WEBVTT Kind: captions Language: en 00:00:04.160 --> 00:00:11.600 Thank you very much. My name is Misha  Chertkov, and I'm telling you about   00:00:11.600 --> 00:00:17.440 graphical models of the pandemic. Also,  I will be mentioning agent-based models,   00:00:17.440 --> 00:00:24.000 so I'm on the applied mass  side of the aisle. Let's go on. 00:00:24.720 --> 00:00:34.080 That's my kind of road map of activities which  we started, well, this pandemic, basically. What   00:00:34.080 --> 00:00:42.400 I'll be telling you about is based on those two  papers which are available on archive. Only entry   00:00:43.040 --> 00:00:51.920 of some portions of this road map is covered, and  specifically I'll be talking about inference and   00:00:54.640 --> 00:01:00.800 prediction and prevention of pandemics. There  are many other aspects which we are planning   00:01:00.800 --> 00:01:06.480 to go ahead with, and one reason for me to  talk here is maybe to look for collaborators. 00:01:07.840 --> 00:01:14.560 It all started with data, and we got quite a  lot of data on how the pandemic expressed. There   00:01:14.560 --> 00:01:20.160 are all kinds of biological, epidemiological,  geographic, environment mobility, which is very   00:01:20.160 --> 00:01:26.960 important for what we are doing, etc. data.  I don't mean to read this table literally,   00:01:27.680 --> 00:01:36.720 but I wanted to emphasize that it's an important  new ingredient for modelers like myself to start   00:01:36.720 --> 00:01:42.160 thinking, and you or for that matter, to  start new projects to model the pandemic. 00:01:43.520 --> 00:01:51.520 On the level of modeling, we talk about very  different levels of resolution and very different   00:01:51.520 --> 00:01:58.000 sources of information, in particular from data  which I mentioned, and different expertises which   00:01:58.000 --> 00:02:04.560 are certainly needed. My position today for  the purpose of this talk will be somewhere   00:02:05.360 --> 00:02:11.840 kind of in the middle in a sense that we'll be  talking about the geographical maps and geographs   00:02:11.840 --> 00:02:19.920 (how we call them). I'll not be discussing a  very aggregated model, what we call compartmental   00:02:19.920 --> 00:02:27.200 models, even though those were very significant  first steps in modeling the pandemic in general,   00:02:27.200 --> 00:02:32.480 not only COVID. By the way, everything which  I’ll be talking about is generalizable to   00:02:33.680 --> 00:02:41.040 other pandemics and other situations, actually,  not only viral but also social. I'll be mainly   00:02:41.040 --> 00:02:45.680 discussing what you're calling graphical  models but also will be mentioning agent-based. 00:02:45.680 --> 00:02:50.560 That's another diagram which  basically puts scales into the play.   00:02:51.680 --> 00:02:58.240 We are mainly interested in these projects on  a neighborhood scale, maybe a city like Tucson,   00:02:58.240 --> 00:03:05.440 or maybe a county, but of course this modeling  or kind of paradigm extends to different scales.   00:03:06.320 --> 00:03:15.520 We are predicting what will happen if there is an  injection of infection in a particular city, for   00:03:15.520 --> 00:03:23.760 example, through a super spreader, and projecting  now what will happen two or three weeks from now.   00:03:24.560 --> 00:03:31.280 They of course also mean not only to predict  but also to prevent. In the first place, I   00:03:31.280 --> 00:03:36.880 mentioned data. We want to learn parameters in our  model, so all of that is a one-to-one umbrella. 00:03:39.040 --> 00:03:46.000 Very high-resolution models are known  under the name of agent-based models.   00:03:46.000 --> 00:03:50.800 Before the pandemic started, we had quite  a lot of expertise on that in the world,   00:03:50.800 --> 00:03:58.320 but very few of those, actually only one, was  an open source. Now you see the list of ABMs   00:03:58.320 --> 00:04:04.640 (that's how we call agent-based models) which  are now all pretty much open source. We can   00:04:04.640 --> 00:04:10.560 all play with them and extend them. They  account for different effects like masking,   00:04:10.560 --> 00:04:18.480 quarantining, etc., and that's ongoing work,  which is very exciting and very important. 00:04:19.360 --> 00:04:27.600 Now we are also developing agent-based model  software. We are not yet public, but we are   00:04:27.600 --> 00:04:34.000 heading towards that. Actually, well, in there  too, we'll put a paper about that on the archive. 00:04:36.480 --> 00:04:42.880 Well, features of agent-based models. They're  basically the working horse of epidemiology.   00:04:42.880 --> 00:04:47.840 They're resolving individual people. We are  talking about the city of Seattle, for example   00:04:47.840 --> 00:04:55.440 700 000 people. So, 700,000 agents,   00:04:56.080 --> 00:05:03.840 it’s extremely heavy. You cannot model  and you cannot prevent this resolution,   00:05:03.840 --> 00:05:09.840 so you need to have reduced models. That's  what I'll start talking about very soon. 00:05:10.400 --> 00:05:17.040 Graphical models is one way of doing this course  graining model reduction. It's macroscopic as   00:05:17.040 --> 00:05:23.760 opposed to ABMs, which were microscopic, they’re  supposed to be efficient. They’re probabilistic,   00:05:23.760 --> 00:05:28.320 so they count. They are not answering  questions affirmatively but giving you   00:05:28.320 --> 00:05:33.680 estimations of probabilities, and they  are data-driven. There are various inputs   00:05:34.480 --> 00:05:37.200 and various questions you can ask, in particular,   00:05:37.200 --> 00:05:42.240 what is the probability of injection of  infection you happen to have as a threat. 00:05:43.360 --> 00:05:50.640 Here I'll put a very schematic slide, and it's  actually based on a paper which is very famous,   00:05:50.640 --> 00:05:57.920 very well known, but not in epidemiology, in  computer science. That paper discussed the spread   00:05:58.480 --> 00:06:05.360 of, well, misinformation or information through  the internet. Now I'm putting it in a little   00:06:05.360 --> 00:06:11.600 bit with epidemiological swing. Imagine  that I have this grid, and each node in a   00:06:11.600 --> 00:06:16.560 grid represents a neighborhood. In fact, that  one neighborhood is at moment of time ‘zero,’   00:06:16.560 --> 00:06:22.560 and red is infected. So, the rule of the game  is that I stay infected only for one step,   00:06:22.560 --> 00:06:28.960 and then I become black. Black means  removed, and otherwise if I am blue, I am   00:06:31.200 --> 00:06:40.240 susceptible to it. This is a probabilistic model.  It basically resolves through connections possible   00:06:40.240 --> 00:06:46.800 spread. You end up with a particular sample,  which is two colors: black and blue. That's   00:06:47.520 --> 00:06:51.520 a sample, which means that there is a  certain probability of this to happen   00:06:51.520 --> 00:06:54.720 depending on initial infection.  You want to answer the question:   00:06:56.240 --> 00:07:01.120 What is the most probable configuration, or what  is the probability of some particular infection? 00:07:01.920 --> 00:07:08.240 That's if I map to the city of Seattle. It's an  illustration of how this type of model would work,   00:07:08.240 --> 00:07:13.520 so parameters are now characterizing these  probabilities of infection between neighbors.   00:07:13.520 --> 00:07:18.240 You need to learn to extract from the data.  I'm putting it all on the rack. I'm showing you   00:07:18.240 --> 00:07:21.520 how I started. Suppose I have  an infection here and that's   00:07:21.520 --> 00:07:25.600 where I end up. It is a number of steps,  so one particular step is intermediate.   00:07:25.600 --> 00:07:32.400 You see that black is quite spread, but not  uniformly, and that's what we want to study. 00:07:32.400 --> 00:07:38.960 So, the model which describes this final  state happened to be a model which is known   00:07:38.960 --> 00:07:44.480 on the statistical side but also the physics  side under the name of ‘Ising model.’ It's not   00:07:44.480 --> 00:07:49.440 exactly the same. It's a graph which is a graph of  this connection between different neighborhoods.   00:07:50.400 --> 00:07:54.880 Those ‘j's’ represent strengths of  interaction— how often people travel and   00:07:54.880 --> 00:08:01.680 how serious, significant [throughout?]. Age is  a local bias. It is how much you protect it,   00:08:01.680 --> 00:08:04.800 how much you are masking,  what is the policy, etc.. 00:08:05.760 --> 00:08:12.720 Now, you can ask questions like I mentioned  before. What is the probability that infection   00:08:12.720 --> 00:08:18.480 spreads? Let's say half of the city of Seattle  three [days?] after this initial infection is   00:08:18.480 --> 00:08:24.880 basically infected (initial infection injection).  There are a lot of different questions,   00:08:24.880 --> 00:08:30.880 a lot of conclusions you can draw. You can  see that basically very often in this densely   00:08:30.880 --> 00:08:36.320 populated city, it goes from either everybody  getting infected or nobody, so there is a sharp   00:08:36.320 --> 00:08:42.560 transition which is called an applied mass physics  phase transition. You depend very much on data.   00:08:43.120 --> 00:08:48.160 Data is how you calibrate your model, and you  can resolve it not only on a city level. You   00:08:48.160 --> 00:08:53.840 can go to Wisconsin, for example, which is much  more rural in comparison with the West Coast. 00:08:55.680 --> 00:09:00.880 Again, different questions, but let me now  jump to what you can do prevention-wise. 00:09:00.880 --> 00:09:06.960 You can put these graphical models in this  prevention framework, and in the prevention   00:09:06.960 --> 00:09:12.640 framework you're basically asking questions  like “how I can change?”, “How I can introduce,   00:09:14.400 --> 00:09:21.840 and if I need to, enforce the mask  mandate if I want to maybe limit traffic?” 00:09:22.800 --> 00:09:27.920 Think about it a little bit abstractly as  this polytope in a space of characteristics.   00:09:29.200 --> 00:09:34.080 If I'm within this polytope, I'm green.  I'm good. If I'm outside, I'm bad,   00:09:34.080 --> 00:09:37.920 and then I need to project myself  back to this point. This is the   00:09:37.920 --> 00:09:41.680 type of mathematical formulation which  you have for this prevention problem. 00:09:42.480 --> 00:09:48.160 We play with that. What we care about there is  a development methodology. Methodology should   00:09:48.160 --> 00:09:53.360 be efficient, and that's what we are testing.  So again, methodology, but we of course want   00:09:53.360 --> 00:09:59.600 to be practical and project real problems, to  real, well for example, the city of Seattle. 00:10:00.480 --> 00:10:04.160 Quite a lot of staff work in progress.  I've been telling you a little bit about   00:10:04.160 --> 00:10:08.880 inference, but I didn't tell you much about  learning and the overall modeling pipeline   00:10:08.880 --> 00:10:14.320 from data, to high resolution, into low  resolution. That's what is ahead of us. 00:10:14.320 --> 00:10:20.240 This is a team not only from  Tucson but also from San Diego. 00:10:21.600 --> 00:10:22.240 Thank you very much. 00:10:22.240 --> 00:10:28.880 Lauren Close:  00:10:28.880 --> 00:10:34.480 Thank you, Misha. That was great. As a  reminder to all of our audience members please,   00:10:34.480 --> 00:10:39.120 remember to share your questions either in  the chat or hang on to them for our moderated   00:10:39.120 --> 00:10:44.880 Q&A session at the end of the presentations.  Florence will collect everyone's questions,   00:10:44.880 --> 00:10:48.400 and we'll talk about them at the  conclusion of today's webinar.   00:10:49.040 --> 00:10:53.840 Next, I'd like to welcome Amanda Leggett, who's  coming to us from the University of Michigan.