 So, yes, the topic for today is deep learning to fight hunger and poverty. A little bit about myself, I'm a principal data scientist at Infinitesum Modeling and also another consulting company called MCG. And I'm also the chapter lead for women in machine learning and data science for Bangalore. And my Twitter handle is you, Rangaraju. I recently created a Twitter account, so do follow me on Twitter so I get enough followers. Yes. I'm also the organizer for NeuroAI, which is India's first ever symposium in the interface of neuroscience and data science. We had this conference in Bangalore on August 3rd and 4th for this year in Enforcers and Manipal Crampus. And it was a not-for-profit conference and it's a free conference to the public. We had around 800 attendees in Enforcers Bangalore on the first day. We will be organizing NeuroAI part two next year as well. If there are any volunteers who are interested, please do reach out to me. Infinitesum Modeling deals with public policy research and economic modeling and technology policy research in the social sector. So, today's talk is related to the work which we do at Infinitesum Modeling. Unfortunately, we haven't published our research work yet. So, I won't be able to talk and my collaborators for the research are not present here. So, I won't be able to touch upon the research we are actually doing, but I'll generally touch upon the generic details of the topic. Yes. So, this is a quote line which I would like to start the talk with. Most of the people in the world are poor. So, if we know the economics of being poor, we would know the much of economics which actually matters. This was told by Theodor Skals in a Nobel Prize-winning lecture in 1979. So, this was told in 1979. I mean, there's a lot changed from 1979 to 2019. So, in 1979, there are a lot of people who were relatively poor. I mean, even in India, if you take 1947, we received independence. And so, even during 1979 and all our living conditions were not that great. We were not having this great technological improvements and technological facilities which we have in 2019. So, in 2019, Dasa's statement stand true, not partially true, I would say. The living conditions of middle-class people has tremendously changed. We have all these technological improvements like mobile phones, Internet, broadband and all of it. But for the poorest people, the relative poverty level hasn't changed much. So, for the poor people, poor people get poorer and poorer, but the middle-class people's, you know, wealth has relatively changed. So, this statement stands partially true in 2019. So, much of the talk will be based on this concept called poverty trap. So, poverty trap is a very vicious cycle. So, if you take, most of our research is in Africa. So, poverty trap is like, these are factors which cause poverty trap. In Africa, if you see, Africa is the richest continent below the ground. So, if you see Africa, particularly a country like Rwanda and stuff, it has worth $24 trillion. It has, it is richest in minerals, Congo, Rwanda and all this. Especially Congo, if you take, it has like $24 trillion worth under the ground. But on the surface, it is the poorest continent in the world. So, and why is it poor in spite of having all the wealth? So, it is because of this poverty trap. So, you have this violence and low productivity which causes poverty and poverty will cause hunger and there is natural resources degradation and this in turn will cause low productivity and this loop keeps going on and on. There are, I mean, it's not like there are not many measures taken up. There are several measures taken up by several NGOs and even the government is actively taking measures. But why is government not able to penetrate through this poverty trap? Why is it Congo even after so many years? Like it's still remaining poor and it's going to be the context of this talk. So, these are some of the images of poverty trap in Africa. Alright, anyone from the audience? So, what are your thoughts on this picture? What is the pollutant which you see in the river? Oil, yeah, oil. Alright, one minute. You're right, this is oil. So, this is an image from Shell. So, Congo had Shell unit. Yeah, is it audible now? So, Congo had Shell Petroleum mining unit and Shell Petroleum mining unit did not have proper outlook, out facility. So, they would drain it into the river and the entire river basin of Nile got ruined, a part of Nile got ruined and Shell did not take any initiatives to clean it up because no one is there to question and they just left the river basin like that. So, you see Africa is already poor and a very big international company like Shell coming and ruining the entire river basin. So, people were using this polluted river with oil for agriculture. So, in turn the agriculture of food crops were there and Africa doesn't have any good medical facilities. So, people who drink this contaminated water and fall even more sick and there's no medical facilities as well. So, Shell got away with it until 2-3 years back when few NGOs started filing a suit against Shell and Shell had to pay around millions of dollars to clean up the entire river basin and also, you know, so the NGOs won the case against Shell. So, this was about Shell. So, which is the mineral? Any guess? No. Cobalt. All right. So, Africa, I mean if you see, Congo was the, what do you say, colony of Belgium. So, King Leopold II literally used to send people into the forest and the deadline given to them by the end of the day they should bring in two bags of minerals. So, if anyone is not bringing two bags of minerals, their hands would be chopped off. We can just Google about, so the Africa is that rich in minerals. So, why is it not able to put its wealth into proper use and become a richer country? What is it stopping? Why aren't the economic policy measures that effective? Why are the economic policy measures being taken in Africa failing again and again and again? Congo is in debt. The debt of Congo is $12.2 billion. Around $12.2 billion and United Nations has waived of the debt because Congo recently waived of the debt because Congo has qualified for the waiver and so the economy is going nowhere. It's already a poor country and there's a debt of $12.2 billion. There's no way to come out of this poverty trap. So, yes, that being said about Congo, this is one of the NGO in Africa. It's called Give Directly. So, what Give Directly does is the use case of Give Directly is it has to identify the poorest people in a particular location, in a particular city and enroll them as registrants for getting money. And on the other side, Give Directly will also look for donors across the world who will be willing to donate the money. So, once for every round of... All right, it's being recorded. I'm sorry, for every round of money. So, once it reaches around $10 million. So, yes. So, the use case for Give Directly is it has to enroll the poorest people in a particular locality. So, it has to somehow find out the poorest people who are poorest or poorest. So, it has to really find out the people who really deserve money and enroll them as donors and get the money it has received. $10 million will be unconditionally transferred to these poorest people. And so, Give Directly has been doing this for several years and they have a reputation of one of the best NGOs in Africa who are very ethical and who have been doing good work in Africa. And so, how can you find out... The whole context of the topic is how can you find out the poorest of poorest people in a particular locality? What is the option? Anyone can guess? How will you find out the poorest of poorest people in a particular locality? What has been traditionally been doing is... We have been taking traditionally surveys. One of the famous surveys in Africa is DHS. So, they have been traditionally going and doing door-to-door surveys and they would find out take the economic wealth of the people who are the people who are earning and who are the people and looking at the economic facilities inside the room and they would do a door-to-door surveys. So, these DHS surveys generally take for a particular region like Rwanda and so on. DHS surveys generally take 2-3 years to complete and it is extremely labour intensive. So, sometimes the budget of these surveys allotting volunteers, setting up the entire team, putting up a regulatory team and all this this DHS survey would take around 2-3 years and it would take 100 million dollars in itself to complete and by the end of 2-3 years the statistics would have again changed. So, how authentic are these DHS surveys we are not very sure. That's where deep learning comes into play. So, using the satellite imageries which are available at free of cost you can download. So, it depends on the use case where you want to use. So, for this particular use case you don't require a resolution of great you don't require a resolution of great kind. So, you can freely download the Google static API satellite imagery which is available free of cost and using satellite imageries and applying deep learning algorithms on those things you can find out the poorest of poorest people in a particular locality and you can actually get this done in a matter of 2 or 3 days with extensive computing power available right now you can even get it done in a day and you don't have to wait for 2 or 3 years to get the statistics and you don't it saves you lot of money and it's less zero labour intensive. So, yes and the one of the ways which they do is using satellite imageries they have to find out the roof tops which has got thatched roofs or a metal roofs using satellite imageries. So, if a house has got a metal roof then it means they are better economically developed and if they have thatched roofs then they are poorest of poor. So, the results are not very accurate because a house in Africa might have some houses in Africa might have both thatched roof and metal roof. So, for some reason they want to separate certain member of the family in a thatched roof they would have a house which has both. So, the results are not very accurate but you get these results in a day and with zero cost. So, it's still better. I mean results the accuracy is only around 75 to 74 percent that's why there's lot of research still being done in this area even in India. In India there's a company called JD Insights which is like actively working with Niti Aayog in some of these use cases. There's another use case in India which is being done. So, Selco Foundation. So, Selco Foundation's main motivation is to power up the rural cities which does not have good access to electricity. So, they would want to find out and identify these rural cities and using they I mean going and taking an accurate measurements of these rural cities and powering of the you know setting of the grid would take longer time but using satellite imageries and getting an approximate you know measurements to place these grids would save their time and less labor intensive for them. They've done a great work in rural cities of Oriya. So, you can go and read a lot about this use case. I'm just skipping that because this is not the main context. So, yes, these are some of the challenges with surveys which I've already talked about and so the satellite imageries are available free of course but if you want to go for a paid resolution provider where you want to get images with a better satellite imageries there are vendors who can provide satellite imageries with much better resolution but at a small minimal cost. So, some of the vendors are digital globe planet labs and sky labs and so yes. So, although we have satellite imageries the problem is not exactly solved. Only for very well developed economically well developed countries like United States, Australia and UK the images are labeled. So, you know, for instance there is a labeling of images available for United States. Okay, this image, this particular region is a river. This particular region belongs to a house. This particular region belongs to a road. So, there are roadways available here. There are waterways available here. The labeling of data is available for only economically well developed countries at this point of time but we don't want to find the poverty level in very well developed countries. We want to find out the poverty prediction, do poverty prediction in African continent or in Southeast Asian countries. So, in these regions the data is not actually available. So, even if satellite imageries are available your problem is actually not solved. To be able to employ a lot of volunteers and certain do the labeling of data at some time and not a lot of money but some money and some time. And so, do we have a solution for it? Do we have something which can, you know, overcome this shortcomings? So, that's where this transfer learning approach comes. So, using transfer learning approaches you can actually overcome this problem of not having enough labeled data and there is one breakthrough paper from 2016 from Stanford University which is still used by many of the countries right now and even used by Indian government right now to do for certain use cases. So, Indian government, you know, they have to find out once in five years how many bills are there in a particular region. They want the statistics for some reason so even Indian government is using a baseline version of this, not exactly this algorithm but a baseline version of this paper and they have done certain optimizations about it and they are using it for predicting they need a statistics of how many bills are there in a particular region for every state. So, this is used by JD Insights and several companies and even in Sri Lankan government and in Manipal filipines. So, what does this paper say is they use nighttime satellite image since there's no labeling of data they use nighttime satellite imageries as a proxy for finding the economic development of it. So, this is an image of North Korea versus South Korea obviously, North Korea is not economically well-developed so there's less light if a country is economically well-developed then there's lot of light in that region. So, increase in the light luminosity corresponds to a country being economically well-developed and decrease in the light luminosity or a lack of light corresponds to an economic development of a country. So, being able to use a nighttime satellite imagery as a proxy helps them overcome this short problem of having unlabeled data and so, how many of you are from alright how many of you are from data science background and stuff so, a lot of them are from data science background so, yes so, whenever I visit conferences there's at least 3 or 4 people in the crowd who know data science so, how many of you are not from data science background? alright, there's an equal 50-50 number so, whenever I recently visited a lot of deaf-fest and stuff so, which is majoritly filled with college students there would be only 3 or 4 students so, the kind of analogy which I give for transfer learning is this, this is Chennai so, yes so, this is the analogy which I give for transfer learning so, this is from a movie called Papana Sam the main character in this movie is actually a movie addict so, he watches lot of movies and then he uses this knowledge to actually solve a problem in his real life whenever he is faced with a problem in real life he would actually apply this knowledge to his real-time problem so, there's a transfer of knowledge to a related task so, he's learning from a particular task and transferring knowledge this is what transfer learning in deep learning also means so, learning from a particular domain and using this knowledge to solve a related task so, the problem statement is divided into 3 sub-problems so, the first problem is training your CNN VGG 16 model on an ImageNet data set and the second model second problem is predicting an item light intensity from a daytime satellite imagery and the third problem is actually predicting this is how I split a problem into 3 types and this is a model where I'm going to walk you through the code of this entire thing so, the key steps are follow as follow is like downloading your satellite image from NOAA NOAA stands for National Oceanographic Atmospheric Administration this is free you can every 30 seconds as an image and available and it's there and DHS data is available for most of the countries in Africa not all, most of the countries and so far it is considered to be the best available you know database with statistics which they have related to the wealth of each of the locality in Africa and for nighttime satellite imagery you can actually download it from the Google Static API you can register for it and you'll get a registration key there is a daily download limit for Google Static API once you cross the daily download Google will return an empty it will not show you an error you will get an empty image and so, the three use cases which we are testing out is we are trying to test out the first use case we are trying to test out trying to predict a wealth just using daytime satellite imagery trying to find out how effective it is and the second we are just using nighttime satellite imagery and we are playing trying to predict a wealth and seeing how effective it is and the third technique we are trying to do a testing how effective transfer learning is in predicting wealth and the final output which you will have is this so ideally this is a wealth distribution map which will the economist or public policy research makers would need to make an economic decision on particular locality they would want to identify the locations with extreme poverty so these once the locations with extreme poverty are identified so just find using satellite imagery and getting micro level insights is not always necessary from satellite imagery you might be able to identify there is some vegetation here in that place but whether that plant is edible or not you need to do a macro level analysis but can you do a macro level analysis for the entire continent of Africa it's not possible we don't have kind of a bandwidth or logistics or this so you can identify the main regions where you want to send your volunteer tea so in Africa identify the 10 most poorest regions which requires immediate identification attention I can send all the I can invest all the volunteers and tea to those regions instead of like doing it for the entire Africa so this is not I cannot rule out saying this probably completely solve the problem of poverty prediction but in this in turn might reduce the amount or might ease out the problem but not completely solve the problem there is lot of research being actively done in this satellite poverty prediction using satellite imagery we are not quite successful with the accuracy or the prediction rate is only around 74 to 75 percent there is still lot of active research done there and a quick walkthrough of it this is the night time satellite imagery from NOAA and DHS will give me from DHS I will be able to get the wealth statistic of every locality I can try to find the clusters of wealth statistics I am trying to find the clusters of wealth statistics in this map so this is a quick demo walkthrough of the entire concept and the conclusions is less time left the conclusion of it we always found the transfer learning approach to be very effective compared to just using the daytime satellite imageries not very effective but better than the other existing methods using deep learning to identify the poverty zone is just one part of the pipeline so once you identify the poverty zone what are the innovations you can do in that particular poverty zone how can you minimize food wastage there is so much of the deep learning problem is a huge pipeline so that is only a tiny aspect of poverty identification or building a poverty pipeline generally this talk is presented as three collaborators there is another person who talks about AI innovation and third collaborator is a person who actually talks about economic and effective technology policy and this is usually done by Dr. Bhathri Gopala Krishnan he is an economist from University of Washington and unfortunately my collaborators are not here and this is about the talk so you have any questions hello we are running short on time so we could take only two questions the question is like how does this model work on as partially located areas and what if people are not placed in places where a lot of vegetations are around okay sparsely located areas so these sparsely located areas will generally be classified as an economically underdeveloped communities because in the transfer learning approach it actually the learns filters like road based things and stuff so it actually classifies as an economically poorly developed country or any area which is sparsely located and most of those sparsely located areas are poor in actual logical context if you see if they are very sparsely located that locality cannot have a district school or a district so logically a sparsely located area will by default be classified as a poor people so I don't think there should be a problem in that context so hi great work yeah thank you so my question is as you said accuracy is some issue like from the satellite data so my question is like can after pinpointing to a certain location okay there is something can the image is from drone can be used to further analyze the area and pinpoint okay this house or this perfectly this village exactly has the poverty issues I think there is a lot of active research being done on this macro level micro level diagnostics of a particular reason particularly from Duke University they have published a lot of research work I would recommend going through I am currently not focused on macro level but I would definitely say from whatever Duke university research is doing they have developed a lot of analytics based mobile phone apps and all that they are not sure of using drones but they are using a lot of analytics based mobile phone apps so there are a bunch of ones the area is identified as an economically drawn trot area there are a lot of analytical tools which they have developed using those analytical tools a bunch of volunteers could quickly go and keep classifying things so which in turn can help them out so there is a lot of active research being done in there but I am not sure of drones but drones is a very nice idea maybe is what I would suggest so is there any other question okay so we are out of time I insist everyone else to direct the questions at the speaker outside the hall we will be starting with the next thank you Usha for the presentation please give a huge round of applause