 Good afternoon, everyone. Can you hear me properly? All right. So thank you all for giving me this opportunity to be here today. I was a little confused as to whether my talk is actually meant for this audience or not, because I'm essentially an ecologist with an interest in big cat research. And I've used some of these tools that you're all talking about here. But at the end of the day, I'm a researcher. And what I talk about and its relevance to what you make out of the topics you're interested in is really up to how we communicate with each other. So just a brief biography of myself. I'll keep it short. About 15 years ago, I finished my engineering and I was like into the software industry like a lot of you. And I thought, well, this doesn't seem to be my cup of tea. And I'm a very outdoorsy person and I need to spend time in the forest. So it looked like wildlife research was something that appeals to my gut feeling. And I thought that's what I should be doing. Ironically, about 15 years later now today, I am spending a lot of time behind computers, writing code, but here's the best part. I think I'm doing it at one-tenth of the salary all of you get here. So it's jokes apart, but it's been a very enjoyable process. And I think being a data analyst, I won't really call it being data analyst in this profession. You have to be a scientist of which data is apart. It's been very rewarding. So the first thing I did while trying to prepare a presentation here was just go through some of the terminologies being used very often. We are talking of big data. Now, how does it apply to the field I'm working in? I mean, the last tiger I did measure was about eight and a half feet in length. Is that big data? No, no, not too sure. You're obviously talking about something else. Then there's something called data science. Again, this confuses me a little bit. What do we mean by data science? You know what science is? At least a scientist would know that. We talk about physics. We talk about chemistry. We talk about, yes, my field ecology. So what is data science? And then there's a subdivision data analytics. And there are also a lot of debates going on as to what machine learning is. How does it relate to statistics? And are they relevant at all? And some say that we don't need statistics anymore. And machine learning is all that we need. We have a lot of data. Put them into your machine. Get predictions. And you'll find the statistics is something of the past now. So I'm installed this confusion. I picked this from Wikipedia, but there are a lot of such flow charts even outside here that I did come across. I mean, and there are as many as the number of organizations out there as to what this data science process is. I felt this was again very confusing as to how do I talk about tigers and lions keeping this sort of conceptual framework in mind. So I said, okay, let's forget all that. Let me just go and stick with what I know. And I've learned science. And let's see if science has some role to play in the sort of work you do, in the sort of analysis you do. And what really matters to the ecologist as to why going through the full process of science is quite is more important than just focusing on data. So what is the scientific method? There are a lot of definitions, but broadly people will all agree with this. And this is something that's been there for centuries. So you observe nature, some phenomenon in the real world. You ask questions and you frame hypotheses. What we say it can be models. You're all talking about various models. You build some mechanisms that you think are actually happening in the real world. And what would you then do? You will design an experiment to see if the mechanisms that you have thought about actually are the realities or it represents reality in that way. Then you go out and gather data and after that you use these data to do analysis. And I think a lot of this conference has been focusing on this aspect, number six. And then you form conclusions because you do assume data are given to you on a platter and what you do with this. This is just broadly. And at the end of this you're meant to know much more. You're meant to have learned something new and you can make predictive analytics as machine learning does or you just understand better the situation that you are looking at. So as an ecologist, we are just interested to know more and more about nature. One variant of what we call the scientific method is what we call the hypothetical deductive method. All I can say is you reduce those number of hypotheses to just two. It's very powerful in medicine. So for example, or even like to see if a new drug is coming out into the market. What would you do? You would have some placebos and then you would have the actual new drug. You conduct tests. You find samples. You test it on various sets and groups of people. And you see if the new drug is working or it's ineffective. This way of teasing apart one hypothesis with the other systematically is very powerful in medicine. A lot of related fields that are very controlled in nature. And what you then get is very reasonably strong inference that, okay, this drug does work. This is what we need to know. But we have other problems in ecology which I'll come to. So in ecology, what do we do? We are just interested to go out to the forests and simply learn about these fascinating mechanisms. We really view nature as our teachers that we need to learn something from it and not the other way around. It's been a relatively young discipline about 200 years, I would say, unlike other sciences. And this is quite important. The growth has been quite slow in this field. But increasingly, a lot of ecological tools is now getting into mainstream business. After the 2008 financial crisis, a lot of ecologists were actually referred to and asked them to solve this problem. This problem that all the banks have gone bust. And we need help. We do not know how does nature work? Does nature have better mechanisms? So a lot of theoretical ecologists actually did come together and solve these problems. And very clearly showed what was pretty obvious, let banks don't become too big, have better governance, create modules. But these all came straight from theoretical ecology. So there is, I would think that if viewed carefully, you will see a lot of application in business. But this is something, this is a problem that actually I think originated in ecology. And I'm pretty right about this, that this process of gathering data, the emphasis we place on gathering data is extremely important. This is not like any other fields. And the reason is, just you take a look at this set of animals. You have tiger and its various prey items. And what do you see? Right now all of them are showing themselves up for you to photograph. But this is not something that usually happens. If you've all gone to the forest, you'll realize that no, actually they don't show up so easily to the naked eye. They are hidden in forest tracks. And we don't get to see all that is out there. So this is extremely important. And I think it is actually increasingly playing a role in a lot of mainstream processes of gathering data. So this actually boils down to a small statistical problem and where we need to estimate two probabilities. I'm not going to get into too much mathematics out here, but this is just to explain this idea to you. One is what we call the spatial sampling issue. This is something that all of you can immediately connect to. You've looked at opinion polls of elections. When you use data to say, okay, it's going to be BJP in power or it's going to be Congress in power. What do we do? We take a small sample of the larger population and draw inference. This is pretty straightforward. But in addition, we have this other problem of what we call imperfect detection. That even when you actually go to a place and look for animals, you do not see all of them. You miss them, even if you've gone to the right place. So it's like this screening effect is happening in our field, which is very important and which has had dramatic consequences. What you see here is pretty much one diagram that summarizes all the statistics that we know. Machine learning is all deeper inside is based on sound statistics, the theory behind machine learning as well. But all the standard statistics, you can pick up books in the market today and all of them will fit into one of these. But because of that problem that I did talk about, the screening effect, actually none of those books are useful to us anymore. And it is now thrown open a whole new world for us to explore. And so what it actually means is very simple. So we are saying that there is something out there called the reality. We call these the process models, which means like you are trying to build mechanisms. Are you trying to get at what causes accidents or what components make a good car? These are the process models, which is the reality. Then you have the observation model, which says actually the data you get probably does not directly reflect the process itself. We are getting a filter and that filter we have not paid attention to. So the moment you remove the filter, you can end up in trouble. And I will give you some examples about this in this talk. So again, don't get too bogged down by the mathematics you see here, but I will very quickly explain this to you. So this is one very, very simple model that we have used mainly to study animals as charismatic as this. And this is what we call the hierarchical model. And this is what we see, the process model. See we are interested to know how many animals are out there. Now this can be some other problem. You are actually looking at at any point of time, if you go to a railway reservation counter, you want to know how many people are using. Should we keep this reservation counter open here or should we shut it down? And every time you go, you're not going to get the same outcome. It's not that you're going to find exactly five people out there. You'll find zero at some times, you'll find eight at some time. So such processes are usually modeled using the Poisson model. We can use any model. This is just an example and it's got some average intensity rate with the outcome. Every one pick could be five, could be eight or any of that sort. Now over and above that we have something called the observation model. And here you might set up a survey to say, look, I'm trying to open a business in, let's say, Bangalore. I know Bangalore has a lot of software engineers. It's quite likely that they may all require glasses or contact lens. And I suspect that they do. One way to do is like set people out there and to just make an observation to see if these people actually wear contact lenses or not. Now here is the very tricky issue that becomes extremely important. So you just set these people out there, they are gathering data for you. They're just not going to tell the people or ask people because people may not tell the truth. They just observe and they say, I saw this person with contact lens or not. If you have not deployed the right people, their ability to detect this process is going to be poor. And here if you see this definition, this is a very, very interesting probability that you will get. You do not know whether some place that you have set a counter actually has a situation where people who come there do not wear contact lenses or whether they did all wear contact lenses but you could just not detect them. And a lot of data are being gathered in this form. Now that is extremely important. If this process is not taken into consideration, definitely in ecology of course we have seen that it's very problematic and I would suspect a lot of the real world problems in the data that we are talking about here also will have this innate problem. So my focus of research has primarily been with tigers for all these years and off late I've been looking at some problems with facing lions and cheetahs as well. And a big deal in all these problems is of course just to know how many there are. And I'll give you, I'll take up four problems that I will illustrate the central point to you and you can see if it actually has some connection to sorts of data you are working on. So one is, well we all know, you know large carnivores like tigers consume a lot of prey and it's pretty, pretty important that you do get this right. It's pretty important that you know how much prey is out there. So this is a problem that we are facing in the tiger conservation world. In about 1.1 million square kilometers of potential tiger habitat today most of this habitat contains animals in very, very few numbers. So one problem was to see, okay let's assess if there is enough prey in all these places. But since even they are really, really depressed conducting studies to actually get these numbers are very tricky. So what do we do? One is you need to carefully design experiments and like a lot of you I would use computers to simulate sort of field survey a data gathering mechanism and evaluate models to see where they perform the best. Sometimes if your survey is bad, now a lot of you are actually assuming you have very good and sound study designs in the data that you are getting but that need not be true. It's probably just people who are giving data to administrative organizations that is coming straight to you. And then we use those simulation results and go straight to the field. This is an area in Bhadra National Park. So we surveyed nearly about 200 to 300 square kilometers. We look for these signs. So you can see ecological data. These are signs of tiger prey and you see it's not very easy to detect them. And it's also you don't find them everywhere when you do the survey. And using this technique the idea was to try and estimate abundance. And it was quite successful and quite well utilized all over Southeast Asia hands. But importantly and this is where I said there is this cross application that immediately happens. Someone approached me to say we have the same problem for detecting malaria. And it's the exact problem that you actually collect samples blood samples from people and conduct clinical trials and actually check to see if a person is infected or not. Many times it shows up as no this person is perfectly fine but actually the person is infected. So how many clinical trials do you need to actually get the correct result? And this is going to vary on a lot of things. It depends on the immune response of the individual. It also depends on how well you've done the test of course. And so here is an example of how actually a problem that was used to count tiger prey got straight away used for malaria. And I could quite personally relate to the situation because I did get typhoid twice in my life. In both times the test didn't reveal that I did have it. It just happened to be an experienced doctor who just looked at me otherwise and came up with a solution. So there was no data analysis. It was just okay we just took to take decisions based on no data. You just take it on other evidences that you have. So then that was one example to show how this idea is moving to other fields. And I hear now that the detection probability problem is now moved into the political field as well. And people are trying to estimate this detection probability as well as trying to get information on the data gathering mechanism. Another very interesting problem very recently is to do with counting tigers at national scales. And so way back so up until say 2004 what happened was the Indian government used what we call the pug mark senses. These were data. Individual different individuals in the forest would be recognized by the footprint by the pug mark left on the on the substrate. Now you can straight away see this is not very easy data. You can straight away see the depending on the substrate type you're going to come up with different numbers. And because of this sort of bad method what happened was tigers went extinct in two very key reserves in India in Sariskanth Panna. And this monitoring system could just not pick it up. Now there is some something very crucial that that I want to point out here. So later on where in 1995 a much superior approach was was studied. And this was the superior approach where you set cameras along trails and you get photographs of tiger individuals. And what happens then you can identify these tigers tiger individuals by their strike patterns. And if you place many cameras this is about a 600 square kilometer area. You will get many individuals and you can count them better. So you just improve the reliability of the field approach itself. But still this is quite expensive and you can't do it at national scales. You can't do it at select reserves probably that's enough. But we are quite ambitious you know we want to come up with a number for the whole country. And so then what happens so the idea was like OK let's see if we can form a relationship with true estimates. Because we have data from one source that's giving you closer to the truth. And then we'll use this what was sort of a bad method but we'll refine it. We won't try identifying individuals here we will just count these tracks and these signs and we will see if we can form a neat relationship. And typically this is how it is so you this is the problem. So you have some select reserves here where you do this sort of calibration exercise. And then you look at a larger landscape to see can I then use that model. And use the cheaper method and then come up with statewide numbers or India's tiger numbers. And well this is quite obvious to most people if you find a relationship that looks like this. Statistics we get something called a high R squared value. You will say oh well this is really good we can it's a it's a great approach. But if you do get a relationship like this you can straight away say it's bad data. We have nothing to tell the world. And now this was very interesting. So there was this one there was this experiment done. Now if you look at this relationship you can apply any machine learning technique to this. This these data come from 19 to 21 different reserves in India. It's very expensive to collect these lot of effort has gone on the ground to collect these data. And now what machine learning tool will you apply to state whether we can use this relationship or not. Now one technique that that's part of the machine learning literature is what we call cross validation. Have you heard of cross validation. So what people do one form is you take one data point eliminate it and you find how good the relationship. You take another data point eliminate it find how good the relationship you do that for all the data points by removing one after the other. And you will come up with some idea of your prediction. This is your predictive analytics right this is exactly what you do in machine learning. But here's the problem you're just relying on this relationship you're just relying on these data as you see them. You're not worried about design you're not worried about the mechanism causing this. Now just see what happens if if we pay that one little extra attention that we had missed previously and let's start to assume. So you can already see here that if I eliminate just that one point that seems to be something called an outlier people people use this word quite often outlier. But that's not what we do in contemporary statistical practice or even machine learning practice. If you looked at textbooks in the 80s and 90s they would say if you find data and you find an odd data point just let's remove that so that the relationship looks good. But today we say no we are going to exactly find a new theory if every new data point tells us a different story. So we need to go back to some theory here. Now as I did explain that the same two step model is a very similar model occurs here and we get some sort of counts and then we have this thing of detection probability. Now you would have never seen that detection probability in in that graph out there you would have just gone tried your various tools and seen which works best. The computer will tell you how good the relationship is. Now if you did just a little bit of theory and worked out how these data are coming you get a completely different story and theory will now explain why you can get such data. And in the process when you actually tried out eight other data points this was how that same relationship looked. The problem was we were eliminating data points and we went overboard with one relationship that probably was not defining the truth. And later on when you when you worked out the theory you come to know that actually no we have a much bigger problem here and the performance or the predictions that we made was not was not good enough. So we came up with conclusions and it became very important to state that again here you need to estimate detection probability otherwise it's not worthy of your time. So that was one problem to do with how machine learning makes sense. And next I take up a sort of third problem. This might actually bear closer resemblance with with a lot of the work that I've been seeing happening here. And that is to do with pattern matching and pattern analysis. This is something you like to do all the time. So this was the problem. So when you did camera trapping you get a lot of individuals a lot of tigers and when you have a few in number you can actually just visually see them and identify them. It'll be fairly foolproof and you do a good job of it. But if these data sets grow in number you start getting five hundred six hundred thousand thousand five hundred and these animals also live for five years seven years. You start seeing the same tiger appearing every time you get a new tiger you need to start going back to years of data and start matching. It tires the eye and you need computers to help you. So a researcher in Cambridge who some of us collaborated with came up with this pattern matching software and the idea is quite simple. This I would say it's somewhat it's it's it's it's like a hybrid between complete machine learning and applying some human intelligence is not completely leaving leaving it to the computers. So this is what happened. So I guess a lot of you are engineers out here and you've done perspective projections you've done engineering drawing. So if you do have a little sense of perspectives actually this approach works very very well when you use computers to help you. So here we do have we have a tiger and as you can see what what we have to do for this to extract this pattern like this is to actually imagine rods going through equal points on the left and the right flank and when that is done and then you mark the above what happens is you will this pattern matching software is able to construct a three dimensional model for you. Now this is very very useful. It got a lot of use in conservation because now you could even find skins that when poachers have caught tigers you could relate them to the tigers in the wild which helped a lot in forensics. But again importantly here is why will a pattern matching software help you in this regard like this. Obviously you want it to be foolproof but it cannot be foolproof. Finally you still have to use use our own sense and see if they've actually matched. And so there were two algorithms applied here. And don't worry about the equations again. You can all I wanted to say these are the two algorithms and what we what we really need is when these two algorithms are put into place. You need a statistical way to say hey I'm going to use information from both these processes to improve the patching pattern matching ability. Then then so most of these problems I've heard people talk about Bayesian statistics. My colleague salmon will talk a little more about the details of this. But essentially this is a very very powerful approach that's now being used where information from multiple sources are multiple sources are coming together and you're able to better do the job. And what you would get is a score you know and this was very interesting. This there is a there is an element of learning here and this is what we want machine learning to do it's a little more than a database. So as the number of images go up as the number of images go up your probability of matching starts going up with that individual because you start getting many representations of the same individual. And here this sort of defeats the human ability and it was so useful for us because we recorded two amazing dispersal events. Dispersals are tigers when they are about 18 to 24 months old the males tend to move away from their resident areas and move long distances. And we found two of these long dispersal events thanks to this combination of a well maintained database linked to this pattern matching software which we would have failed otherwise. This is very very similar to the classic example that you hear on about chess right. I mean every anyone any time you hear about artificial intelligence people refer to the 90s when Gary Kasparov split the chess world world championship world into two and he took on the blue at that time. A lot of debate on whether man can beat computers or computers will beat men. This this went on today we don't ask that question anymore because today's world number one Magnus Carlson will lose straight away to the best chess engine out there. But there's a small point to note here that even today even today if you remove the opening theory from these chess engines humans are doing pretty well. And I think that was this this this application very closely resembles that because there's a power of the database which is helping you in the process and more powerful computation to work out the combinations that's helping you. And finally I come to this topic that I've been looking at very recently these are what we call spatial capture recapture models estimating density of big cats. Since working on this problem I've met a few people in this field and they said they find a lot of use for this. Of course it didn't make sense to me because I was just focusing on its approach to count tigers or lions or cheetahs. But actually it seemed to have a much bigger application than I did conceive of. Now here's the big deal that all carnivores and especially large felids they tend to be very territorial in nature. And so they distribute themselves in space following certain rules certain societal rules that we are not too clear about. But they do form territories they are more closely associated with space. So for example you have a tiger territory here in this part of Bangalore it's very unlikely that you will find the same tiger let's say in Malaysia for example. Of course we are working in different scales here. Now what again what's important to know is most of the statistics that you have learned doesn't account for this. You need to again create statistics from the start build models right from the first principles and then confront them with data. Fortunately now we have these abilities. So again don't worry so we use the same approach. Some realities out there animals have distributed themselves in space and then what we get is a thin down set of data. And with the thin down set of data we need to form conclusions of what's happening above. And here now we can due to these Markov chain Monte Carlo techniques which Sauman will talk about in stock at four. This is a very very powerful tool. But it does come with some problems. We tend to build very complex models to explain the phenomena. But it takes a very very long time when we use the computers. And typically one analysis takes four days for a tiger data set. And but we get very very interesting results. We are actually able to say where which part do we get very high tiger density is where do they occur. And and where should we focus our attention in trying to say recover tiger numbers. And someone told me that they're using a very very similar approach in Scotland yard to find thieves. And how does that work. It's because no matter what you do somehow these tend to have some sort of territory. And when their crimes they detect like each each each robber has his own ways. So when they pick up the clues they're able to say it's from this one person and then they boil it down. Now again if you look at these data you're not probably picking up all out there. You're going to get a sparse sense of of these people. You will need a hierarchical model like this. And later I also found actually we took this model and we converted it into a very simple software package that that people could use. Regular ecologists could use. And it came to my surprise that in Sweden people were using the software to actually take tools to show where dolphins are. I said actually this sounds extremely commercial to me but but then but then that's how it is. So ecological research is increasingly finding commercial applications is what I find. The problem became something else when I started looking at cheetahs. This was a collaborative study in Kenya and cheetahs are quite different from tigers. They move much more. They move much more relative to their body size because they avoid lions. And in doing so you see we drove about 8400 kilometers in the Masaimara reserve and we did estimate but this would take a computation time of about 15 days. And this was the time I was forced to actually use high performance computing. And it made a lot of sense because though when we use MCMC techniques we are a little limited in how much we can parallelize operations because the operation itself is meant to be sequential. We can run multiple chains multiple analysis on high performance computers that but the bigger deal is as as field biologists. We want our petty little computer to be free of computation. You don't want it to do that task. So if you are in the forest and you look and you're working here in the middle of Masaimara how convenient it was at least for me to get all data here. I would just send this off to the Oxford supercomputer. It would do all the analysis for 15 days. Let it take as much time. But I think that culture I do have some ecologists in the audience here today. I think this culture of transferring this whole issue of computation to these increasing sources that we have. We now have a lot of commercial sources giving it out for cheap. And I think it's good to explore here also. It's it's really helpful that we get into this culture of shipping our computation to some place where you don't have a problem of power failures and things like that. And then also looked at lions. This caused more. This was even more tricky because you obviously have to go very close to lions to identify individuals. We use these spots here by the whiskers to identify them. But we're already running into the same problem as we did with tigers. And we have no solution to this because you see a lot of lions in the Masaimara. And for every one day we sample out there you come back and you need to spend two days just trying to identify individuals now. And we don't have any pattern matching software for this yet. This needs to be developed. And I think that's that's that's very useful and that only then will we get any information on their populations in the Masaimara. So this was that that software that that some of us developed that runs these fairly complex models. But it's got a user friendly database and user friendly setting. So anyone who knows how to use Windows little bit of our can can run their analysis. It became surprisingly very popular. And this is again we are all really not computer programmers. You know, so we are not good at optimizing things. So they all run very slow. We are researchers at the end of the day. We just wanted to work. But I think there's a lot of scope for to optimize these things and actually use them. And it will really benefit our field. So my probably just my take home messages here. I think there's a great advantage of sticking with the formal scientific approach. I know a lot of you are data analysts here. But I would probably encourage you to get into the process of also seeing how data are coming in. And when you do this, you probably will be surprised by the answers that you get. It might be quite different from the analytics generated by machine learning that you see machine learning that you see now. And as scientists, we are really interested in the underlying mechanisms. Probably machine learning people are not too interested because they just want predictions. But they are two things that you cannot separate. They are closely tied to one another. And if you miss one, you can really misinform decision makers as we saw with those tiger data. And I think it calls for a lot of skills to help wildlife. And I think machine learning increasingly has its role. As I said, there's already a problem out there for Lyons. And I think it takes a lot of us to help these charismatic species. And I just hope I'm first very thankful to the organizers for having an ecologist here. I hope this becomes a very regular feature of the fifth elephant where people talk about a lot of other problems. And I think there will be a lot of synthesis in the approaches we take. Thank you very much. I'll take any questions. Questions? No questions? Are there too much maps? Oh, sorry. So, Boston was going to ask a question. Characterizing science and the hypothetical deductive things. I would say that part of science is reduction of arbitrariness so that you find hypotheses that are fewer and give you more results rather than ad hoc this and this and this and this. General principles, which is very much the line that you're talking about rather than just throw it all in a big data heap and let the algorithms run. And along this line, one could say that Panini was the first real scientist because he had a theory of grammar and it was a saying that the grammarians when they eliminated one syllable from the slokers of Panini's theory but still got the same result, they celebrated more than the birth of a son. Reduction in arbitrariness, getting the conclusions conveyed in fewer principles is very essential. And data analysis can only look for the sort of patterns that it's taught to look for. And the methods behind it have to come in to make the software better because it should be capable of recognizing more patterns, more subtle things. It's mostly still really statistics as you're... and the non-linearity of the real universe is about too much. So it can only get so far with basically statistical models. So I've talked too long. Did you have a question? Yeah. Can I... Hello? Question? Yeah, I had a question. Animals need to... Yeah, it's here. Yeah, obviously regarding the estimation of... You mentioned that estimating the lions is difficult but obviously animals need to eat, they need to drink water near a pond and they also have excrements. So is it possible to look at the forest floors, scrape it, sample it and do DNA analysis on that to find out the density of animals, like find unique animals or something from their DNA patterns? Yeah, can you hear me? Sorry, this thing is slipping. Yeah, in principle we would ideally like to identify individuals however they come. Now you spoke about DNA and we do use that for tigers as well. We do use that for a lot of other charismatic wildlife as well. But even with DNA you cannot just take data as individuals without going through the careful detail on how these DNA data are being collected. One approach that people have used was something like you take regions of the DNA, we call them the microsatellite loci and what you do in these places it's like you don't get all the information. They are like a set of piano keys and you get tones among them. So you have C, D, E, you eliminate D, you have E, you have C and you need to guess the complete tune from these sorts of data. The idea being that the larger you increase and you repeat that you are able to fix these patterns. But even here you will have to build the probability statements like I showed you within that model so that you are sure of what level of uncertainty is there in the individual ID. So DNA is used. We are increasingly using it. We are also using newer technologies now called single nuclear polypeptide approaches, SNPs they call. So yes you are right. So any form of individual identification will help. I have a question. Yeah. Oh sorry. Does your study have any implication on the conservation of these wildcats and helping anti poaching in some sense? Helping anti poaching. Well, cameras also detect some poachers. But whether that actually converts into something in the real world whether it actually helps in booking cases and making an impact that some people have done in certain parts. But if you are talking about tigers those examples that you saw there were instances where actually the skins that were seized by the forest department claimed by poachers who said like no this never belonged to the wild. This was something that was inherited to me. You are able to fix all those false claims and actually use it in a verdict in court. So that way it has been useful. Hi. So I also have with me Dr. Ula's Karanth here who is absolutely a legend when it comes to tiger conservation issues. I hope he has something to say on the conservation aspect of that. Just to make the point that these metrics, these measurements of numbers reflect whether your conservation is succeeding or not. When you plot a variety of parks and you are seeing densities of 15 tigers in Nagarhole versus three in some other place whereas ecologically it should be 15 in both places clearly you have a tool to identify where things are not working and then go and fix the problem. So the methods that he talked about have two purposes. One is to answer questions, basic science for the sake of advancing knowledge but to us more importantly the same answers can tell you where there is a problem and how to fix it and it's through these studies across the country that we figured out the idea that the tigers range had shrunk by 93% in the last 200 years and the prevailing dogma in conservation in 1995 till we did the studies was that it's the Chinese who are eating the tigers and that's what's compressing the tiger range and we found actually it's the depletion of the prey because we found a 1 to 500 correlation between tiger and prey numbers once we had the methods and it very clearly showed that entire business of tiger numbers was dependent on prey species being abundant and the prey in turn were not being eaten by Chinese they were being eaten by our local villagers so what you needed to do was a very different solution so these tools are very useful for advancing conservation. Hi, you had two models, one is a process model which has a Poisson distribution and other is an observability model so the probability in the observability model you have you parameterized A and B, I think it's A and B right I was wondering what exactly do they match to I would have thought it would match to the vegetation or the landscape and were able to match the parameters to something physical in terms of landscape or the density of the actual prey or the you know predator So were you particularly asking about that beta KB? Yeah, see for that that was a different problem it was not like the others and there it was just this when you sample you are assuming that the sampling you are doing is random if you do not take that randomness into account we will get various values for those parameters A and B so what those parameters A and B really are doing is what is the fluctuation in this unknown detection probability that's out there that you want to utilize or that represents the reality so I gave you two examples the one example where you did filter out data it looked like there is no not much fluctuation in A and B so you would assume in a simplistic sense if A equals B in the beta distribution and it's very high you will probably get such a relationship the moment you added a few more data points the truth out there has A and B that is completely different the important thing to note here is you should estimate those A and B first if you have a process that explains this for example if you go out and do tiger surveys very clearly you will see that in reserves you have dusty substrates you see tracks very clearly no doubt they are also very high in abundance in these areas the moment you go out outside these reserves the substrates are very hard and you don't see the tracks but even if there were tigers there you would have actually not recorded them accurately enough using that method then what happens is you have this problem where abundance and detection are related to one another you are not able to tease them apart so you need one of those process models and observation models they don't actually work independently they work together and you design your survey in such a way that you tease them apart so the point I was trying to make here is how to tease apart these two processes that will then give you much stronger inference so the resulting data you get is not biased or it's not being over presumptuous on how much is the variance around if it's unreliable you have to say it's unreliable but usually it doesn't work like that when you don't take these two processes apart thank you thank you Arjun