio of the outcome of the policy that we have chosen, to what we could have chosen optimally had we known nature’s parameter settings in advance. So I’m not going to go into details, but basically by solving these kinds of algorithms we can get a robust policy.
The second problem I’ll discuss is sampling of social networks. You know, we could imagine sending our social work colleagues to a homeless shelter sitting, they can sit there, they can sit there for days, they can talk to all of the youth and kind of figure out the entire social network within a shelter as to who’s who’s friends with whom. This is very costly and if you want this work to be applied in different cities in Los Angeles, San Francisco, etc, then you know, this is not going to scale up. We need a fast, simple way for this technique to work. And to that end we assume that we can only query, let’s say 15 percent of the nodes, we can go get our social work colleagues to go to a homeless shelter, ask questions about say 15 of the youth that may show up in the center that day, that’s it. And they’ll ask a question who are your top five friends, something like that, just note that down. That allows them to essentially get the understanding of the social network. The key here for us is the right sampling algorithm. You know, who do you sample? That’s the secret sauce that we bring to the table. And so the idea is that once we sample exactly this small fraction of the population, then we output the K peer leader nodes and at the same time we guarantee that the performance of our algorithm, even though we have chosen peer leaders from a small sample of the network, is similar to the optimal possible if we knew the full network ahead of time. And this is a done by sampling nodes, randomly estimating sizes of their communities, and then choosing seeds from largest communities. There’s a triple AI 2018 paper, should you want to get into details of how this gets done.
All right so, having addressed some of these challenges, we built a system that I’ll refer to as sampling healer. Sampling, because it’s sampling the network. It is generating a robust policy and what it does is it selects a set of peer leaders. It says: bring in these five leaders, these five youth, and then after seeing who’s present and who’s absent, it’ll say bring in the next five youth, and so forth. So at this time, it was time to do a pilot test with the sampling healer. In each, there are three arms to this pilot test sampling healer which is our sample network. This is our actual algorithm, we recruited 60. Youth healer is just like sampling healer except that it knows the full network, it doesn’t have just the sample of the network, and degree centrality is the traditional approach to bring in the most popular youth. It makes sense right, to bring in the youth with the highest degree, the higher degree centrality- again about 60 youth in that. And we then selected 12 peer leaders in each case as per the recommendation of that algorithm, and this is an actual picture of our social work colleague educating the youth who were identified as peer leaders by these algorithms.
There were obviously 12 different peer leaders in each of these pilot tests, and then the question was at the end of a month how many of the non-peer leaders, the ones we did not bring in for the education session, got informed about HIV? And what we find, is after a month with degree centrality 25 percent of the non-peer leaders got informed about HIV. With healers 75 percent of the non-peer leaders got informed about HIV. So our influence maximization algorithm is clearly far more effective in spreading HIV information, compared to traditional approaches. With sampling healer we seem to be even higher than healer, and that just could be this particular sample that was chosen, because we are doing a small pilot test at this point. But given the performance of sampling healer we decided to now compare it on a larger sample. So in this actual final test we are not only interested in the actual ‘how much information was spread’, but whether there was actual change in behavior. And this is work done with professor Eric Rice. Of social work as far as we know, this is the first large-scale application of influence maximization in social networks for public health, and this was done in collaboration with three non-profits: My French Place, Los Angeles LGBT center, and Safe Place for Youth. We recruited 750 youth in the study, 215 each of the three arms. On the first term was our sampling healer algorithm, the second was degree centrality, the third was no intervention at all. In each case we recruited peer leaders as per the recommendation of the algorithm. 15 of the youth were chosen as peer leaders in each case, and then we wanted to understand what was the change in behavior after a month, and then after three months. And here’s what we found at the end of one month in terms of reduction in condomless anal sex, which is one of the HIV risk behaviors. With sampling healer there was more than 30 percent reduction in condomless anal sex. With degree centrality and control group, there was no reduction at all at the end of one month. At the end of three months we see that degree centrality begins to catch up to sampling healer but still, sampling healer is still better. What’s important here is that the speed at which sampling healer is able to cause changes in behavior is important because this is a HIV risk behavior. So having this behavior change occur faster is important, and also because youth in this community come and go. And therefore, having this behavior change happen faster is important. Also, look at a second HIV risk behavior reduction in condomless vaginal sex, and again sampling healer is seen to perform better. Our colleagues were clearly happy with these results, our collaborators.
I hope this video will play the audio… not working. So I think that trick, the toggle button. All right, I guess I’ll just move on. So the next step here is for those who may be familiar with reinforcement learning, we have been exploring uses of reinforcement learning for influence maximization, rather than our hand-drawn algorithms, and shown that in simulations it performs better. But to test it out in the field is an exciting area for future work. So I’m going to go through a couple more areas of work for about the next 10 minutes and then I’ll stop and take some questions. So I’ll move on to health program adherence.
So remember that we have this data to deployment pipeline. So first stage is collect the data and make a predictive model, making predictions of which of our cases are high risk versus low risk. So here I’m going to look at work that we’ve been doing at Google Research India on maternal and child care. A woman dies in childbirth in India every 15 minutes. 4 out of 10 children are too thin or short. We are very fortunate to be working with a non-profit called Armman which has 18 million women enrolled one of their programs is called Mitra. Basically it’s a weekly three-minute call in a local language to a new or expecting mom. So it may be a call that says in three minutes you know, you are in sixth week of pregnancy you should take this type of health supplement. And they have shown that in randomized control trials, this program leads to significant benefits. So women who listen to the 150 calls that come in, have significant benefits to them and their children in terms of health, compared to women who don’t listen. 2.2 million women have enrolled in this program, so one challenge unfortunately that Armman faces is that women enroll, and then 30 percent or so of these women will become low listeners or drop out of the program. So the first job in terms of prediction, is can we predict which beneficiaries, which of these women are going to drop out. So if we can predict who is going to drop out before they drop out then Armman can call them live or go to their homes and persuade them to not drop out of the program. So here’s how this is supposed to work. You have mom number one, she does not pick up the automated voice call about health in week one, but yes in week two, yes in week three, yes in week four, yes in week five. Mom number two picks up the call in week one and two, but not in week three, yes in week four, not in week five, and so on and so on. For hundreds or thousands of these women it turned out that in this particular case, mom number two and four dropped out of the program. If we could inform Armman ahead of time before they drop out, that mom number two and four are at high risk of drop out, then Armman can do live-calling or other ways of trying to prevent them from dropping, or persuade them to not drop out. So we built a classifier. We did a test with 18 000 beneficiaries. This was done in fall of 2019. We made predictions and indeed in the spring of 2020, verified that women who we had predicted would drop off. We have a high recall, high precision, more than 0.8. And so this seemed like a good system in terms of making predictions about who’s likely to drop off. Okay, we can make predictions, but is that really helpful? So we did another test to actually intervene, to engage in phone calls – live calls – with these women to say persuade them to not drop out. And it turns out that our call intervention indeed causes 44 percent of these women, stopping them from drop out, as opposed to the control group which in this case is about 22 percent. So this call identifying women at high risk and calling them and stopping preventing them, does seem to help. Unfortunately what we still have is that – and this is really in service of helping these three hundred thousand beneficiaries – thirty percent of three hundred thousand, is a hundred thousand beneficiaries. And so if we say to our mom that well, we’ve identified 100 000 women at high risk, now you figure out who to call, that’s not going to work. So we have to prioritize exactly who to call week by week. That’s what I’m going to talk about next, but before I go there let me introduce another domain. This is tuberculosis, it causes half a million deaths in India every year. 3 million are infected and one of the challenges is to cure TB, a person has to take this medicine for six months, and this has got a lot of side effects. So people don’t want to take it for six months. They start and then after a while they’ll drop out of treatment, which is bad for them and bad because it also leads to drug resistant bacteria. So again there are these phone systems to try to track who’s taking their medicine or not, and again we can make predictions based on these phone call patterns as to who’s at risk of drop-off, so that there can be early intervention before they drop out. With data from a non-profit called Everwell, with 15 000 patients and 1.5 million phone calls from these patients, we are able to show that our prediction algorithms are more effective in terms of identifying who’s at risk of drop-off compared to the rule-based system that this non-profit had. So, as I said, we’ve made a prediction but we may have predicted and identified a hundred thousand women who are at risk of drop out, and how do we actually then decide who to call, because you cannot call 100 000 people very quickly.
Because we have limited resources and so here’s a challenge. You can imagine that you have a health worker who has a large number of high-risk patients under her care and she can call only say 10 patients per day. Which 10 patients should recall? So you can imagine in the TB domain she chooses the first three patients, calls them saying hey did you take your medicine? And the first to say yes, the last one says no I didn’t take my medicine, and then she encourages them to take their medicine. And now she has to decide which three patients should I call tomorrow, and she has to keep doing it every day for 180 days because the TB treatment is 180 days. And so you can imagine this health worker trying to figure out who to call every day, and there’s a particular recommendation to do this in a round robin fashion called the first three, then the next three, then the next three. But that’s wasteful because there may be patients who are taking their medicine regularly and you just keep calling them again and again, when they don’t need to be. And there are patients who really need more calling, but we are not calling them because their turn hasn’t yet come. And so these kinds of problems of figuring out who to call can be addressed in terms of what’s called a restless bandit model, where each arm of this bandit is a patient. And we have two states; whether there are deering or not. Again I’m going to skip over a lot of details here because this bandit model would require a lot of time for me to explain the details, which may not be as interesting, but we have techniques to solve this problem so that there’s more efficient policies, and in simulations in this case- this is not real – the simulations we’ve shown, that our algorithm compared to baseline. So the baseline is shown in orange, and our algorithm is in blue. You can see that our algorithm is much faster than the baseline without losing much in terms of solution quality. So, what’s exciting here is we have a 20 000 subject trial ongoing in India, so every week we recommend who to call in terms of these beneficiaries to Armman call them, and they give us results as to what happened. So, we are comparing our algorithm for restless bandits, with the baseline which is based on round robin, and then a third arm where there’s no intervention, and we expect these results to be available very soon. And so it’ll be exciting because it’s the first large-scale application of restless bandage available in the area of public health. So I’ll quickly cover a third topic and then I’ll stop for questions.
This is agent-based modelling for Covid. So in March of 2020 at least in the United States, everything started shutting down due to Covid. So as an agent-based modelling group we started looking at agent-based modelling for Covid19 dynamics first. And this is an article that appeared in the proceedings of national academy of sciences. In our AI 21 paper we looked at tracking disease outbreaks. So if you imagine it’s Harvard or it’s RMIT, and you’re testing students to see if there is an outbreak of Covid, and you sample a small fraction of the students, and you get noisy data of their tests- the results as to whose positive and whose not. And you’re trying to figure out is the reproduction rate going up by one? The traditional epidemiological tools to do that, we showed compared to them that our new approach is provides much more accurate tracking results as shown in this case. But the work that I thought I would cover in a little bit more, is this work with Professor Michael Mina in Harvard school of public health on Covid testing policy, so there are a range of tests that have entered the market with varying sensitivity and cost. Everybody rushes to the PCR test which is a gold standard. Basically it detects very low viral concentrations but the cost is high, and it takes at least 24 hours to get results back. Antigen strip requires higher viral concentrations, it’s less sensitive but has a low cost, and you get results back quickly even in 15 minutes. And so question is again if your Harvard you’re RMIT, or you’re some kind of campus, what test should you use? So you can imagine the model here is that we test all of our students and then whoever is positive we isolate them, and we can do this either by using our less sensitive antigen strip on every student, or the more sensitive PCR test. So if we could do both sorts of tests with equal frequency every three days, and you get results back instantaneously and isolate the student who – or the individual who test positive – then it turns out if you can indeed do this, you get your results back quickly and do this equally with equal frequency then indeed the more sensitive PCR test is better. Because on the y-axis here is total number of infections, and with the less sensitive test more people get infected. However if we just again go back to the fact that the more sensitive test there is a delay in getting results back, a day’s delay for example, then all the advantages of this more sensitive test are lost. There’s a significantly higher rate of infections with the blue, with the more sensitive test. Furthermore, because of cost, if you can run the more sensitive test only every five days instead of every three days, again, all the advantages of the more sensitive tests are lost. So the key result of this modelling was that rapid turnaround time and frequency is more critical than sensitivity for Covid19 surveillance. These results turned out to be picked up by New York Times and Washington Post, and many other places. We were very delighted that Dr Anthony Fauci talked about this paper, this was a great honour and of course our collaborator Dr Michael Mina is consulting with the Biden Covid task force, president Biden’s Covid task force, and so we give all credit to Michael Mina of course, but we think our modelling helped him make this case to advocate to FD and CDC that this rapid test really works. And today as a result of his advocacy, rapid tests are now available in pharmacies locally so that people can use these tests in their homes and isolate themselves.
So, I’ll stop here and take questions and then we can you know, take questions for about three four minutes, and then I’ll do about 10 minutes of talking about conservation.
Okay so I’ll stop for questions. Please go on.
Prof Flora Salim:
Well, it’s been brilliant so far. There are a couple of questions already on the chat and let me just – Sneha, you’ve got a question, I’m just wondering which project you’re referring to, how long did it take to come up with the algorithm after you got a data set in hand. Maybe if you could clarify which project?
Okay let’s get back to that question later on. There’s the second question about the drop-off from Michael. I think this was to do with the TB. What’s used to identify who will drop off- was this NLP from calls? The general information about a person?
Prof Milind Tambe:
I think the more important features in the TB domain were the particular pattern of calls. So you know, they basically have information as to the duration of the call every day and whether the call was picked up or not. So you can imagine that that’s the information we have. So it’s: today this person you know, called at 9am and the call lasted for so much time., something like that. Tomorrow you know, the next day we didn’t get any call at all, third day we got a call, it was at 11 am. So we have this kind of information about the call and that’s what gets used in making predictions. We’ve also looked at using demographic information at least in the TB case, and it helps somewhat, but the calling pattern is more important to make predictions.
In the other case which is the Armman mothers dropping out, there the demographic information is actually quite useful in addition to the call patterns, in order to improve the accuracy of making predictions.
Prof Flora Salim:
Great, thank you. So there’s another project, mean there’s another question about, I think it may be to do with simulation in general. So how would you model the real world entities, especially in a very dynamic environment. What other steps to make sure that models generalize.
Prof Milind Tambe:
So, I guess this probably refers to our Covid19 simulations which are the ones that I mentioned in terms of you know, campuses and so forth. And so, for example in our paper that I mentioned here for agent-based modelling, essentially we are really trying to get at those elements that are relevant for the particular disease spread. So in this particular case there’s a contact matrix of people in terms of who talks to whom with what frequency. We have household general information about the sizes of the household, and then in the household we assume that there is frequent contact. And so in this agent-based model we modelled the entire cities, 2e modelled families and the families had people who are going to work. At work there is a certain contact pattern and so we have information about – just from normal surveys that is available – about the contact that may happen at work. And so all of this kind of information gets taken into account in order to build this model. And then there are more details here about how the infection spreads and so on. But in a sense, in all of these simulations we are trying to make sure we subtract out things that are not important and only keep those things that are important. In this case for example, for disease spread.
Prof Flora Salim:
Great, thank you. I’ve got a question to do with the influence maximization paper. So, it is really great work and I’m actually curious to do with your randomized control sampling, because you mentioned you could only sample a few, not all right? And you want to maximize that. Also, how did you ensure that although it’s the goal to maximize influence, what about if there’s another objective which is to reduce bias in sampling. How would you deal with that? So that is one question I have and the second question I have is to do with the behaviour change tracking. So how did that happen? Is that you know, is someone asking them the changes over time? And I really like the fact that you’re actually asking after three months, after the intervention as well. Because in a lot of behaviour change study, they only do the first one week, or three weeks, and then it went off.
Prof Milind Tambe:
Right, so, with respect to bias in sampling we have done this analysis to try to figure out if it is the case, you know, the different kinds of racial groups. There are groups, there are black, white, Hispanic, mixed race, all of that. And whether the influence was spreading without much a difference. Basically if everybody got a roughly similar proportion of influence. And so for the networks that we had, we did not have disparities along racial groups or gender and so forth. However we can certainly have networks which are not part of our particular study, but other networks where we can show that indeed there can be significant disparities along racial groups, for example in terms of benefits of influence maximization. And so the question you raise about bias- so how to sample to make sure that there are no such disparities is actually an important one. And it’s an area of future work. So we have done work when the network is known for fair influence maximization, but when the network is not known how to ensure that the benefits of influence maximization are spread in a fair way across different communities, that’s an important open question that has not been addressed in literature. With respect to behaviour tracking, we engage in normal social work practices in order to engage with the youth. So basically behaviour change is tracked by them voluntarily giving this information, but the information is given via a computer. So basically it’s a computer interface. It’s not somebody, some person asking them to reduce bias so that they don’t feel like they have to give a favourable answer. So there are other kinds of standard social work practices used in order to understand and track this behaviour change of youth.
And I just remember one thing to do with that also, with your randomized control trial with the three different methodology. You know whether you’re using your sampling method degree centrality, how did you separate the groups there on a control and treatment group?
So basically these are three separate shelters in three different parts of the city, and then we collected you know – so there’s 250 – so basically 80 youth from one centre, 180 from another, 80 from another. This was done so the youth actually, the whole population roughly changes every six months or so, the whole population begins to change so we did this once in January, then you know, six months later July. August again, you recruit 80 youth in each of the three centers, and then again six, seven months later. And so this was you know, the whole thing lasted for about two years, in order to track this. So, basically there are completely different youth communities that get engaged, and also when we recruit we make sure that the youth who are recruited have not engaged with us in the past, so there are all these you know, safeguards, trying to put into place to make sure that there’s no contamination. And also everybody you know, all of the different arms, got a chance at each of the three different places, at each of the three different time slots and so forth. So everything is equitably spread across the three arms of the experiment.
Prof Flora Salim:
Excellent, thank you. Just wondering if you have time for one or two more questions before you move on. Sure, although you’ve got a question do you want to just turn on your microphone and ask this question?
Yes, thank you. I mean now I might misuse my opportunity and ask a different question as well, but I think when you were talking in the Armman project about women like receiving a call and trying to just kind of make them stay with the program, my first reaction was that if I would receive a call to encourage me to stay, I would definitely leave. And I just kind of was thinking – would it be possible to change the model, in a way that you just, the content of the conversation and the potentially emotional reaction to that, can be reflected in that zero one type of model that you had the rest at standard.
Prof Milind Tambe:
That’s an excellent question. So what happens is that these are trained callers who are calling from hospitals to these mothers, and there are you know, let’s say half the women or something like that may not pick up. They make three tries to try to get to the moms, and they engage in an extended conversation with these mothers. And so you’re absolutely right, that this is a sensitive conversation, and so some others will say hey, you know, you’ve been calling me at the wrong time slot, I’m busy at that time, you have to change when you call, something like that. And some of these cases get resolved in terms of if there’s some kind of technical difficulty, like the calling time is wrong or you know, the phone number that was registered was the husband’s phone number, and it really should have been the wife’s phone number, or something like that. So some of these difficulties get resolved. There is a certain component where a woman unfortunately has lost a child and in this case this call is important because otherwise there’s this traumatic experience, because of this automated voice call keeps on coming. And so this is a benefit, but what you’re saying with respect to other reactions that the women may have of annoyance, and then leave – that’s an interesting one. And I should talk to Armman to see what their reaction is to understand what happens. My expectation was that, okay so the trained nurses would deliver standard calls, but the reaction to those standard calls can be different, and how we can model that? That was basically what I was just thinking. Good question and we haven’t done that work.
I would misuse this opportunity- do you have, I know you shared your three lessons at the beginning, but just for someone who’s young and wants to start, what would be a way of approaching NGO’s and convincing them that you can do the job that they want. I mean because they have to share their data, and it’s pretty sensitive. I just thought if you have any more kind of advice, that would be appreciated.
Prof Milind Tambe:
Thank you. Absolutely, thank you for all these very important questions. So I want to offer a few pointers with respect to starting this type of work. One, at Google Research India we have a program for matchmaking, so basically we come out with a call, we came out with one in 2019, there’s one in 2020, here’ll be one in 2021 again. So, essentially we invite university researchers from around the world to apply to the program, we invite NGO’s to apply simultaneously. So last time in 2020 there were 180 faculty members from around, researchers from around the world who had expertise, who applied. We have 158 nonprofits who applied and then we did matchmaking. So every non-profit got to meet with three AI researchers. Every AI researcher got to meet with three non-profits and then they came up with a proposal after their meeting, and then we are essentially going through the proposals and then we’ll select 25 or 30 of these teams, we’ll fund the NGO, we’ll fund a faculty member to go do this work jointly, so this is one avenue. You know, to sort of be introduced to an NGO and continue, to start over, so that’s one thing.
Another possibility which we’ve done locally here at Harvard is that we just invited NGO’s to our University and then basically we invited our few researchers to be in the same – now it’s all in zoom so it’s not physically in the same room – but basically on one zoom call, and we did breakout rooms and again same thing, you know, matchmaking. And then finally out came some projects after some discussion and so that’s another possibility, but what I agree with you, that if you just go to an NGO and say you know, I’ve had these students who after this one meeting in the breakout room, say we’ll work with this NGO and then they expect that the NGO is just going to give them everything in one shot. So it takes, I agree with you that it requires some kind of relationship building and so forth, but you know maybe that you talk to two or three NGO’s, and not everybody agrees but there’ll be one of them who’ll say sure they have this problem, they have this data set, and then after some kind of relationship building they’ll say sure let’s work together. I wonder if you know, we are this class I’ve also actually in my class on AI for social impact here, invited NGO’s to come and lecture in the class and then allowed students to talk to the NGO’s, and form projects that way, and so that might be another avenue. So, I wonder if this particular class that you’re part of could invite NGO’s and maybe that’s a way to build relationships. So, these are some ideas I have, I’d be happy to discuss as more it’s such a important question.
Prof Flora Salim:
Excellent, thank you. We have a hand raise- Javinda do you want to have a quick one? Or maybe we can continue and get Javinda to come in at the end.
Yeah. Hi Milinde, hi Flora. So I’m not sure whether this question is already answered but my question revolves around on the fact of the HIV, and the question is on what basis and features, or background study, did you already predict or choose the best volunteers or peer education educators for HIV? And secondly, in the same context, how did you keep in touch with everyone and see if the volunteers are doing their job to their best?
Prof Milind Tambe:
So with respect to the second question we have in this case, like 250 youth who are recruited as part of each arm of the experiment and so we know who they are because they have been recruited. You know obviously with all the IRB approvals and everything into the experiment. So we know how to contact them and we can follow up with them and so on. So it’s just part of the study, that all of these youth are part of the study. So it’s easy to follow up with them in the sense that we know who they are and so on. Of course there’s some attrition, some people leave the city, all that. But otherwise you know, we have their contact information to follow up with them.
With respect to identifying the volunteers, the peer leaders, that was the algorithm that you know is trying to find the key peer leaders by sampling the network and then figuring out who’s the best peer leader by that robust influence maximization, by playing a game against nature to figure out who are the best people to spread information. And so that’s sort of our algorithm to figure out who are the best peer leaders. Hopefully that answers your question.
Yeah, it does thank you very much. Thank you, wonderful.
Prof Milind Tambe:
So let me proceed. At this point I’ll spend about 10 minutes just giving you a quick overview of some of the work that we’ve done in conservation, and end with some of the lessons that we’ve learned so you know, all over the world in conservation areas there are wonderful wildlife, but their threats to the wildlife snares are traps by the thousands that are placed in order to maim and kill wildlife. So you can imagine the job of a ranger to search these forest areas trying to look for traps, and you can see that this is a very difficult task if you have a very small number of rangers who are in charge of thousands of square kilometres of National Parks that they have to protect. So for example this is the Queen Elizabeth National Park in Uganda where we’ve done a lot of work, and so in order to figure out where to send rangers, you can divide up the park into a one kilometre by one kilometre grid square. Each grid square is so-called target areas where there is water. You can imagine there are more animals, so this is a more important area for poachers. So we are trying to learn poacher response model at each target to figure out how they respond to our patrols. So based on that we can recommend patrolling strategies to these rangers. We have 14 years of data from Uganda when we started this work. So for example it has ranger patrol frequency per each grid square, animal density per each grid square, this obviously we know from maps- what is the distance to nearby rivers, roads, villages, etc. And so based on these different features, we are trying to predict the probability of finding a trap per one kilometre grid square of the park. We did do this prediction using an ensemble of classifiers. In the interest of time I’m going to skip the details of that we produced all kinds of results in the lab, to show recall precision this, that, etc, it doesn’t matter. Our colleagues in the field are not convinced they wanted to do an actual field test, so we selected two nine square kilos. This is our first pilot test- two nine square kilometre areas in Queen Elizabeth National Park, that were infrequently patrolled. These are shown by the green dots. The red dots are where previously a lot of snare traps had been found by the rangers. We are asking the rangers to go to completely new areas, very infrequently patrolled there. One found any snares there and we told them once you patrol them, our model predicts you are going to find snares there. And so the rangers patrol for a month and every day they would report back an email as to what they saw. Initially there was nothing but then they found a poached elephant with its tusks cut off and soon thereafter a whole elephant snare roll as pictured here, that they removed. So poachers were active in the area. They were killing elephant, but before they could kill the next set of elephants, we were able to remove this elephant snare role, hoping that this would or has saved lives of elephants. Then we had 10 antelope snares that were found and removed. So this pilot test was seen to be a success and at this point there was a longer test that was done in Queen Elizabeth and Murchison Falls National Parks in Uganda, and Sripak wildlife sanctuary in Cambodia. In each park we selected 24 areas in frequently patrolled areas of nine square kilometre each. We predicted some of these to be high risk- more snares will be found, some of these to be low risk- less snares would be found. So basically we are not saying go to you know, infrequently patrolled areas and you’re going to find snares. We are saying even there we can discriminate there are some areas at a high risk and some are low risk, and then rangers went out for six months and came back with what they found, and where we predicted high risk indeed more snares were found compared to where we predicted low risk. And this was seen across all of these parks today. What we find in Sripak wildlife sanctuary in Cambodia, these are pictures of the snares that were found because of our predictions. The number of snare captures jumped from 101 to 521, more than five fold increase once our system, which is called Pause, started getting used, and is now in regular use in 2021.
Just in March they found what 1000 snares in the park, and so this is just showing that these techniques that you are all learning about can be very useful for the fight against poaching. And so today Pazuse has gone global, we have made integrated laws with the platform called Smart that is a collaboration with WWF, WCS and 13 other wildlife conservation agencies. Smart is what is active across the globe in hundreds of national parks by rangers, and so by make integrating Pause with smart we’ve made Pause available to all of these rangers. We are testing Pause out in all of the many national parks, and so this very exciting time in terms of Pause getting used in order to make predictions to save wildlife.
So I’m going to just say one more thing which is that there are drones that get used for conservation and we’ve built image recognition to automatically detect poachers in these images. Human beings who shouldn’t be in the park, who are there illegally, and wildlife. So with that I’m going to stop, I’m going to skip over some other details and come to the last part which is key lessons. I wanted to repeat achieving social impact an AI innovation goes hand in hand. We don’t have to sacrifice AI innovation when we try to work on social impact. We want to ultimately empower non-profits to use AI tools and avoid being gatekeepers to AI technology for social impact. Non-profits is where we are going to get our AI for social impact problems. It’s a partnership we have in AI for social impact. We have to look at the whole data to deployment pipeline. It’s no use just trying to concentrate on improving algorithms, because that doesn’t achieve social impact. It’s important to step out of the lab and into the field. I can give you many instances where sitting in the lab we had no clue what was going on and once we went into the field is when the problem became clearer. To embrace interdisciplinary work, whether it’s social work with conservation and others, and of course lack of data as I mentioned is the norm, it’s a feature of this problem set and it should be part of our project strategy. So I wanted to mention many of our collaborators who have mentioned who’ve been co-authors on the papers I’ve mentioned, and with that I’m going to end, and thank you for listening to me. Happy to take any more questions.
Prof Flora Salim:
Beautiful, thank you so much for the great talk again. Scott has a question- do you want to turn on your microphone and ask Milind directly.
Yes. Won’t there be a problem that as your model drives success by targeting certain areas with enhanced patrolling, that you effectively drive poachers into other areas and then that starts to invalidate the results of the model? How do you keep the model trained as poacher behaviour changes?
Prof Milind Tambe:
That’s a really excellent question. So first of all we use the game theoretic model in order to generate patrols. So the most obvious issue where you know, if you change patrolling patterns the poacher will change their reaction, that is taken into account by the game theory. So it sort of is calculating its recommending patrols based on this calculation. So there’s some element of that, which is a very important question you asked, that is already built in. However if you know we have learned a model and the poachers now change that behaviour, that’ll indeed become a concern. In Cambodia for example after our model started getting used just before we went to Cambodia wildlife sanctuary, there was a fire fight between the rangers and the poachers. One of the rangers got shot and had to be moved to the hospital and one reason they told us is because of the increased patrols and loss of snares the poachers had become more aggressive. And you know, they wanted to fight back and so this is clearly a change in behaviour that we had not anticipated. So no, these are all very important and interesting open research questions. Short answer is that using game theory we can handle some, we can anticipate some poacher behaviour changes and reactions, but if the poachers completely change their behaviour, the learn model gets invalidated frequently or the poachers actually engage in active deception, then those become very interesting open challenges that are yet to be addressed.
Prof Flora Salim:
Thank you. That’s really great. Just wondering if there’s any more question from the audience. Deandrea.
Thanks Milinde for the talk, that’s really inspiring. I’ve got a more general sort of question around working with NGO’s. You know, seemingly AI and NGO’s are as far as you can imagine on some spectrum of technology use, and how do you sort of bridge this gap, and what’s your experience with engaging with them, talking the same language as they talk, and in general do you sort of engage in some process of co-development of these models with these NGO’s or what’s the process? How do you go about it?
Prof Milind Tambe:
So that’s a beautiful question. It’s a deep partnership that has to be developed. There has to be trust, there has to be time spent in building up a relationship, and for example with these wildlife conservation organizations it took me years just to initially, you know, I was sort of thinking well I’ll be sitting in my office in the US, I’ll be sending a few emails and then they’ll just say great! We’ll just send you our data, and now let’s just work together. That’s not the way things happen. So this group in Uganda kept saying that you have to come down here and then we can talk and so finally I was saying you know, I was sort of reluctant but then I just flew down to Kampala City, I met with officials there, I went to Queen Elizabeth actually. I went to Murchison falls national park in that trip and met with people in the park and then you know, there was some kind of an understanding, some kind of a relationship that got built. And then after that I received data from Uganda, same with Cambodia too. I mean you know WWF Cambodia, they’re very interested in working, but they wanted us to see actual situation on the ground. And so we went there, we visited with them and after that a very nice relationship got built up, and then we have been working with them throughout. So yeah I agree it takes a little time to build up that trust, build up that relationship, but the end result I feel is very powerful and it’s worth it. And you know for AI researchers I sometimes have to sort of drag people out of the lab to say hey you know, you can’t just assume that people will just come to the lab and say here USB key all my data, process everything, give me back everything, and then I’ll just use it, and you know wonderful. That’s not the way things will happen. And it’s true I mean, they also want to make sure you understand the context in which they’re working and going to Cambodia made us appreciate what a difficult job – I mean the pay for the rangers is low, poachers are shooting at you, the place is like, well I stayed there in the park for one or two nights and it’s just like they gave us the best accommodation and even then it was, I can’t imagine how people can stay there for six months at a time. They’re away from their families. So it gave us a deeper appreciation of the hard work that goes in and when they say something that you know, saying that this patrol is hard, we will change our algorithm to make sure that the patrols are better than you know, what we had given you rather than just saying to just do it. You know, something like that. So yeah, it’s a very good question thank you.
Prof Flora Salim:
Thanks. I’m aware that we’ve taken a lot of your time, are you happy to take one more question Milinde?
Prof Milind Tambe:
Prof Flora Salim:
Natasha would you like to use your microphone?
Hi, my question is just around data security and how much you have to focus on protecting sensitive data, in particular in human focused projects like Armman or whatever.
Prof Milind Tambe:
Yes, that’s also very important and you know earlier we used to work with the federal air marshals and the TSA and the US coast guard, and that was even more sensitive and protected data. So, there have to be very secure protections in place so that you know, there’s things that we will not do for example we will not print a picture in our paper: and here’s all the hot spots in Queen Elizabeth national park in Uganda, with a map showing the poachers where we think their hot spots are or things like that. Those are sort of obvious things, but I mean it’s a real concern, and we have done our best to try to protect data we’ve gone through. Right now at Harvard we’re going through a lot of these negotiations about where the data will be stored so the University satisfied that the data is protected and so on. And so we try to go through all of the data protection protocols and everything else to make sure we keep the data protected, and then try not to put anything, not to make predictions and all these things available easily, not to put them on the cloud, not to put them in the papers, things like that. So, reviewers will say why don’t you print a picture of them where all the poaching hot spots are in national parks, then we can sort of see what is going on, and you have to say sorry that’s not going to happen. So it’s a very good question. I don’t have a much better answer than that other than we take data protection protocols of the University seriously. Okay, thank you, great.
Prof Flora Salim:
Thank you so much for the talk. I just have one final question for you with regards to going to future, I really like your slides today I’m just wondering is there any new way of AI research and innovation that integrates this kind of like societal implications or human-centered behaviour into the way we approach the research because? I mean I think you started you know, how many years ago with NBDI agents and all those things, but how have these things changed because as you mentioned if you go into the real world there’s no perfect data, there’ll be perhaps, even data set for you ready to go.
Prof Milind Tambe:
Yeah so, what is I guess the key hot future research area in this topic right? I think one of the things that I feel is a recognition that the entire data to deployment pipeline is important because when we talk about hot research areas in AI we often focus on just algorithmic portion, you know. This reinforcement learning is now hot, versus something else being hot etc. And so at least for AI for social impact, my view is that the idea that we actually have to see impact in the real world is something that will become more important. What this means is that understanding how to efficiently measure impact, how to actually do clever experiments to measure impact, because these are things like in public health there’s a whole field called implementation science as you may know, whose job it is to actually measure impact and how to actually do clever experiments. So you can see differences, you know does something work in Australia may not work the same way in Los Angeles, in India, or something like that. So these are all questions of culture context impact. So these kinds of things we often say ‘oh that’s just like you know, somebody doing experiments in the field’, but that’s science. I mean that’s like testing an AI algorithm versus some other algorithm understanding. Why something worked, why didn’t it work, all these sorts of things. So that’s on the experimental side. And on the data side I mean there is sort of this idea that there’s so much data like we are drowning in data, everybody you know, but in our domains this is not true, and so we have to come up with all these clever techniques of sampling social networks. And you know this restless bandit, because we can’t quite figure out who’s in what state and all of these kinds of things so these are all outcomes of the fact that we don’t have adequate data, and we have to be sensitive and selective and all these kinds of things about data. So at least in AI for social impact I feel paying attention to the whole pipeline will be an important next step and I think I just feel like there needs to be, and there will be, a greater recognition that we want to bring AI’s benefits to the large sections of the world, you know, who have not benefited from AI benefits, have largely been confined to a smaller section of the planet. And in order to actually show that impact in the field, and not just write papers about it, we have to actually engage in experiments on the ground and measure impact, and deal with the fact that the data are not there and all these kinds of wonderful things.
I think all of these are going to become important. I feel they should become important. Hopeful they’re going to be important. So thank you, that’s I guess where I stand. Thank you so much, hopefully there will be days and time in our interval AI, all those conferences where accuracy is not the goal right. That’s right, and social impact is the goal. And you’re not just measured on you know – did you improve the algorithmic efficiency or whatever, but rather what impact did you actually achieve on society.
Prof Flora Salim:
Thank you so much. So, big round of applause for Milind, and I really appreciate that. I’m very inspire and I’m sure everyone else is inspired too. So and the talk is recorded. I’m going to have it uploaded and I’ll share it with everyone. Thank you.
Prof Milind Tambe:
Thank you so much, thank you for inviting me. Bye.