 It's a pleasure to have you here today. I know the faces I've been seeing since last week, and four of you have been with us since the current thing, the last week, right? Where we had Dr. Carpenter, since before some of the people in here didn't quite know what to do with Python or with Git, so the few things that we find very fundamental within data science. And some of you were not here yesterday? You guys? I'm a bit closer. If it's too loud, let me know, as well. Great. So what happened yesterday? I took some notes here, so we learned some fundamental tools to analyze matrices, the basis of representation of image data, right? We went through NumPy, SciPy, to ending up with how we can clean pictures, remove artifacts, noise up to how we define convolutions, how we put together feuders. We learned about Gaussian feuders. And then we ended up with the image partitioning using algorithms that allow you to recover borders, as well as ransacking which Stefan introduced to some of us as well. So what's today? Today we're going to hear about problems and how we can use some of those building blocks to construct the data analysis pipelines. And what's this talk about then? Well, at this talk, I'm going to elaborate on what's this face detection to faces of a scientific image. So what is a scientific image? Let me start with an aspiring scientist here. This picture is from 2008. And although this girl now is 15 years old, she has no pimples and a few accidents on her face, stitches, we can still recover her face really quickly. And given that, we can do such a great job with recognizing faces. This computer vision is solved. So we're basically wasting our time here. It's all done. In fact, there is a lot to do. And what I want to explore with you guys today is what if we could do such a good job that they have done with the faces, but instead look for the faces of scientific data? Wouldn't it be wonderful if we could put together a because of experimental data? Somehow, automatically retrieve those characteristics or faces of our experiment that could be brains. There are particular two more in the brains, and I would like to be able, just by feeding into several of these images, recover all those images that are very similar. That's one of the tasks within computer vision. And hopefully, at the end of this presentation, I will have shown you a few of the tools that allow to get much closer to the because of experiments. All of the problems within computer vision these days is definitely the data explosion and the need for automated processing. We had lots of upgrades in the detectors, particularly within Berkeley labs. We hear every year that there is a new instrument that has a better CCD, so pictures that are even larger than we had before. Not only that, the ability to record these images really quickly is also increasing. Some of the problems actually is a combination of these two requirements. We got it to automate. Once upon a time, we would hire summer students to be just manually segmenting images. Well, they don't want to do that in first place anymore. And second, this is just too time consuming. We need to come up with better tools to do data curation. So what I'm going to tell you today is we're trying to improve whatever are the tools to do with the image across domains, both at the realm of data curation as well as image understanding. I'll talk to you about some problems in material science and geology, in chemistry and health, in trying to make sure that I'm telling you the story. What's common among them? There are different science questions associated to them, for sure. But are there algorithms that we can use for all of them? And hopefully, I'm going to have time to tell you a little bit about the future. What does it look like dealing with a hybrid data, hybrid hardware, and hybrid expertise? So just to get started, my presentation will work sort of a top down. I'll go really broadly into what they have in common to then drill into the specifics of each of these science domains. Just to give you a flavor of some of the things that my team and I have been doing at Berkeley Lab and now here a bit the past two years, ranging from the scales of the nanometer to the kilometer scale here on the x-axis and a y-axis. Some of these phenomena happening really quickly at the order of nanosecond up to taking perhaps years or a million. The first example here, and of course, I'm showing when all is solved. We could red separate what was the internal part of this nanoparticle in the outer layer. This is a combination of a platinum and a palladium. What we received was a bunch of STM images. Those are gray-scale images. This is a stack, and we need to recover where are these nanoparticles to do calculations, for example, of the geometry. Depending on how the outer surface is organized, this particle will have a specific application for one thing or the other. So our work here was doing the segmentation and using technique called front propagation. And the same technique was also used for this. This is a collaboration, oh, by the way, nanoparticle collaboration with NANOSAM, which is the National Center for Electro-Microscopy. And the second one, collaboration with the optometry department. Picture of the back of the eye, completely different, right? The first, we're talking with the chemists, and here we're talking with the optometry. The problem that they want to solve is, can you tell that this retina is associated to retinopathy or not? My job, as well as my team, is to recover where are the vessels of the back of the eye? Again, what we're using here for this problem was just coming from internal seeds and then propagate outwards. We can define it in this case because we have the seeds, and these are very thin. They are monotonically increasing outwards. Second example. This is a collaboration with the Helen Wills Institute, and the question seems very simple. Can you calculate the volume of the school? Well, if we had the school it's empty, we just put it in the bucket of water and eventually we'll do that, right? Except these people are alive. We're using in this case CT and we want to recover. It seems that there's a pretty good sharp edge and we learned about sobelly yesterday, right? We can use that information to tell what's inside and outside. But one of the problems that we had, particularly with the front propagation that would escape the eye, so there was a lot of further work to close the orifices of the school so that we could calculate the school volume, yeah? Sure. Awesome, yeah. Better now? Great. Yeah, let me know if now it's a bit too loud. And the last one here at the kilometer scale, the idea was to find where are the vessels on a very grainy looking image. The graining associated here is a kind of noise called a speckle noise. Again, using front propagation. So for all of these we're using a single segmentation algorithm, if you will, except in this case it's not only monotonically increasing, it has the chance that the front will propagate outwards and inwards. Okay, this is all great. There's even more science that we've been doing at LBL, our different teams working in different domains, but what's that? I mean, how big needs to be a team so that you can go from radar images to nanoparticles to chemists to neurology? Actually, we're a very limited. And we're even looking into lots of different information. I think one of the key things is definitely participating in multidisciplinary meetings like this one, get to learning a little bit of the science domains, and work very close with the domain scientist. So that we can better understand chemical, electronical and structural properties of these different materials. And by doing that, it's quite amazing how much we can constrain the computer vision problems. And we get to the details of each one of them. So what I did, I divide all these problems that I showed on the previous slide into energy and health related. Just in a nutshell, what am I using? What kind of images are these? So these are the specimens I'm working, for example, with the geological specimens. We're using micro CT, collaborating with the people that are actually working with Venkat and Lewis that are here up front. And the goal is segment phases. Ceramic composites also with tomography, detective fibers, with this particular composite using another instrument of LBL to quantify porosity. That's the goal that mostly we get from the domain scientist. But as a computer vision expert, what we're trying to do, we're trying to do segmentation. How we can put these different parts apart? Let me go to the first example. In here, what we receive from the domain scientist are 3D representations of an specimen. This is a ceramic composite. There are several sections associated to them. I'm showing just a few of them. But believe me, we've got to process the whole thing. Some of these stacks may have something between 30 to 60 gigabytes. So quite challenging to do that in our laptop. Why anybody would try to study this material? So just to give you a flavor of what the kind of science behind it, they want a very tough material. So what they did, they weave it in these fibers inside the material. And then they put pressure, different conditions of oppression and temperature and see, when is this material going to break? This is a collaboration. Between LBL, UC Berkeley, NASA, and the DOD. They don't want to build an airplane and see, well, does it fall or not? You want to check the material first. So that's what they did. They put in different conditions of pressure and temperature an hour ago is to separate the different phases of the material. In this case, the matrix itself, the fibers inside, and in blue, when this material is going to break. It matters. And also track these breaks across time. There are measurements that we need. Visualization is great, and we love that. We're part of the data analytics and visualization group. But the domain science want the measurement. Can you give me the crack open displacement associate to these different planes that I'm showing here? So here's more about the work. What we're trying to do is to quantify this microstructural damage associated with these materials, 3D stacks, evolving in time, using a combination of algorithms. And some of them have a lot to do with one of the algorithms that we saw yesterday, which was the segmentation one that uses the region adjacent graph. So that is a key in one of the algorithms that you use here, which is the statistical region merging. You construct a graph, construct the super voxels, if you will, and then go merging according to similarities across the different nodes that represent your graph. The theory is great, but we have a lot of push from our domain scientists to make the software available. So I'll also tell you, could you test the software? The answer is yes. We are creating software tools and also with an attention to interfaces. So let me make sure I tell what is the main advance of the technique for that. Absolutely teams. And you can see some of the people in the team in here. Tolita Perciano is in here as well. Hari Krishna and a few other folks at this lab. So these are the computer scientists working in this project as well as the material scientists, people that are designing what it is about this new composite and the person that is acquiring the image at the tomographer. It seems like, OK, this is a great story. But at the end of the day, there is so much that you can talk to them. Well, then you're going to spend much more time modeling your computer vision problem to solve such a generic work that didn't need to be. In our case, here are the fibers. These fibers were unidirectional. Oh, this thing is very simple. So why don't you take that into account? The cross sections, when you have a cross section of those material, you have little circles. So perhaps you could even use in that ransack algorithm that we learned yesterday to look for several of those circles. That would help you. So absolutely, learning about the domain problem, we will allow you to come up with algorithms that are more efficient. And you've got to be. You've got to be timely. That's the problem. Interpretation of the data in a time before the user that came to the beam line and went back home. So how it works, people come to the instrument, they have a few hours at the instrument. And after that, some of them need to go back to different countries. It would be great if they could have a feedback, at least that they were able to capture the mechanical phenomena that they are after, just after they do the experiment. We do not have such tools. It's up to you guys, ourselves, to construct those tools and make those available. The tool. What's about these is about one of the tools that we put together is called F3D. This is a plugin that works inside a tool called Fiji. And you may be asking, well, we've been hearing about Python now these days. Could I use that within Python? Yes, you can. And I can teach you. I can even put in our etherpad how you could call macros built completely in Fiji through Python. So you let me know if I forget that. For those that are up, so right now, all of you know how to play with GitHub. That's great. You can get cloned there and be a collaborator at that. We didn't solve all the problems. At the moment, what we did, we worked really hard to speed up the processing of the several nonlinear edge preserving filters as well as a bunch of morphological operators so that you can do a much faster job. How faster? Well, we have users complaining that we're spending the whole weekend doing one filter. Very similar to what we were doing yesterday with that Gaussian filter. There are other ways in which, instead of blurring the edges, you preserve the edges. One way was actually introduced by Professor Malek of its spironomalic filter that helps you to preserve edges while you're smoothing homogenous areas within particular grains. One of the filters that is available in this package is the bilateral filter. What kind of hardware we're using for these tasks, in which I'm saying we're processing 100 gigabytes in approximately 40 minutes. It's not even a supercomputer. This is just a cluster with a few TASLA cards. This is our GPU cards and allowing then really real time feedback, particularly on the processing. Enough of the ceramic and matrix composite. We're going to talk about a different science domain, but you're going to recognize algorithms that were used in this first problem, also in the second problem. Why somebody would work with films? When we imagine films, we imagine the spasm very thin, very smooth, and then I learned that wasn't really that way. Not only that, they need to understand about the porosity of the film. It was very important for me to understand how the process was performed so that I could design the descriptors for exactly that problem. And I boil down to a very simple explanation how to prepare films. You started with the porogens, which works like a baking powder. You put in a block of copolymer, which, you know, just the dough in here. And please forgive me, all the chemists, if we have chemists in here, by disimpliating this block of copolymer is a mix of polyester and a few other chemicals. So you mix that block of copolymer and come up with a perfect PMO, which is actually the prime material for the film itself. PMO stands for Periodic Mesoporous Organicillica. Why would you want to memorize this name? Well, periodic, something is repeating, right? Mesoporous, these pores are sort of a large in comparison with how thick this material is. So interesting to learn is a few words so that you can establish a conversation with the main scientist. And who is at the end of the line? We have Intel, and I hear it with a, does it have the right properties? Well, because if it doesn't film, it doesn't have the right properties. You cannot constrict the micro-electronic with this material. Where our teams get in, we get in for the quality control in collaboration with Anne Sim and Molecular Foundry. These are two laboratories up Berkeley Lab that have the instruments. They can collect pictures really quickly, but they cannot process it very quickly, right? Then the computer vision come to the game and try to design features to help understand what are the differences between films? Again, without all the jokes, what's the process? First, the porogens are introduced in the block of the polymer. You can assemble an advance of the film. This film can assume different geometries. We need to understand of these geometries. There are different instruments to go after that. GXXX was one of the instruments used in another one was STM. That's what we used. The problem is, out of this film, out of this STM, we cannot exactly separate each of the pores, individualize them as we were doing with the fibers. Remember we had the fibers in red and that was great. We cannot do that in here. We need basically to understand the intensive variations across the different slices of the sample. So what we went after was texture analysis and then get to the porous film. A little more about the composition just in case you're curious how this is performed. They are also playing with the different conditions temperature here to cook up the perfect, the perfect espacem. We have here in a porous film with 58% porosity at the bottom with 73% porosity and just by looking at the pictures, several of the reviewers couldn't tell any of the difference. That was also one of the motivations of our team to keep kin. Can you give me some of the descriptor that will allow me to put them apart? That's when we started exploring a technique called gray-level coacuricin matrix. This is as old as I am, right? 1976, this is coming together and like, what's the big T of them? Why they didn't think about that? Well, there's a lot of conversation between computer vision folks and a material sciences that is to happen yet. And how many of you is familiar with a gray-level coacuricin matrix? Okay, so after this explanation, you are going to be an expert. So basically what you do, you get, you have your picture, right? It's just a matrix. You go to a position in that matrix, X, Y, that is an I, right? This is one position in your matrix and then another position in that matrix. What's the relationship? But suppose Dawn is I and we have our colleague here. Yeah, he's J, right? And so the distance between them is just one, right? And a particular direction. In this case, zero degrees. That's exactly what we're computing. So suppose you all are part of this matrix. What we're doing is computing at different directions, giving a particular distance. What is the difference between your intensity and the other? So if you're a pixel and your value is zero and your colleague at zero degrees and distance one is also zero, so there is no variation between the two of you. So probably we're talking about a homogenous patch. Can you see the picture? I also put a picture to help to understand that but what we're doing at the end of the day is to calculate a joint probability distribution. We're just calculating a 2D histogram, right? With the intensities. How large is going to be this matrix? As large as your pixel depth. We saw two to the power of eight yesterday, right? Two was pixels could vary between zero to 255. And the example that I'm going to give to you here. Our pixels are varying. We have a two power three, right? Pixel depth. It varies from one to eight. This is my picture, if you will, as it would come from a microscope or whatever is the instrument. And you check just the variations between them and give a distance one in this example. While there is one variation from one to one and so on. And you keep computing. So picture, you calculated the gray level co-curriculum matrix. And then on top of this 2D histogram, as we were used to do with the 1D histogram to calculate entropy, energy, right? We can also do that for 2D histograms. And that was exactly what we did. We computed different features from these gray level co-curriculum matrix and we found that pictures, features associated to how random that material was, the image of that material was showing up was very relevant. Some of the theory behind what they were looking for is a very low dielectric constant. So the lower, the better for them. That's what they were looking for. This is what the results that we got. Not only that, you need to come up with a constant. You need to find a material that obey that constant. What we got was, given the prior work, it wasn't achieved before, it was the first time that it was possible to come up with a very high porosity material that has a very low dielectric constant so you can make the perfect film for microelectronics. That was a collaboration with Intel. Apparently they were happy, so happy that they decided that this should be in a journal, the Advanced Functional Materials. I invite you to take a look at that paper if there is something interesting to you. But let me just step back here. I started talking about ceramic materials and now this is a complete material. I don't get it how I could use these features for my problem. Well, perhaps there is an intensive variation associated to separate tumor from the tumor and there are lots of work using heralic descriptors for that kind of job. That could be one potential application of these features across domains. Just blowing up what we found before, how we're separating these two different films using different descriptors. Okay, getting to health now. And lots of the things that I talked about before, particularly in finding where the fibers were, I use it for cytology. And what was interesting, sometimes doing one, you have freeing storms, how you could potentially solve the other. So science at Cresemains can be very fruitful in that regard. What kind of instruments our teams have been working on? From a light microscopy to MRI, with the samples that range from cervical cells and you've heard a few people here in the audience work on that. I've worked with the human memory epithelial cells and I actually show a few movies back there. We can get into the details if you're interested and bring skeins more recently. And these are because we're doing brain skeins and we may ask, well, how you do light microscopy in cytology? These people are our dad, right? Just in case you're asking yourself. There are different goals and the first one wouldn't be great if I have a cervical cell and I can't tell it's malignant or not. And because I'm gonna drill into the details of the third one, let me tell you, in here we're looking for correlations between different imaging modalities. So how many of you know what cervix is? I'm really impressed. I went to a conference in the last month. I mean, we had 200 people and I saw three girls in the corner like this, right? So cervix is at the bottom of the uterus and women need to go through cervical cancer exams if everything's going well every three years. It's not a fun exam. And what happens is fortunately here in the US the methods are very advanced but for the rest of the world they rely on conventional pap smear. That's the kind of image we've been looking on. Here I'm illustrating how does this collect. They basically scrub you and you get a bunch of cells put in a glass slide and they will look similar to this. There are lots of variations of these cells. And fortunately by participating in events like this I got the chance to meet a few immunopathologists that were looking into methods how we can do a very simple task. Find the cell. I want you to find the nucleus and the cytoplasm of these cells and individualize. Give me the counting. Very simple question, right? Accepted for your computer vision system to do that with a natural routine exam cells. It wasn't that simple. Well, our teams started working on this problem and after a while we decide we need a very large data set. We found a data set in the web that was part of the code challenge. Wow, okay, let's get into a code competition and see how well our current algorithms are doing. And what happened by participating of the International Symposium on Biomedical Imaging in 2014 we exposed our codes and we learned, guess what, we're doing pretty well on this. That's when we decided to apply for funding in continuous processing research in this area. What I'm showing here are some of the preliminary results that we calculated, completely automated. Nobody went there in manually segment except for the, to generate the simulated data sets. And I can give you more details about that. We can even make that one of the breakouts if you think this is interesting. But the goal is very large data set. We can simulate how these cells can show up and then we can, the goal is to separate each one of them. Here's some pictures of the teams involving large teams from pathologists, biologists, computer scientists and among the several algorithms that we were doing for the segmentation and classification we've also been working on a searchable database. Remember, I promise you that we're trying to work on the picas of our faces of experiments. We're trying to get very close to that. We do have a prototype, I'll be happy to give you a demo if you want in which we have real cells being all segmented through this remote system in collaboration with these folks in Brazil. And also let me point out, yeah, Rome and Flavio who are sitting somewhere here in the audience also working in this project. Enough of the cervical cells, now we're going to a much larger, we're talking about tissues, but also here it's going to involve how we're going to better find enhanced features. So all the goodies that we've been talking before, theaters, how we find descriptors so that we can classify different problems. For those that have been working with Neuroscience for a while, you've heard of, oh, we want to register MRI with PAT, positive emission tomography. This is very standard, people have been trying to do that for a long time. One of the motivations is MRI has a course resolution in comparison with other techniques, for example, PAT. Well, we knew that in addition to PAT, one of the ideas that Professor Lea Grimberg from UCSF had, what about if we had the histology? Well, but you need to have then a bank of dead people to chop around just like a salami and you can register that to the real MRI. So she did. Ask Graze and this may sound, you know, a bunch of scientists got to get it, we got to do this just because it's freaky. No, because you can sting and also know about the proteins associated to different portions of the brain. So this is the whole of Grail, folks. With all is asking, what's the function associated to the protein associated to the part of the brain? So if this is possible, we're gonna understand much more of what's going on in the brain. This is the work I want to point out to one of the folks here in the audience. Mariana Legro is leading a lot of this work and put together a pipeline to do the registration between the histology and MRI of these images. And this was just accepted as a paper at CVPR and just in a nutshell, what's done, why it's challenging. You may ask yourself, well, I imagine you're chopping as sharp as it is the knife that they have and I got the chance to see it. There's deformations associated with cuts. So you need to take those into account. You need to fix those deformations. That comes, Mariana work in collaboration with the team of LBL in several processes. So now you can put the two of them together, the MRI and the histology. The histology. We're looking at the tissue level here. These are some of the preliminary results as well in which I'm showing MRI in the back and in purple here, the different segments of the brain on the cytology. So challenging here, how we can fuse different modalities. You know, about you to talk to Mariana about more methods and the area. Another reason it's challenging. You cannot serve this in your laptop. These data sets are incredibly large and she has counted with the supercomputers at LBL, particularly using computers at NERSC to do a lot of these registrations. Okay, and I think I'm just two of them in my time. So, okay, I'll jump into what do we think about feature? And that's a very personal take into what's about the feature. I think we're going from the one off to definitely analytical motifs, if you will. Things that serve for several different domains. And the way it grew into me was just by exposing myself. I've been working in this area for almost 20 years, 19 years working with the computer vision and get to a point that there's so much that it can do. So curious, I wanna work with all these different science domains, so I need to find these basic motifs that will allow me to spread across different areas. One of the fun things that I've learned and I had the chance to play around particularly two weeks ago in IBM was with convolution neural networks. By no means IBM invented that, right? This was invented a long time ago. Passed for several processes. One of the key people working in this area is Yann LeCun, who is also involved in one of these data science centers. We also have a few folks here working for a long time with the CNN. And I can point out Stella Yu here, who is one of the experts in the area. I'm sure Jitanda Malik also has a lot to talk about all these different areas that I mentioned in terms of algorithms. But if you've never used or you would like to learn something, in a nutshell what it is is you have a picture and you pass through several layers that extract characteristics out of this picture. Very similar to the operations that we saw yesterday during the Python hands-on. Do you remember you were working with convolutions? Yeah, this convolution here is the same convolution of there, except that this is done in a way that you don't need to be fine-tuning yourself each of those parameters. You let the data speak to itself, which means that by feeding into this supervised classification system, you're able to recover very different classes associated to your problems. Through munching to several layers, so deep, you also may have heard the term deep learning because you have several layers in this neural network for processing so that you can classify your different themes or different problems. That's pretty much the path of the processing that I'm illustrating in green in this image. What I did. One collaboration with Talita that is also here, well, we put together a very large database. That's one of the things that you needed to play with our written such as CNN, in which we have approximately 300,000 samples of fibers and no fibers. And the goal is, can you pull apart this picture from this picture? And what I noticed is after I thought I was too tired, I was looking at thousands of this pictures. And after a while, I could not tell for myself that those were fibers, particularly because of this is all gray level. And humans tend to have difficulties after a while at looking at gray levels. So okay, let's see how well this convolution neural network running in a wet. Why was I at IBM again? Well, they are testing a new chip. So you're gonna have a device is specifically doing a kind of, I'm gonna call computer vision here as well. A particular kind of machine learning computer vision algorithm with CNN. So pretty soon we're gonna be able to have a microscope. We're gonna plug in here my trained CNN and then have the results, right? At my computer, what's this that I have? Am I grabbing the fiber or not the cell or not? So I see that coming pretty soon. Through neuromorphic computing, it may be through IBM or not. There are several other neuromorphic chips I had the chance to run in this particular one called IBM True North. How good is it? We were impressed to learn that we got a 99.78% accuracy with those fibers and no fibers. So we're really impressed with what we got. I think this is too very preliminary, but we're able to do that in a single chip. So you can put several chips together. We did in one chip. One chip means that you have a 4,065 cores. We did in about 3,200 cores. Miniature. This another fun thing about this chip, it works with a little more than a battery that you use for your watch. So you can put that in a drone if you want. Do some processing in real time. That may be a way to reduce the data as we are collecting so much data these days. So perhaps we will need to embed layers in between the instrument that is collecting the data and what we're keeping in our records. While there are many other things that we're doing up in the lab, particularly with this chip, I showed you the problem with the fibers, but there are different instruments that can be also tailored into this pipeline and you're going to solve with a CNN. CNN is gonna save us all. I don't think so. There are particular problems that you can tailor in such a way that will be perfect at the end of the day. Yeah, you're gonna get 99%. But there are others that I believe that you will not. There's so much pre-processing or so much data that we will require that may not be the way to go. This is part of an LDRD, a project sponsored by LBL Director. A few other folks involved in that. One of them is Van Cat, who is also sitting here among us. You can ask him more about how to do a simulation of a crystal lattice. And perhaps they can share with you this very large data sets with interesting applications. One of them is in the characterization of the material properties. And with that, I wanna end with, yeah, the science and scientists. For me, it has been great to participate in bids. I've been learning a lot throughout my career, both on my own as well as in teams, but definitely very powerful with community. There's too much to learn. You've got to leverage from your colleagues and this is one of the people that I've been relying upon a lot. This is the data analytics and visualization group. I've been with them for almost 10 years now and learn a lot with them and by work with them, I really have been able to do much more if you're interested to spend some time with us. As a visiting professor or a summer student, you name it, come talk to me. Camera, we've been talking, what's this camera? A camera stands for the Center of Applied Mathematics for Industry Research Application. These was a dream that started long ago when people would say, man, this is not gonna work. You want different areas within the Department of Energy to talk to each other. You want material science talking to computer science at the level of funding. This is not gonna happen. Well, guess what? Camera is the realization that, yeah, it can happen. You need to come up with a very good idea in a wonderful team that is willing to work hard and talk to each other. So this is part of the team. There are more of us, and you may actually find a few others in here in the audience. Another team, and this started just like this. I wanna show this picture because about a year ago, we put together hands-on, a workshop, and a hackathon. Pretty much like we're doing this week. And after a while, what happened? We have a few hundred, a thousand dollars later, we are sponsored, right? We have papers together. I didn't know most of these folks before. By coming together, we were able to put packages that we didn't think of before, right? And I need to point out that two of these folks of the 10 pounds slimmer, I would say, are here today, thanks to this hackathon that started just like this that we're currently going through. And I would like to end with a big thank you to the Beats family here in the pictures of probably most of us. Definitely it wouldn't be possible to pursue all these different science domains without a very large community. Thank you.