Audio player should automatically calculate it's heading levels

Root Level h3

Transcript should be h5, cues should be h6

Code Comments

Bringing Deep Learning to Enterprise Applications

About the episode

There are a lot of publicly available data sets out there. But when it comes to specific enterprise use cases, you're not necessarily going to able to find one to train your models. To realize the power of AI/ML in enterprise environments, end users need an inference engine to run on their hardware. Ryan Loney takes us through OpenVINO and Anomalib, open toolkits from Intel that do precisely that. He looks specifically at anomaly detection in use cases as varied as medical imaging and manufacturing.

Want to learn more about Anomalib? Check out the research paper that introduces the deep learning library.

Ryan Loney Product manager, OpenVINO Developer Tools, Intel®
Subscribe

Subscribe here:

Listen on Apple Podcasts Listen on Spotify Listen on Google Podcasts Subscribe via RSS Feed
Hi, I'm Burr Sutter. I'm a Red Hatter who spends a lot of time talking to technologists about technologies. We say this a lot of Red Hat. No single technology provider holds the key to success, including us. And I would say the same thing about myself. I love to share ideas. So I thought it'd be awesome to talk to some brilliant technologists at Red Hat Partners. This is Code Comments, an original podcast from Red Hat. I'm sure, like many of you here, you have been thinking about AI, ML, artificial intelligence and machine learning. I've been thinking about that for quite some time and actually had the opportunity to work on a few successful projects here at Red Hat using those technologies, actually enabling a dataset, gathering a dataset, working with data scientists and data engineering team, and then training a model and putting that model into production runtime environment. It was an exciting set of projects and you can kind of see those on numerous YouTube videos I have published out there before. But I want you to think about the problem space a little bit because there are some interesting challenges about AI/ML. One is simply just getting access to the data, and while there are numerous publicly available datasets when it comes to your specific enterprise use case, you might not be to find publicly available data. In many cases, you cannot, even for our applications that we created, we had to create our dataset, capture our dataset, explore the dataset, and of course train a model accordingly. And we also found there's another challenge to be overcome in this AML world, and that is access to certain types of hardware. If you think about the enterprise environment and the creation of an enterprise application specifically for AML and users need an inference engine to run on their hardware, hardware that's available to them to be effective for their application. Let's say an application like computer vision, one that can detect anomalies in medical imaging or maybe on a factory floor, You know, those things are whizzing by on the factory line. They're looking at them and trying to determine if there is an error or not. Well, how do you actually make it run on your hardware, your accessible technology that you have today? Well, there's a solution for this as an open toolkit called Open vino. And you might be thinking, hey, wait a minute, don't you need a GPU for a I inferencing a GPU for artificial intelligence machine learning? Well, not according to Ryan Loney, product manager of Open Vino Developer Tools at Intel. I guess we'll start with trying to maybe dispel the myths, right? I think that CPUs are widely used for inference today. So and if we look at the data center segment, you know, about 70% of the A.I. inference is happening on Intel Xeon on our data center CPUs. And so you don't needed a GPU, especially for running inference. And that's part of the value of open vino, is that we're you know, we're taking models that may have been trained on a GPU using deep learning frameworks like PyTorch or TensorFlow and then optimizing them to run on Intel hardware. Ryan joined me to discuss AI/ML and the enterprise across various industries and exploring numerous use cases. Let's talk a little bit about the origin story behind Open Vino. Tell us more about it and how it came to be and why it came out of Intel. Definitely. So we had the first release of Open Vino was back in 2018, so still relatively new. And at that time we were focused on computer vision and pretty tightly coupled with open CV, which is another open source library with origins at Intel. You know, it had its first release back in 1999, so it's been around a little bit longer. And many of the software engineers and architects at Intel that were involved with and contributing to open CV are working on open Vino. So you can think of open vino as complementary software to open CV. And we're providing like an engine for executing inference as part of a computer vision pipeline, or at least that's how we started. But since 2018, we've we've started to move beyond just computer vision inference. So when I say computer vision inference, I mean like image classification, object detection, segmentation. And now we're moving into natural language processing, things like speech synthesis, speech recognition, knowledge, graphs, time series forecasting, and other use cases that don't involve computer vision and don't involve inference on pixels. Our latest release, the 20 22.1 that came out earlier this year, there was a most significant update that we've had to open vino since we started in 2018, and the major focus of that release was optimizing for use cases that go beyond computer vision. And I like that concept that you just mentioned right there, computer vision. And you said that you extended those use cases and went beyond that. So could you give us more concrete examples of computer vision? Yeah, sure. So when you think about manufacturing quality control in factories, everything from ARC welding, defect detection to inspecting BMW cars on assembly lines, they're using cameras or sensors to collect data. And usually it's cameras collecting images like RGV images that you and I can see. And looks like something taken from a camera or video camera, but also things like infrared or computerized tomography scans used in health care, X-ray, different types of images where we can draw bounding boxes around regions of interest and say, you know, this is a defect or this is not a defect. And also, is this worker wearing a safety hat or did they forget to put it on? And so you can take this and integrate it into a pipeline where you're triggering an alert if somebody forgets to wear their safety mask or if there's a defect in a product on an assembly line, you can just use cameras and open vino and open CV running these on Intel hardware and help to analyze. And that's what a lot of the partners that we work with are doing. So these independent software vendors and there's other use cases for things like retail. You think about going to a store and using an automated checkout system. You know, sometimes people use those automated checkouts and they they slide a few extra items into their bag that they don't scan. And it's a huge loss for the retail outlets that are providing this way to check out real time shelf monitoring. We have this bear on one of our is fees that helps keep store shelves stocked by just analyzing the cameras in the stores, detecting when objects are missing from the shelves so that they can be restocked. We have Vistry, another ISP that works with quick service restaurants. So when you think about automating the process of when do I drop the fries into the fryer so that they're warm when the car gets to the drive thru window, you know, there's quite a bit of industrial health care retail examples that we can walk through and we should dig into some more of those. But I got to tell you, I have I have a personal experience in this category that I want to share with, and you can tell me how how silly you might think at this point in time. It is. We actually built an AI keynote demonstration for the Red Hat big stage back in 2015, and I really want to illustrate the concept of asset tracking. So we actually gave everybody in the conference a little Bluetooth token, but a little battery, a little watch battery and a little Bluetooth emitter. And we basically tracked those things around the conference. We basically put a Raspberry Pi in each of the meeting rooms and up in the lunch room, and you could see how the tokens moved from room to room to room as a relatively simple application. But it occurred to me after we figured out, okay, how to do that with Bluetooth and triangulating Bluetooth signals by looking at relative signal strength from one radio to another and putting that through an Apache Spark application at the time, we then realized, you know what, this is easier done with cameras and just simply looking at a camera and having some form of animal or machine learning model that would say, Oh, there are people here now are there are no people here now. What do you think about that? Yeah, I mean, what you just described is sort of exactly that the product that either one of our partners is offering, you know, that they're doing it with computer vision and cameras. So when partner tries to help retail stores analyze the foot traffic and understand with Heatmaps, where people are spending the most time in stores, how many people are coming in, what size groups are coming into the store, you know, and trying to help understand if there was a successful transaction from the people who entered the store and left the store so that you can, you know, to help with the, you know, retail analytics and marketing sales and positioning of products. And so they're doing that in a way that also protects privacy. And that's something that's really important. So when you talked about those Bluetooth beacons, probably, you know, if everyone who walked into a grocery store was asked to put a tracking device in their cart or on their person and say, you know, you're going to be tracked around the store, they probably wouldn't want to do that. The way that you can do this with cameras is you can, you know, detect people as they enter and, you know, remove their face. Right. So you can ignore any biometric information and and just track the person based on pixels that are present in the detected region of interest. So they're able to analyze, say, a family walks in the door and they can group those people together with object detection and then they can track their movement throughout the store without keeping track of their face or any biometric or any personal identifiable information to avoid things like bias and to make sure that they're protecting the privacy of the shoppers in the store while still getting that really useful marketing analytics data rate so that they can make better decisions about where to place their products. So that's one really good example of how computer vision AI with open vino is being used today. And that is a great example because you're definitely spot on. It is invasive when you hand someone to Bluetooth devices, say, please keep this with you as you go throughout our our store or our mall or throughout our hospital, wherever you might be. Now, you mentioned another example earlier in the conversation which was related to like worker safety. Are they wearing a helmet? I want to talk more about that concept in a real industrial setting, a manufacturing setting where there might be a factory floor and there are certain requirements, or better yet, there's like a a quality assurance requirement, let's say, when it comes to looking at a factory line, I run to that use case often what some of our customers. Can you talk more about those kinds of use cases? Yeah. So one of our partners, Robuchon and we you know, published a case study I think last year where they're working with BMW at one of their factories and they do quality control inspection, but they're also doing things related to worker safety and analyzing. You know, I used the safety had example. There's a number of of our ISP's and partners who have similar use cases. And it comes down to there's a few reasons that are motivating this and some are related to like insurance, right? It's important to make sure that if you want to have your factory insured and that your workers are protecting themselves and wearing the gear. Regulatory compliance. Right. You're you're being asked to properly protect from exposure to chemicals or, you know, potentially having something fall and and hit someone on the head. So wearing a safety vest, wearing goggles, wearing a helmet, these are things that you need to do inside the factory. And you can really easily automate and detect and sometimes without bias. I think that's one of the interesting things about the robots on BMW example is that they were also blurring sort of blocking out and so drawing a box to cover the face of the workers in the factory so that somebody who was analyzing the video footage and getting the alerts saying that, hey, you know, Bay 21 has a worker without a hat on, that it's not sending their face and in the alert and potentially, you know, invading or going against privacy laws or just the ethics of the company. Right. They don't want to introduce bias or have people targeted because it's much better to to have it be, you know, blur the face and alert and have somebody take care of it on the floor. And then if you ever need to audit that information later, they have a way to do it where people who need to be able to see who the employee was and look up their personal information, they can do that. But then just for the purposes of maintaining safety, they don't need to have access to that personal information or biometric information, because that's one thing that when you hear about computer vision or object person tracking, object detection, there's a lot of concern, and rightfully so, about privacy being invaded and about tracking information, face ID, identifying people who may have committed crimes through video footage. And that's just not something that a lot of companies want to you know, they want to protect privacy and they don't they don't want to be in a situation where they might be violating someone's rights. Well, privacy is certainly opening up Pandora's box. There's a lot to be explored in that area, especially in a digital world that we now live in. But for now, let's move on and explore different area. I'm interested in how machines and computers offer advantages, specifically in certain use cases like a quality control scenario. I asked Ryan to explain how AML and specifically machines computers can augment that capability. I can give a specific example where we have a partner that's there doing defect detection with and looking for anomalies in batteries. So, you know, sure, you've heard there's a lot of interest right now in electric vehicles, a lot of batteries being produced. And so if you go into one of these factories, they have images that they collect of every battery that's going through this assembly line and through these images, people can look and see and visually inspect with their eyes and say, this battery has a defect, send it back. And that's one step in the quality control process. And there's other steps, I'm sure, like running diagnostic tests and, you know, measuring voltage and doing other types of non-visual inspection. But for the visual inspection piece where you can really easily identify some problems, it's much more efficient to introduce computer vision. And so that's where we have this new library that we've introduced called Anomali, that's open vino. While we're focused on inference, you know, we're also thinking about the pipeline or the funnel that gets these models to open vino. And so we've we've invested in this anomaly segmentation, anomaly detection library and that we've recently open source and there's a great research paper about it about Anomali. But the idea is you can take just a few images and train a model and start detecting these defects. And so for this battery example, that's a more advanced example. But to make it simpler, you know, take some bolts and, you know, take ten bolts. You have one that has a scratch on it or one that is chipped or has some damage to it. And you can easily get started in training to recognize the bolts that do not have an anomaly. And the ones that do, which is a small data set and I think that's really one of the most important things today. Challenges is one is access to data, but the other is is needing a massive amount of data to do something meaningful. And so we're starting to try to change that dynamic with Anomali. So you may not need 100,000 images, you may need 100 images, and you can start detecting anomalies in everything from batteries to bolts to, you know, maybe even the wood varnish use case that you mentioned. That is very key point because often in that data scientist process, that data engineer and data scientist process, right. The one key thing is can you gather the data that you need for the input for the model training? And we've often sat at least people I've worked with over the last couple of years, you know, you need a lot of data. You need tens of thousands of correct images so we can sort out the difference between dogs versus cats, let's say, or you need dozens and dozens of situations where if it's a natural language processing scenario, you know, a good customer interaction, a good customer conversation, and in this case, it sounds like what you're saying is show us just the bad things, right? Fewer images, fewer incorrect things, and then let us look for those kind of anomalies. Can tell us more about that because that is very interesting. The concept that I can use a much smaller dataset as my input as opposed to gathering terabytes of data in some cases to just simply get my model training underway. You know, like you described, the idea is if you have some good images and then you have some of the the known defects and you can just label here's a set of good images and here's a few of the defects and you can right away start detecting those specific defects that you've identified. And then also, you know, be able to determine when it doesn't match the expected appearance of a non defective item. So if I have the undamaged screw and then I introduce one with some new anomaly that's never been seen before, I can say, you know, this one is not a valid screw. And so that's sort of the the approach that we're taking. And it's really important because so often you need to have subject matter experts often like if you think the take the battery example, there's these workers who are on the floor in a factory and they're the ones who know best when they look at these images, which one's going to have an issue, which one's defective? And then they also need to take that subject matter, expertise and then use it to annotate data sets. And when you have these, you know, tens of thousands of images you need to annotate, it's asking those people to stop working on the factory floor so they can come annotate some images. That's a tough business call to make, right? But if you only need them to annotate a handful of images, it's a much easier ask to get the ball rolling and demonstrate value. And maybe over time you will want to annotate more and more images because you'll get even better accuracy in the model. Even better, even if it's just small incremental improvements. You know, that's something that if it generates value for the business, it's something the business will invest in over time. But you have to convince the decision makers that it's worth the time of these subject matter experts to stop what they're doing and go and label some images of the things that they're working on in the factory. And that labeling process can be very labor intensive of the annotations, basically saying what is correct, what's wrong, what is this, what is that? And therefore, if we can minimize that time frame to get the value quicker, then there's something that's useful for the business, useful for the organization long before we necessarily good. There are huge model training based, so we talk about labeling and how that is labor intensive activity. But I love the idea of helping the human and helping the human models specifically not get bored. Basically, if the human is eyeballing a bunch of widgets flying by over time, they make mistakes, they get bored and they don't pay as close attention as they should. That's why the concept of Amazon specifically computer vision, augmenting that capability and really helping the human identify anomalies faster, more quickly, maybe with greater accuracy could be a big win. We focused on manufacturing, but let's actually go into health care and learn how these tools can be used in that sector and that industry. Ryan talked to me about how Open Windows runtime can be incorporated into medical imaging equipment with intel processors and better than c.T. MRI and ultrasound machines. Well, these inferences, this AML workload can be operating and executing right there in the same physical room as the patient. We did a presentation when she last year. I think they said there's at least 80 countries that have their X-ray machines deployed and they're doing things like helping doctors place breathing tubes in patients. So during COVID, during the pandemic, that was a really important tool to help with nurses and doctors who were intubating patients sometimes like in a parking lot or a hallway of the hospital. And, you know, when they had a statistic that you said, I think one out of four breathing tubes gets placed incorrectly when you're doing it outside the operating room, because when you're in an operating room, it's much more controlled and there's someone who's an expert at placing the tubes. It's something you have more of a controlled environment than when you're out in a parking lot, in a tent. You know, when the hospital's completely full and you're triaging patients with COVID, that's when they're more likely to make mistakes. And so they had this endotracheal tube placement model that they trained, and it helped to use an x ray and give an alert and say, hey, this tube is placed wrong, pull it out and do it again. And so things like that help doctors so that they can avoid mistakes. And, you know, having a breathing tube placed incorrectly can cause collapsed lung and a number of other unwanted side effects. So it's really important to do it correctly. Another example is Samsung Medicine. They actually are doing estimating fetal angle of progression. So this is analyzing ultrasound of pregnant women with that, being able to to help take measurements that are usually hard to calculate that can be done in an automated way. They're already taking the ultrasound scan and now they're executing this model. They can take some of these measurements to help the doctor avoid potentially more intrusive alternative methods so the patient wins. It makes their life better. And the doctors is getting help from this A.I. model. And those are, you know, just a few examples. Those are some amazing examples when it comes to all these things. We're talking like CT scans, right, and x rays, other examples of computer vision. One thing that's kind of interesting in the space, I think whenever I get a chance to work on, let's say an object traction model and one of our workshops, by the way, is actually putting that out in front of people to say, Hey, look, you can use your phone. And it basically sends the image over to our OpenShift, right, with our data science platform and then analyzes what you see. And even in my case, where I take a picture of my dog as an example, it can't really decide is it a dog or a cat? I have a very funny looking dog, and so there's always a percentage outcome, you know? In other words, I think it's a dog 52%. So I want to talk about that more. What how important is it to get to 100% accuracy? How important is it to really, depending on the use case, to allow for the gray area, if you will, where it's an 80% accuracy or 70% accuracy? And where are the trade offs there associated with the application? Can you can you discuss that more? Accuracy is definitely, you know, a touchy subject because how you measure it makes a huge difference. And then I think with like what you were describing with the dog example, there's sort of a top five potential classes that might may be identified. So let's say you're doing object detection and you detect a region of interest and it says 65% confidence. This is a dog. Well, the next potential label that could be maybe 50% confidence or 20% confidence might be something similar to a dog or in the case of models that have been trained on like the image net dataset or on cocoa data set, they have like actual breeds of dogs. So if I want to look at the top five labels for a dog, for my dog, for example, she's a mixed mostly Labrador retriever. But I may look at the top five labels and it may say 65% confidence that she's a flat coated retriever and then confidence that she's a husky, as you know, 20% and then 5% confidence that she's a Greyhound or something. Those labels, all of them are dogs. So if I'm just trying to figure out is, is this a dog, I could probably find all of the, you know, classes within the data set and say, well, these are all, you know, class ID 65, 132, 92 and 158 all belong to a group of dogs. So if I wanted to just write an application to tell me if this is a dog or not, I would probably use that to determine if it's a dog. But how you measure that is accuracy. Well, that's where it gets a little bit complicated, because if you're being really strict about the definition and you're trying to validate against the data set of labeled images and I have specific dog breeds or some specific detail and it doesn't match, well, then the accuracy is going to go down. That's especially important when we talk about things like compression and quantization, which, you know, historically has been difficult for to get adoption in some domains like health care, where even the hint of accuracy going down implies that we're not going to be able to help in some small case, maybe if it's even half a percent of the time we want to take that that tube is placed incorrectly or that, you know, that patient's, you know, lung has collapsed or something like that. And that's something that really prevents adoption of some of these methods that can really boost performance like quantization. But if you take that example of sort of different from the dog example and you think about like segmentation of kidneys. So if I'm doing kidney segmentation, which is, you know, taking a CT scan and then trying to pick the pixels out of that scan that belong to a kidney, how I measure accuracy may be how many of those pixels I'm able to detect and how many did I miss? Missing some of the pixels is maybe not a problem, right, depending on how you built the application because you still detect the kidney and maybe you just need to apply padding around the region of interest so that you don't miss any of the the actual kidney. When you compress the model and when you quantized the model. But that requires, you know, data scientist and email engineer somebody to really they have to be able to go and apply that after the fact, after the inference happens to make sure that you're not losing critical information, because the next step from detecting the kidney may be detecting a tumor. And so maybe you can use the more optimized model to detect the kidney, but then you can use a slower model to detect the tumor. But that also requires somebody to architect and make that decision or that tradeoff and say, well, I need to add padding, or I should only use the quantized model to detect the region of interest for the kidney and then use the model that takes longer to do the inference just to find the tumor, which is going to be on a smaller size. Right. The dimensions are going to be much smaller once we crop to the region of interest. But all of those details, that's maybe not easy to explain in a few sentences. And even the way I explained it is probably really confusing. I do love that use case. Like you mentioned, the cropping even in one such an area that we worked on for another project, we specifically decided to pix like the image that we had taken because we knew that we could get the outcome we wanted by even just using a smaller or less having less resolution in our image. And therefore, as we transferred it from the mobile device storage device up into the cloud, we wanted that smaller image just for transfer purposes and it still we could get the accuracy we needed by a lot of testing. And one thing that's interesting about that from my perspective is if you're doing image processing, sometimes it takes a while for this transaction to occur. Like I, I come from a traditional application background, you know, where I'm reading and writing things from a database or a message broker or moving data from one place to another. Those things happen subsequent. Normally, even with great latency between your data centers, you know, it's still subsequent. In most cases, while on a transaction like this, one can actually take 2 seconds or 4 seconds as it's doing its analysis and actually coming back with, you know, I think it's a dog, I think it's a kidney, I think it's whatever, and provided me that accuracy statement. So that concept of optimization is very important in the overall application architecture. Would you agree with that or how do you think about that concept? Yeah, definitely. It depends too on the use case. So if you think about how important it is to reduce the latency and increase the number of frames per second that you can process when you're talking about a loss prevention model that's running at a grocery store. So you want to keep the lines moving. You don't want every person who's at the self-checkout to have to wait 5 seconds for every item they scan. You need it to happen as quickly as possible. And if sometimes you, you know, the accuracy decreases slightly or the I'd say the accuracy of the whole pipeline. So not just looking at the individual model or the individual inference, but let's say that the the whole pipeline is not as successful at detecting when somebody steals one item from the self-checkout, it's not going to be a life threatening situation. Whereas, you know, being in the hooked up to the X-ray machine with the two placement model, they might be willing to have the doctor, the nurse wait 5 seconds to get the result. They don't need it to happen in 500 milliseconds. So they're willing their threshold for waiting is a little bit higher. So that, I think, also drives some of the decision, like you want to keep people moving through the checkout line and you can afford to to potentially if you lose a little bit of accuracy here and there, it's not going to cost the company that much money or it's not going to be life threatening. It's going to be worth the tradeoff of keeping the line moving and not having people leave the store and not check out at all. And to say, I'm not going to shop today because the line's too long. There are so many trade offs and enterprise AML use cases, things like latency, accuracy and availability. And certainly complexities abound, especially in an obviously ever evolving technological landscape where we are still very early in the adoption of AML. And to navigate that complexity, the direct feedback from real world end users is essential to Ryan and his team at Intel. What would you say are some of the big hurdles or big outcomes, big opportunities in that space? And do you agree that we're kind of still at the very beginning in our infancy, if you will, of adopting these technologies and and discovering what they can do for us? Yeah, I think we're definitely in the infancy and I think that what we've seen is our customers are evolving and the people who are deploying on Intel hardware, they're trying to run more complicated models. They're the models that are doing object detection or, you know, detecting defects and, you know, doing segmentation. You know, in the past you could say, oh, here's a generic model that will do face detection or person detection or vehicle detection and license plate detection. And those are sort of like general purpose models that you can just grab off the shelf and use them. But now we're moving into like the anomaly scenarios where I've got my own data and I'm trying to do something very specific and I'm the only one that has access to this data. And you don't have a public data set that you can go download that's under Creative Commons license for, you know, car batteries. It's, you know, it's just not something that's available. And so those use cases, the challenge with with training those models and and getting them optimized is the beginning of the pipeline. It's the data you have to get the data you have to annotated and the tools have to exist for you to do that. And that's part of the problem that we're trying to help solve. And then the models are getting more complex. So if you think, you know, just from working with customers recently, you know, they're no longer just trying to do image classification and, you know, like is it a dog or a cat? They've moved on to like 3D point clouds and, you know, 3D segmentation models and things that are like the speech synthesis example, doing things these GPT models that are generating, you know, you, you put a text input and it generates an image for you. It's just becoming much more advanced, much more sophisticated and on larger images. And so things like running super resolution enhancing images, upscaling images, instead of just trying to take that, you know, 200 by 200 pixel image and classifying if it's a cat. Now we're talking about gigantic huge images that we're processing and that all requires more resources or more optimized models. And, you know, every computer vision conference or A.I. conference, there's there's a new latest and greatest architecture. There's new research paper, and things are getting adopted much faster. The lead time for a nurse paper or CV PR for a company to actually adopt and put those into production. It's like the time shortens every year. Well, Ryan, I got to tell you, I could talk to you literally all day about these topics, the various use cases, the various ways models are being optimized, how to put models into a pipeline for average enterprise applications. I've enjoyed learning about pop and vino and anomalies, but I'm fascinated by this because I will have a chance to go try this myself. Taking advantage of Red Hat OpenShift and taking advantage of our data science platform. On top of that, I will definitely go be poking at this myself. So thank you so much for your time today. Thanks, Burr. This was a lot of fun. Thanks for having me. And you can check out the full transcript of our conversation and more resources, like a link to a white paper on open vino and normal lib at Red Hat dot com slash code Comments Podcast. This episode was produced by Brant Seminole and Caroline Prickett. Our sound designer is Christian. From our audio team includes Lee Day, Stephanie Wunderlich, Mike Esser, Laura Barnes, Claire Allison, Nick Burns, Aaron Williamson, Karen King, Booboo House, Rachel Artell, Mike Compton, Ocean Mathews, Laura Walters, Alex Trabelsi and Victoria Lutton. I'm your host, Burt Sutter. Thank you for joining me today on Code Comments. I hope you enjoyed today's session and today's conversation, and I look forward to many more in.

Even with more headings

This last heading is allowed to go back up to h2, but the player should still take h3 as its root heading level.