AWS re:Invent 2021 - Accelerate innovation with machine learning

Aug 16, 2023

AWS re:Invent 2021 - Accelerate innovation with machine learning

With the rise in compute power and data proliferation, machine learning has moved from the peripheral to being a core part of businesses and organizations across industries. AWS customers use machine learning and AI services to make accurate predictions, get deeper insights from their data, reduce operational overhead, improve customer experiences, and create entirely new lines of business. In this session, explore how AWS services can help you move from idea to production with machine learning.

Learn more about AWS at https://bit.ly/3IQa1L0

Subscribe:
More AWS videos http://bit.ly/2O3zS75
More AWS events videos http://bit.ly/316g9t4

ABOUT AWS
Amazon Web Services (AWS) hosts events, both online and in-person, bringing the cloud computing community together to connect, collaborate, and learn from AWS experts.

AWS is the world’s most comprehensive and broadly adopted cloud platform, offering over 200 fully featured services from data centers globally. Millions of customers—including the fastest-growing startups, largest enterprises, and leading government agencies—are using AWS to lower costs, become more agile, and innovate faster.

#AWS #AmazonWebServices #CloudComputing

Content

0.35 -> [music playing]

1.59 -> Please welcome Vice President of Artificial Intelligence and Machine

4.78 -> Learning Services AWS, Bratin Saha.

8.65 -> [music playing]

13.295 -> [applause]

18.49 -> Good afternoon, everyone, welcome,

21.11 -> and thank you for joining me for the AI/ML Leadership Session.

24.75 -> I’m Bratin Saha, VP of AI/ML at AWS.

28.89 -> When I earned my Ph.D. in Computer Science,

31.62 -> machine learning was just starting its evolution

34.88 -> from an academic pursuit to what it is today,

38.54 -> a critical component of every company’s business strategy.

42.97 -> Now at Amazon, we have been at the forefront of this machine

46.48 -> learning evolution through Alexa,

49.19 -> Amazon Go, Amazon Prime, Amazon.com, and others.

53.97 -> In fact, every time someone buys something from Amazon.com.

59.07 -> it goes through one of our machine learning services.

62.25 -> I am incredibly proud of what our machine

64.62 -> learning teams have done at a scale and complexity

68.1 -> that is truly unprecedented.

72.15 -> Now at AWS, we have channeled this deep expertise of deploying machine

77.23 -> learning at scale to create the AI and Machine Learning Services

82.01 -> for our customers,

83.64 -> and today more machine learning happens at AWS

87.78 -> than anywhere else, and I have had the privilege

92.28 -> of helping our customers use our AI and Machine Learning Services

97.61 -> to extract uncommon insight from their data,

101.18 -> and then use those insights to drive better business outcomes.

106.39 -> In this journey, I also had the opportunity to help build

110.6 -> one of the fastest growing services in AWS history,

115.17 -> and learning many valuable lessons along the way.

118.78 -> One of those lessons is that machine learning

122.91 -> is not the future that we need to plan for,

126.48 -> machine learning is the present that needs to be harnessed now,

132.96 -> and so in this talk I would like to talk about why machine learning

138.91 -> is critical to innovation today,

141.91 -> how we think it’s going to evolve in the future,

145.5 -> and what that means for every company in every industry.

151.69 -> Let’s start with some data on the progression of machine learning.

155.82 -> According to IDC, the global spend on enterprise AI/ML

162.11 -> has gone from virtually zero in 2013 to 50 billion dollars in 2020,

169.06 -> going from zero to 50 billion dollars in just 7 years.

175.27 -> By comparison, cloud computing went from zero to 50 billion

180.73 -> in almost 12 years, almost twice as long as machine learning.

187.37 -> Now, this investment in machine learning is happening

190.78 -> because it will help us solve important economic and social issues,

195.75 -> and this investment is apparent in the way

200.37 -> our customers are adopting machine learning.

203.84 -> Today, more than 100,000 customers

208.305 -> across virtually every industry –

211.85 -> Capital One and Fidelity in financial services,

215.27 -> Philips and Novartis in healthcare, Amazon.com

219.21 -> and Mercado Libre in retail, Formula 1 and NFL in sports,

224.77 -> Bayer and Siemens in industrial – and many, many other companies

230.9 -> are using the AI and Machine Learning Services on AWS

236.18 -> and getting significant business results.

240.8 -> So as machine learning takes off, many people ask us,

247.12 -> how will this all play out?

250.46 -> So I want to use the rest of this talk

253.39 -> on the four key drivers of machine learning innovation,

258.42 -> why we think these are the key drivers for machine

261.62 -> learning innovation, and what that means for all of us.

268.6 -> First, machine learning will solve problems

272.67 -> that could not be solved before with software and analytics and big data,

277.56 -> and in fact, machine learning will push the frontier

281.11 -> in ways that many of us could not even imagine before,

286.48 -> making the world a safer, smarter, and healthier place,

292.57 -> which leads to the question,

295.11 -> what is it about machine learning that makes this true?

299.59 -> Now, before machine learning,

301.75 -> you could analyze tabular data and extract information from it.

306.36 -> For example, you could look at your historical sales data

309.97 -> and predict your future sales,

312.38 -> and you could do that because tabular data has a nice structure to it,

316.38 -> so you can write a software program,

318.14 -> you can write programmatic rules to extract information from the tabular

322.55 -> and to act on that information.

325.6 -> However, most of the data in the world today is not tabular data.

332.72 -> In fact, more than 80% of the data generated in the world today

337.7 -> is unstructured data –

339.66 -> it’s audio, it’s video, it’s specs, it’s images, it’s 3D point cloud.

345.6 -> So most of the information and most of the insights

349.39 -> that we want to extract today is embedded inside unstructured data,

355.12 -> and as a result,

357.07 -> you cannot use software and traditional analytics to do that,

360.68 -> because it’s very hard to write programmatic rules

364.12 -> to extract information from unstructured data.

367.77 -> Think of a physician.

369.66 -> A physician needs to extract insights from MRIs, from x-rays,

375.3 -> from patient prescriptions.

377.66 -> You can’t write a software program to do that,

380.38 -> and so these had to be done manually.

383.99 -> Now with machine learning, not only can you extract insights

389.47 -> from structured data, but more importantly,

393.85 -> machine learning can help detect patterns in unstructured data,

398.01 -> and at a fundamental level, machine learning takes inspiration

402.32 -> from the learning process in the human brain,

406.53 -> and so just as we humans can read text,

411.17 -> can look at videos and look at images and listen to audios,

414.71 -> and extract information and insights from it,

417.65 -> and then act on that information, just like that,

422.76 -> machine learning can read text, can look at images,

425.68 -> can look at videos, can listen to audio,

428.7 -> and extract information and insights from it,

431.02 -> and then act on that information,

434.15 -> and that is what make machine learning uniquely

437.69 -> capable of solving problems that could not be solved before.

443.68 -> To illustrate, let’s look at a few customer examples

448.72 -> of deploying machine learning at scale in the real world,

453.6 -> and how these customers are pushing the frontier in their domain.

458.24 -> Let me start with computer vision,

460.72 -> which deals with extracting information from images and videos.

466.16 -> In many parts of the world, the kinds of government-issued IDs

470.2 -> that are used for loan applications aren’t easily available,

475.38 -> and so servicing the loan applications of these individuals

479.16 -> who do not have government-issued IDs

482 -> can be very difficult and can get dragged on for weeks,

486.35 -> and it’s to serve these people that Aella Credit,

490.32 -> a digital financial services company operating

492.81 -> in Sub-Saharan Africa, was created.

497.61 -> By using facial recognition technology from Amazon Rekognition,

503.24 -> Aella Credit is able to verify applicants

506.54 -> even when those applicants do not have other kinds of IDs.

512.28 -> In fact, by using Amazon Rekognition,

517.09 -> Aella Credit has been able to extend credit

520.36 -> to more than 2 million individuals and microbusinesses,

525.53 -> many of whom would not have an access to credit,

530.33 -> and this is making a big change to their lives.

535.01 -> Now customers are pushing the frontier

537.75 -> in computer vision in other ways as well.

542.04 -> For example, many customers today use automated ways

547.07 -> of extracting information from different kinds of IDs,

550.07 -> like your passport and your driver’s license,

553.77 -> but the problem is today’s automated solutions use specialized templates,

558.74 -> and so these automated solutions

560.55 -> do not work well across different kinds of IDs,

563.32 -> like across your passport and your driver’s license,

566.97 -> and that is because these IDs have different formats,

570.02 -> so the template does not work well across different formats.

574.76 -> Consider the case of Curative, it’s a COVID-19 testing company.

579.87 -> Curative needs to extract information

582.2 -> from both your ID and your medical insurance card

587.22 -> to be able to process insurance claims in compliance with HIPAA.

591.82 -> Now, Curative wanted to build an automated solution

595.2 -> but found that it doesn't scale,

597.34 -> because it doesn't work well across different kinds of IDs,

600.95 -> which differ in format from state-to-state

603.52 -> and sometimes even year-to-year,

607.25 -> and so I am very happy to announce an extension to Amazon Textract

612.68 -> that lets customers extract information

616.21 -> automatically from different kinds of IDs.

621.57 -> With these enhancements to Amazon Textract,

625.22 -> customers will now be able to extract information

628.23 -> from different kinds of IDs, like passports and driver’s licenses

632.42 -> issued in the United States, in a fully automated manner,

637.12 -> in near real-time and with very high accuracy.

642.95 -> These features use machine learning to understand the context of the IDs,

649.21 -> and then use that understanding to extract different information

653.79 -> like a name, your date of birth,

656.31 -> the expiration date of the ID, and so on, and using these features,

661.58 -> companies like Curative are now looking to automate ID analysis

666.68 -> and improve their business process workflows.

671.47 -> Another big domain that is being transformed by AI

675.75 -> is natural language processing,

677.33 -> which deals with extracting information from texts and documents.

682.16 -> In fact, natural language processing

684.41 -> was a largely unsolved domain before AI came into the picture,

689.74 -> but since the advent of AI, there has been a sea-change

695.26 -> in how customers are starting to use natural language processing.

700.34 -> Take the case of Slack, Slack has a feature called Clips

705.88 -> that users can use to upload voice and video clips

709.21 -> into any Slack channel.

711.77 -> Slack uses Amazon Transcribe, one of our language AI services,

716.65 -> to caption these clips so that users who are hard of hearing

722.6 -> can have a more meaningful Slack experience.

726.46 -> And not just that, these captions are searchable,

729.41 -> which means they increase the organization’s knowledge base.

736.47 -> Getting to healthcare now, there are so many frontiers to push.

742.71 -> For example, Cerner is using machine learning models in SageMaker

747.95 -> to predict patients who may be at risk of opioid use disorder.

752.77 -> To talk more about this, I would like to welcome

755.43 -> Ashleigh George from Cerner.

757.59 -> [music playing]

761.895 -> [applause]

766.75 -> Hello, hi, I’m Ashleigh George,

769.23 -> Vice President of Clinical Products at Cerner Corporation.

773.18 -> Cerner is responsible for helping

775.21 -> to provide healthcare information technology

778.57 -> to caregivers around the world.

781.32 -> We have been working to improve the electronic health record

785.36 -> for over 40 years,

786.96 -> helping to ensure that caregivers have the right data,

790.15 -> the right information, to best care for their patients.

795.41 -> As we approach machine learning,

797.65 -> some of the things that we think about at Cerner

799.78 -> is to ensure that we are thinking about machine

801.97 -> learning to provide new insights and information

805.76 -> that could be presented to our caregivers in near real-time.

810.28 -> We look at all the information

811.77 -> that’s coming from the vast information about our patients,

815.47 -> and how could we use this information for new insights

819.03 -> about our patients to be able to help predict

822.28 -> or even prevent treatment in the future.

826.15 -> As we think about this as well,

828.14 -> we’re thinking about how can we use this predictive information

832.72 -> to lead us to new diagnostic and information for our patients,

837.77 -> and ultimately, we want to make sure

840 -> that any of the machine learning information we’re bringing forward,

843.64 -> we present it in a way to our caregivers at the point of care

847.43 -> so that they can intervene right at that point of time.

852.17 -> So let’s now talk about one of the real-life examples

857.19 -> around opioid use disorder.

860.6 -> For those of you who have probably seen in the news,

865.32 -> maybe you have loved one or family members

868.07 -> or even you have been a caregiver

870.3 -> for someone who has been impacted by the opioid crisis,

873.88 -> it’s very a sombering topic but very real and relevant.

878 -> It is estimated that over 16 million people around the world

883.1 -> are afflicted with an opioid use disorder,

886.67 -> and yet of that 16 million,

889.544 -> less than 10% are receiving treatment

893.97 -> or being able to intervene for this disorder,

897.35 -> and of that then we see a staggering number of deaths

900.05 -> that continue to occur because of the opioid crisis.

904.56 -> So at Cerner, we asked ourselves could we apply machine learning

909.81 -> and be able to think about this differently

912.44 -> or how can we best intervene and be able to help advance

916.67 -> and think about the ways that we can use machine

919.03 -> learning as a predictor, so that's what we did.

922.73 -> We created an opioid use predictor disorder for us

927.36 -> to be able to think through

929.4 -> and be able to have the information come forward to our caregivers.

934.22 -> Built within the Amazon SageMaker,

937.61 -> we took variable different types of data,

940.87 -> over 40 different datas from the electronic medical record,

944.74 -> and we’re targeting the emergency departments

947.45 -> so that when a clinician is presented

950.51 -> with an information from their patient,

952.92 -> they are able to see this information and see if they are at a high risk

956.74 -> for needing to have any type of treatment.

960.02 -> This is currently in validation at our testing partners,

963.32 -> and we anticipate that it will be generally available in 2022.

968.49 -> So now let me take you through a little bit of the workflow.

973.25 -> If a patient comes in through the emergency department,

976.99 -> what is occurring is the patient is being cared for

980.5 -> and triaged by the caregiver, information is already collected

984.91 -> about this patient, and through this,

987.5 -> an algorithm is generated through the machine learning

990.44 -> and presents a score to that caregiver.

993.45 -> From here, based on the risk or how high the score is,

997.57 -> the caregiver then is able to intervene at that point in time

1001.69 -> and recommend the appropriate treatment path for that patient.

1006.31 -> This is for us the ability for how we can once again intervene early

1011.36 -> and help reduce any type of opioid disorder events

1015.54 -> or any other types of patient harm.

1019.19 -> Once again, this is just one example in terms of how we are thinking

1023.28 -> and approaching and using machine learning.

1026.15 -> We believe that there are endless amounts of possibilities

1028.87 -> of how machine learning can help to continue to improve outcomes,

1033.55 -> not only for our patients, but help our organizations

1036.4 -> to also improve operational and financial outcomes.

1040.94 -> Thank you.

1042.62 -> [music playing]

1048.79 -> [applause]

1054.66 -> Thank you, Ashleigh, truly extraordinary work at Cerner.

1059.4 -> Now, part of having a good life is not just having good health,

1064.95 -> but also having fun, relaxing,

1068.01 -> and knowing more about the world around us,

1070.87 -> and as you may have guessed,

1072.21 -> machine learning is also pushing the frontier there.

1076.55 -> Discovery is a great example of a company that's using machine

1080.68 -> learning to push the frontier on user experience.

1084.61 -> Discovery recently launched

1086.25 -> their first direct-to-consumer service called Discovery+,

1090.16 -> and the team had a challenge.

1093.25 -> They had 20 million customers, more than 2,500 shows,

1099.46 -> in 220 countries, and 50 different languages.

1104.88 -> Discovery knew that their human editors alone

1107.76 -> could not curate the content

1109.27 -> and provide a good tailored experience to the users,

1113.19 -> and therefore Discovery turned to Amazon Personalize,

1118.94 -> a fully managed service that lets you build

1122.17 -> your own recommendation system in a fully managed manner.

1126.1 -> Discovery used Amazon Personalize

1129.76 -> to build a recommendation system for their Discovery+ app

1133.71 -> and was able to increase user engagement by 3x.

1143.12 -> Like Discovery, many other customers

1146.5 -> have similar ambitions of increasing their audience engagement,

1150.81 -> but these customers tell us that today’s personalization tools

1154.66 -> that they use are just not accurate enough

1158.12 -> because they do not keep up with changing user preferences,

1162.29 -> and therefore we are launching two new additions

1165.02 -> to Amazon Personalize that address the needs of these customers

1169.19 -> and further push the frontier on user personalization.

1176.79 -> The first is prebuilt recommenders for Amazon Personalize

1182.14 -> that provide popular recommendation capabilities

1184.98 -> in a fully turnkey manner.

1187.61 -> These are recommendation capabilities like top picks for you.

1191.99 -> Think of a situation where a person has seen a movie,

1194.75 -> and you want to get a list of movies

1196.53 -> that would be most relevant for this person.

1200.1 -> Now, customers want their marketing professionals,

1202.69 -> their website managers,

1204.07 -> their online merchandising managers to use personalization

1208.78 -> without having to become an expert in machine learning,

1212.86 -> and to address these personas,

1214.87 -> we added these prebuilt recommenders into Amazon Personalize

1218.97 -> that can be used in a fully turnkey manner

1222.39 -> without needing to know any machine learning,

1226.17 -> and Amazon Personalize does all of the heavy

1229.52 -> lifting of training models on your data,

1232.81 -> and then hosting your recommendation system

1236.14 -> so that your product recommendations remain fresh,

1239.16 -> even when the user behavior changes or your product catalog changes.

1251.6 -> The next is user segmentation in Amazon Personalize.

1258.27 -> Many users want to provide

1261.988 -> highly-tailored marketing messages

1266.28 -> so that they can engage their users

1268.4 -> and increase the user conversion, however unfortunately,

1273.01 -> today’s user segmentation tools are based on static predefined rules,

1278.88 -> which means today’s user segmentation rules

1281.25 -> do not keep up with changing user behavior,

1284.43 -> and as a result, customers have to spend a lot of time

1288.12 -> tweaking these rules to keep up with changing user behavior.

1292.76 -> With user segmentation in Amazon Personalize,

1296.09 -> you can now engage your users

1298.26 -> even when their behavior changes dynamically,

1302.35 -> and you can use user segmentation to categorize your users

1306.16 -> into different categories

1308.43 -> based on their preferences in things like movies

1311.28 -> or even product genre or even product metadata,

1315.24 -> and you can use this to create highly-personalized marketing campaigns

1320.44 -> that engage your users more and ultimately

1323.68 -> improve your user conversion.

1328.06 -> Now, machine learning can be equally powerful

1332.2 -> in running cloud applications,

1334.76 -> because applications running the cloud

1336.86 -> can sometimes have anomalous behavior like increased latencies

1340.79 -> or increased error rates due to a variety of factors.

1344.78 -> Now in the cloud, it’s easy enough to log all events,

1349.59 -> but it can be daunting to use these logs,

1353.41 -> to use these event logs to find root causes of anomalies

1357.93 -> because you may end up logging millions of events per hour,

1362.94 -> and so essentially you end up trying to look for a needle in a haystack,

1368 -> and that is also where machine learning shines.

1373.95 -> That is why Amazon DevOps Guru uses machine learning models,

1378.18 -> informed by years of Amazon and AWS operational experience,

1383.34 -> to automatically find anomalies in your AWS applications,

1388.11 -> and not just that.

1390.12 -> When Amazon DevOps Guru finds anomalies in your applications,

1394.46 -> it gives you a list of potential root causes

1397.81 -> and it gives you a list of possible remediations

1401.14 -> so you can fix your applications quickly.

1405.17 -> Let’s look at a customer example.

1407.65 -> 605 is a TV advertising measurement company

1413.64 -> that helps customers optimize their TV advertising

1418.08 -> and reaches more than 21 million U.S.

1420.7 -> households.

1422.31 -> 605 had more than a dozen AWS accounts

1425.72 -> and tens of thousands of AWS cloud resources,

1429.2 -> and they were having a hard time

1431.11 -> correlating the metrics across all of these AWS resources,

1435.32 -> and therefore they turned to Amazon DevOps Guru,

1438.14 -> which uses machine learning models,

1440.27 -> to help them root cause issues in their applications,

1443.62 -> and using Amazon DevOps Guru, 605 was able to reduce,

1448.57 -> significantly reduce the meantime to application recovery.

1454.95 -> Another industry that generates a lot of data

1459.17 -> is the manufacturing industry.

1461.21 -> In fact, applying machine learning to sensors

1464.56 -> and other data generated by industrial equipment

1468.79 -> can be a game-changer in the manufacturing industry.

1473.97 -> Let’s look at a customer example.

1478.48 -> Koch AG & Energy is a wholly owned subsidiary of Koch Industries,

1483.58 -> one of the largest private companies in the world.

1487.63 -> Koch AG wanted to be able to proactively detect

1490.82 -> potential failures in their equipment,

1493.92 -> and therefore they turned to Amazon Monitron

1497.27 -> and Amazon Lookout for Equipment, two machine learning services

1502.16 -> that can proactively detect potential failures in their equipment

1506.23 -> and alert you to them before it impacts their users.

1510.72 -> In fact, by using Amazon Monitron and Amazon Lookout

1515.16 -> for Equipment, Koch has been able to find potential failures

1520.03 -> in their equipment hours before any other monitoring method.

1524.82 -> For example, Amazon Monitron was able to alert Koch

1530.09 -> to a potential issue in one of their nitrogen

1532.94 -> producing units by detecting abnormal vibrations using machine learning.

1541.14 -> More and more, the use cases that we have been talking

1544.83 -> about rely on data from sensors, from cameras, from robots,

1549.32 -> and from other edge devices that are located in places

1552.65 -> far away from a data center,

1555.91 -> but these devices capture massive amounts of data –

1559.48 -> audio, video, images, and so on –

1562.26 -> and so customers are very interested in doing machine

1565.41 -> learning on the data being captured in these devices,

1569.11 -> but unfortunately,

1570.29 -> it’s often not feasible to upload this data to a data center

1574.43 -> because of bandwidth limitations in a remote places,

1578.17 -> and so AWS provides you services

1580.9 -> that lets you do machine learning on edge devices,

1584.17 -> the newest member of which is AWS Panorama.

1589.74 -> AWS Panorama is a machine learning appliance

1592.72 -> that lets you do computer vision on on-premises IT cameras,

1597.97 -> and lets you analyze your video feeds in just milliseconds.

1603.4 -> Take the example of the Port of Vancouver,

1605.88 -> which is the third largest port in North America.

1609.91 -> The Port of Vancouver is using AWS Panorama

1613.48 -> to automatically track containers

1615.66 -> through the entire inspection process in the port.

1620 -> Now before starting to use AWS Panorama, the Port of Vancouver

1623.79 -> used to get a lot of complaints from the users

1626.91 -> because of delays in the inspection process

1629.6 -> or because these users did not have good visibility

1632.94 -> into the status of the inspections,

1635.54 -> and that was because the inspection process was largely manual.

1640.28 -> So the Port of Vancouver, in partnership with Deloitte,

1644.51 -> decided to automate this inspection process,

1648.25 -> and since this inspection requires looking at the containers

1651.85 -> and its contents and its conditions,

1655.08 -> the Port of Vancouver decided to use on-premises cameras

1659.54 -> and turn to AWS Panorama.

1663.36 -> By using AWS Panorama, the Port of Vancouver was able

1667.11 -> to automate the tracking of containers

1669.73 -> through this inspection process,

1671.9 -> and therefore reduce the amount of manual data entry

1675.92 -> in the inspection process, was able to improve the data

1679.63 -> reconciliation during this inspection process,

1683.34 -> and provide near real-time updates to the key stakeholders,

1687.96 -> and all of this reduced the wait time

1691.3 -> and improved space utilization at the port.

1697.52 -> Now, I wanted to discuss these real-life customer examples

1703.24 -> of using machine learning at scale to explain why machine

1708.93 -> learning is critical to innovation today,

1713.1 -> but as more people start doing machine learning,

1718.14 -> we need the fuel to keep the fire going,

1723.15 -> and for machine learning, that fuel is data.

1727.49 -> All of this machine learning will need tons of multi-modal data –

1732.02 -> audio, video, text, tabular, images, 3D, and so on –

1737.96 -> which brings me to the second key driver of machine

1742.02 -> learning innovation, and that is enabling the processing

1746.44 -> of massive amounts of multi-modal data.

1750.9 -> Let’s look at some customer examples of multi-modal data

1753.8 -> processing at work.

1757.62 -> The NFL has partnered with AWS to create the Digital Athlete Program

1763.45 -> that uses machine learning to track helmet collisions

1768.05 -> and identify risks during games.

1771.21 -> This requires labeling hours of video footage

1774.4 -> so you can train machine

1775.79 -> learning models to automatically track helmet collisions

1779.32 -> and identify risks during games.

1783.43 -> Thomson Reuters has more than 150 years of rich data

1789.12 -> on tax, on law, on news and other aspects,

1794.07 -> and they want to use this data to train machine learning models.

1797.8 -> For example, Thomson Reuters is today

1800.39 -> using tens of thousands of documents to train machine

1804.63 -> learning models to do natural language question answering,

1809.1 -> and Intuit has more than 275 million minutes

1814.81 -> of customer conversations every year

1818.29 -> that they want to analyze to gain insights,

1822.46 -> and Aurora uses simulations to generate massive amounts of video

1826.84 -> and 3D data that they use

1829.32 -> for training highly accurate perception models

1831.68 -> for their autonomous driving.

1835.4 -> Now, to create machine learning models from this data,

1838.53 -> all of these customers need to label this data,

1842.56 -> and as the demand for machine learning has grown,

1845.62 -> so has demand for data labeling.

1849.17 -> Now today, customers often use teams of data operations managers

1853.63 -> or program managers to do the labeling work for them,

1858.71 -> but these teams have to manage the labeling workforce,

1862.12 -> they have to set up the data labeling jobs,

1864.29 -> they have to validate the quality of the labeled data,

1868.07 -> and all of this can be daunting,

1870.26 -> especially when the volume of data has grown,

1873.9 -> and therefore, we are launching SageMaker Ground Truth Plus,

1877.7 -> which provides a fully turnkey experience for data labeling.

1885.2 -> Here is how SageMaker Ground Truth Plus works.

1891.18 -> As a customer, you bring your raw data

1894.72 -> and you give us your labeling instructions,

1897.46 -> and then Ground Truth Plus takes it over from there.

1901.23 -> It looks at your label instructions, for example, if you want your data

1905.3 -> to be labeled by experts in video labeling,

1908.38 -> Ground Truth Plus will only send your data

1911.17 -> to workers who are proficient in video labeling.

1915.01 -> It then manages the workforce on your behalf,

1918.13 -> it does the automation of the label data on your behalf,

1921.95 -> and then hands back the label data to you.

1924.64 -> In essence, you handed the raw data, and Ground Truth

1928.2 -> Plus gives you the finished validated output back to you.

1932.91 -> And not just that, Ground Truth

1935.29 -> Plus embeds machine learning into data labeling,

1939.45 -> so for example, Ground Truth

1941.5 -> Plus uses machine learning models to prelabel the data,

1946.42 -> so that human labelers don't have to do any more labeling,

1949.45 -> they just have to verify that the labels being done by the machine

1954.42 -> learning models are correct,

1956.69 -> and this can reduce the cost of data labeling by up to 40%.

1965.04 -> Now as we think of all these transformative use cases,

1970.2 -> it’s natural to ask

1973.734 -> how do we make sure that more people can do ML?

1979.9 -> How do we make sure that machine learning

1982.35 -> is not just restricted to machine learning scientists

1985.43 -> and data scientists, but is more broadly accessible

1989.65 -> so that more employees can be part of the machine learning transformation?

1995.32 -> And that brings me to the third key driver for machine

1999.68 -> learning innovation,

2001.42 -> which is empowering more people to do machine learning.

2008.64 -> According to the LinkedIn Jobs Report in 2020, the demand for AI

2014.03 -> and ML practitioners has been growing by 74%

2017.79 -> annually for the last four years,

2020.87 -> more than 2x that of any other job category,

2027.1 -> and given the early nature of machine learning,

2029.68 -> it’s no surprise that customers are having a hard time hiring

2034.32 -> all the people that they need to hire to get all of the machine learning done.

2040.05 -> So these customers asked us,

2043.48 -> “Can you expand the audience for machine learning?

2047.21 -> Can you provide us tools that make it easier for more employees

2051.17 -> to do machine learning?”

2054.25 -> And that got us thinking, how do we change the paradigm fundamentally

2061.14 -> so that machine learning can be done by data analysts,

2063.9 -> by sales and marketing professionals, by HR and Finance professionals,

2068.77 -> employees who use data, who understand data,

2071.57 -> who will benefit from machine learning insights,

2076.44 -> and we came up with a novel solution, but before I get into the solution,

2080.37 -> let me give you an analogy to explain how we have been thinking about this.

2085.48 -> If I can take you back to the late nineties,

2087.43 -> when the internet boom started, creating a website then wasn’t easy.

2093.91 -> You needed to write a bunch of HTML code,

2096.8 -> and only coders could do it,

2099.99 -> and then came a set of no-code tools that let you build websites.

2105.8 -> You could just point and click, and drag and drop,

2108.07 -> and build your own website,

2110.27 -> and boom, there was an explosion in the number of websites.

2113.48 -> Some estimates say today there are more than 2 billion websites,

2119.3 -> and that got us thinking.

2121.55 -> How do we change the paradigm

2123.34 -> and make it possible to build and deploy machine

2126.24 -> learning models without having to write any code?

2130.77 -> And that gets me to Amazon SageMaker Canvas,

2134.16 -> a no-code extension of Amazon SageMaker

2137.75 -> that lets you build and deploy machine learning models

2141.38 -> without having to write a single line of code.

2147.81 -> Canvas make machine learning accessible to data analysts,

2152.02 -> to other employees in the company

2154.53 -> who may not be proficient with coding,

2157.07 -> who may not be proficient with machine learning,

2160.33 -> but who nevertheless want to use machine

2163.23 -> learning and benefit from its insights.

2167.11 -> With SageMaker Canvas, users can access multiple sources of data,

2171.95 -> on premises or in the cloud,

2175.27 -> and then Canvas will automatically build the right models for them,

2180.41 -> will create the models for them, will deploy the models for them,

2185.63 -> and all of this without having to write a single line of code.

2193.41 -> In effect, we completely redid SageMaker

2198.524 -> for a new customer persona.

2201.75 -> To show how Canvas works, I would like to invite Kimberly Madia

2205.68 -> from the AWS Product Marketing team.

2209.657 -> [music playing]

2214.92 -> [applause]

2220.96 -> Thank you, Bratin.

2222.67 -> As a Product Marketing Manager, I deal with a lot of data every day.

2228.31 -> I need to use this data to measure the impact

2231.35 -> and effectiveness of my sales and marketing campaigns.

2235.6 -> For example, I look at visits to my webpage

2239.26 -> and try to figure out among all those who visited,

2242.54 -> who are the right people to offer our promotions to.

2245.62 -> Now, I know machine learning is the way to go,

2249.04 -> but I’m not a machine learning practitioner,

2252.29 -> so I don't typically build, train,

2255.11 -> and deploy machine learning models as part of my everyday job.

2264.11 -> More specifically, here’s my problem: I need to be able to predict

2265.93 -> if leads in my marketing pipeline will convert

2268.53 -> and become paying customers or not, and now with Amazon SageMaker Canvas,

2274.23 -> I am able to generate these predictions all on my own

2277.78 -> even though I don't have much experience with machine learning,

2281.49 -> so let’s dive in and see how it works.

2285.41 -> The first step is for me to connect to the SageMaker Visual Experience

2290.34 -> and connect to the data sources

2291.77 -> containing my sales and marketing data.

2295.19 -> SageMaker Canvas will automatically discover

2297.94 -> the data sources that my account has access to,

2300.86 -> such as Amazon S3 and Amazon Redshift.

2304.51 -> I am also able to drag and drop files from my local computer

2308.29 -> into SageMaker Canvas, and that's not all.

2311.28 -> Canvas also comes with built-in connector

2313.83 -> to third-party data sources.

2317.32 -> Now that I’ve connected to my data sources,

2319.68 -> the next step is for me to create a single unified data

2323.25 -> set that I can use to train my prediction model.

2327.6 -> In this case, I am going to join some web traffic data

2331.64 -> with customer information in Amazon S3,

2334.5 -> such as the unique lead identifier.

2337.69 -> I am able to visualize this data join to make sure that it was correct,

2342.9 -> and that my data is ready for machine learning,

2346.27 -> and SageMaker Canvas makes this very easy for me to do

2349.18 -> because it will automatically detect and correct errors,

2353.23 -> such as filling in missing values

2355.49 -> or removing duplicate rows and columns.

2360.06 -> The next step is for me to specify the target that I want to predict,

2365.02 -> in this case, if a lead will convert and become a paying customer or not,

2369.38 -> and I’m able to do that very simply

2371.17 -> from a pull-down menu in SageMaker Canvas.

2375.67 -> Now that I’m ready to go, SageMaker Canvas

2378.27 -> will automatically generate my model for me,

2381.06 -> based on my use case and based on my data.

2384.94 -> I can see in this case that my model has an accuracy of about 90%,

2390.1 -> and I feel really good about this for my use case

2392.6 -> so I’m going to go ahead

2393.81 -> and put this model to work for my sales and marketing campaigns.

2397.81 -> The first step is for me to go into the analyze view

2400.93 -> and explore all the different model

2402.66 -> inputs that went into making the prediction.

2406.35 -> This is known as model explainability,

2409.02 -> and model explanability is super important for me

2411.41 -> because I want to earn trust and better collaborate

2414.06 -> with my stakeholders

2415.6 -> by explaining the how and the why of my prediction.

2419.47 -> In this case, I can see that if a lead participated in a promotion,

2423.43 -> they are very likely to convert and become a paying customer,

2428.7 -> so let’s see what happens if I offer a promotion

2431.48 -> to a particular lead in my pipeline,

2434.38 -> and I’m going to do this by performing a what-if analysis.

2438.12 -> I simply change the promotion field from no to yes,

2442.88 -> and I can see that the prediction changes

2445.13 -> from not converted to converted, therefore it makes a lot of sense

2450.06 -> for me to work with my sales and marketing colleagues

2452.96 -> to put together a promotion for this lead in my pipeline.

2456.89 -> And that's not all, I am able to share my models

2461.36 -> and my data sets with the data science teams

2463.93 -> who are using Amazon SageMaker Studio.

2466.79 -> This is really important because I definitely want to make sure

2469.21 -> that my models are compliant with corporate standards and guidelines,

2473.43 -> and I also want to get some really helpful information

2475.61 -> so that I can improve my model.

2477.84 -> In this case, the data science team realized

2479.96 -> I was missing an important data set,

2482.03 -> so I am very easily able to add that to SageMaker Canvas

2485.35 -> to improve my model.

2488.38 -> So overall with SageMaker Canvas,

2490.39 -> I have everything I need to build machine

2492.53 -> learning model predictions all on my own,

2494.38 -> and I am able to collaborate with my stakeholders

2497.21 -> by explaining the how and the why of my predictions

2500.44 -> so that I can better meet my business goals,

2502.96 -> and Bratin, I’m sure you will be very happy to hear

2505.26 -> we now have a machine learning-based approach to our marketing campaigns.

2509.24 -> Thank you very much.

2511.46 -> [applause]

2517.37 -> Thank you, Kimberly.

2520.13 -> I think Canvas will be a game-changer

2522.47 -> in making machine learning accessible to more employees,

2526.67 -> but we are doing even more to help people

2529.98 -> get started with machine learning, and so therefore to help students

2535.64 -> and other people who want to get started with machine learning

2539.23 -> and just want to experiment with machine learning,

2542.41 -> we are launching Amazon SageMaker Studio Lab,

2547.14 -> a no setup, no charge machine learning environment.

2554.77 -> Studio Lab providers you a Jupiter notebook,

2560.03 -> integrated with GitHub,

2562.09 -> and then packaged with all the popular machine learning tools

2566.34 -> so that students and others can quickly get started.

2570.81 -> In fact, you don't even need an AWS account to get started.

2575.99 -> You can just use your email address to get started with Studio Lab,

2581.8 -> and Studio Lab not only gives you free compute,

2585.99 -> it also gives you free storage,

2589.6 -> and then when you are done with your work,

2592.95 -> you don't have to worry about shutting down your instances

2595.98 -> or saving your model or saving your data

2599.18 -> because Studio Lab does all of that for you.

2603.1 -> It’s as simple as closing your laptop,

2606.45 -> and then coming back to it again

2607.92 -> and resuming your work when you want to.

2611.72 -> We have many ways to help you get started with Studio Lab,

2615.38 -> including a chance to enter the Guinness Book of World Records.

2619.61 -> Starting today, you can enter the Studio Lab Hackathon,

2623.82 -> and this is a special event because we are trying to create

2627.32 -> the world’s largest virtual hackathon.

2631.34 -> We are looking for 5,000 hacks, and I hope you will join us there.

2637.64 -> I’m excited about how Canvas and Studio Lab

2642.96 -> are going to make it much easier for people

2646.18 -> to get started with machine learning,

2648.18 -> especially those who are early on in the machine learning journey,

2653.47 -> but to get to business value,

2656.52 -> machine learning needs to get integrated

2659.03 -> into every aspect of a company’s operations,

2663.27 -> and that gets me to the final key driver of machine

2667.44 -> learning innovation,

2669.8 -> and that is industrializing machine learning to scale its deployment.

2676.57 -> We have seen this industrialization play out in other industries as well.

2682.91 -> The automotive industry is a great example.

2688.48 -> The assembly line industrialized automotive design and manufacturing,

2694.18 -> and moved those from one of hand-made cars to mass-produced cars,

2699.03 -> effectively launching a revolution in transportation.

2703.55 -> The software industry went from a few specialized business applications

2708.21 -> to becoming ubiquitous in our lives through automation,

2713.2 -> through standardized tooling and standardized processes,

2716.68 -> in effect through the industrialization of software,

2720.77 -> and just in that way machine learning also needs to industrialize,

2725.84 -> and at AWS we have been at the forefront of that machine

2729.94 -> learning industrialization.

2733.29 -> To set some context, let’s look at how machine learning has grown.

2739.07 -> On AWS today, customers deploy millions of models,

2745.55 -> and they train models with billions or tens of billions of parameters,

2750.55 -> and they make hundreds of billions of predictions per month,

2757.22 -> so when we are talking about millions and billions

2759.77 -> and hundreds of billions,

2762.5 -> that's says machine learning is no longer a niche

2765.49 -> and has to industrialize.

2769.19 -> Now, machine learning industrialization

2770.92 -> has three components.

2773.09 -> First, is an infrastructure that is tailor-made for machine

2776.14 -> learning so that you can get the best performance and cost.

2780.66 -> Remember, if we are talking about producing million

2782.91 -> and billions of something, you need an infrastructure

2786.08 -> that can produce it quickly and product it at low cost.

2790.4 -> The second is machine learning tools,

2794.5 -> and these are tools that are purpose-built for machine

2797.57 -> learning so that they can reduce the heavy lifting

2800.7 -> in the machine learning workflow.

2803.2 -> These are tools like machine learning IDEs,

2805.55 -> like machine learning bias detection tools,

2807.71 -> like machine learning project management tools,

2810.36 -> and debuggers, and profilers, and so on,

2813.99 -> and all of these are integrated into a cloud-based machine

2817.66 -> learning service so that you get the best

2820.2 -> and most performant cost-effective ML infrastructure.

2826.48 -> The third is automating your machine learning workflow,

2830.28 -> also called MLOps.

2833.58 -> Let’s look at each one of these

2835.17 -> starting with machine learning infrastructure.

2838.77 -> On compute infrastructure, we have been releasing more performance

2843.19 -> and better GPUs and custom processors

2845.97 -> so customers can do faster and cheaper machine

2849.87 -> learning model training and inference.

2853.8 -> On training, just yesterday we launched AWS Trainium,

2858.319 -> the second custom-built machine learning chip from AWS.

2864.63 -> AWS Trainium will provide the best cost performance for training machine

2870.22 -> learning models in the cloud, and I think it will be a game-changer

2874.35 -> for training deep learning models in the cloud.

2877.95 -> For GPUs, our P4D instances continue to remain

2881.88 -> the most powerful GPU instances in the cloud,

2885.32 -> improving performance by up to 2x and reducing cost by up to 60%

2891.65 -> over the previous generation P3 instances.

2895.2 -> And just this year, we also launched the EC2 DL1 instances,

2899.57 -> which are based on the Gaudi processors from Habana Labs,

2903.48 -> which reduce price performance by almost 40%

2908.18 -> over comparable GPU instances.

2911.54 -> Now for inference, many of our customers

2914.6 -> continue to benefit from AWS Inferentia,

2917.57 -> a purpose-built chip from AWS for machine learning inference.

2922.51 -> In fact, customers like Airbnb, Prime Video,

2925.66 -> and Sprinkler get up towards 70% lower cost

2931.19 -> and 2x higher throughput than comparable GPU instances

2935.45 -> by using AWS Inferentia.

2938.34 -> And finally, our G5 instances that we also launched this year

2944.38 -> provide up to 3x higher

2945.72 -> throughput than the previous generation G4dn instances.

2952.02 -> But machine learning infrastructure

2954.23 -> is not just about hardware optimization.

2957.88 -> It also requires software optimization.

2961.65 -> Think of customers who have intermittent workloads,

2964.54 -> like restaurant recommendations from an online food delivery service,

2969.42 -> they don't need that recommendation to be running at midnight

2972.47 -> when no one is ordering food,

2974.93 -> and so these customers today create complex workloads

2979.12 -> to shut down their instances when there no traffic,

2982.64 -> and then to quickly bring these instances back up

2985.69 -> when there is traffic.

2988.34 -> Therefore, I am pleased to announce

2989.88 -> SageMaker Serverless Inference,

2993.74 -> a fully managed serverless offering from SageMaker.

3000.11 -> SageMaker Serverless Inference will automatically allocate

3003.97 -> an instance for you

3005.17 -> based on the needs of your machine learning model,

3008.16 -> then elastically scale it up and down based on the traffic,

3013.04 -> and best of all it’s pay per use,

3017.19 -> which implies no traffic means no charge.

3023.89 -> Now, infrastructure makes your machine

3026.71 -> learning faster and cheaper,

3029.29 -> but machine learning tools and machine

3031.4 -> learning ops make it your machine learning easier,

3034.93 -> they make your ML developers much more productive,

3038.91 -> and SageMaker provides you an integrated set of capabilities

3043.09 -> inside SageMaker Studio that make it much easier

3046.84 -> for developers to do machine learning.

3050.5 -> Let’s see how customers are using these.

3056.11 -> Vanguard has been able to automate their entire machine

3060.36 -> learning workflow using SageMaker,

3063.34 -> and they have been able to reduce model deployment time by 20x.

3070.54 -> AstraZeneca, by using SageMaker,

3074.51 -> has been able to reduce the time to deploy their machine

3077.94 -> learning environments from one month to five minutes,

3082.4 -> from one month to five minutes.

3086.99 -> NerdWallet has been able to reduce their machine

3090.06 -> learning training time by 75%,

3093.49 -> even when they have increased the number of models they are training,

3097.58 -> and Zendesk has been able to reduce their cost

3100.66 -> by almost 90% for inference.

3105.24 -> So customers are clearly getting a lot of benefit even today,

3110.1 -> but we want to do even more to help customers

3113.98 -> get more productive with machine learning,

3117.68 -> so let’s start with model training.

3120.76 -> Today, many customers spend a lot of time, weeks to months,

3126.65 -> to optimize their models, and they have to do this

3129.81 -> because there’s a lot of craftsmanship

3131.73 -> involved in optimizing these models,

3134.54 -> because these models can have billions of parameters,

3138.16 -> and therefore we are launching the Amazon SageMaker Training Compiler,

3144.6 -> which automatically optimizes your models for you

3149.79 -> and gives up to a 50% performance improvement.

3155.9 -> The SageMaker Model Compiler comes integrated

3158.89 -> with the default versions of TensorFlow and PyTorch,

3162.13 -> so that users of these popular packages

3164.73 -> can use the compiler by default and get its performance benefits.

3170.31 -> Now, I talked about machine learning building and optimizations,

3176.66 -> but customers still have to do a lot of manual work

3179.83 -> for machine learning deployment.

3183.11 -> For example, today machine learning engineers

3185.88 -> have to choose from over 70 instances to figure out which instance

3190.85 -> they want to do machine learning deployment on,

3193.68 -> then they have to configure that instance,

3196.37 -> then they have to run load tests on that instance,

3199.62 -> and then they have to optimize their model on that instance.

3203.6 -> All of this takes weeks of manual effort,

3206.21 -> ultimately delaying model deployment and time to market.

3211.79 -> Therefore, we are announcing the SageMaker Inference Recommender,

3216.18 -> that reduces the time to deploy models from weeks

3220.29 -> to just a few hours.

3223.51 -> The SageMaker Inference Recommender

3225.53 -> will automatically pick the right instance for you

3228.34 -> that gives you the optimal cost and performance,

3231.88 -> optimize your models for you,

3234.38 -> run load tests for you on the instance,

3237.12 -> and then configure the instance for you,

3240.06 -> and so all of that work that took weeks of manual effort now

3243.94 -> gets done automatically in just a few hours,

3247.52 -> which makes machine learning deployment much,

3250.46 -> much easier for customers.

3255.27 -> Finally, let’s look at the third component of machine

3258.15 -> learning industrialization, which is MLOps,

3261.69 -> and MLOps lets you do several key operational requirements.

3266.28 -> For example, it let you automate your entire machine learning workflow

3270.3 -> and make it repeatable, so that you don't introduce errors.

3274.71 -> It lets you enhance your governance and audit posture,

3278.32 -> so that you make your machine

3279.48 -> learning more transparent and increase confidence.

3283.69 -> Then finally, it fosters the responsible use of machine

3287.48 -> learning so that you can do things like bias protection,

3291.27 -> model explanability, and model monitoring.

3296.03 -> SageMaker supports all of these ML operational requirements

3299.82 -> through SageMaker Pipelines.

3303.73 -> SageMaker Pipelines is a fully managed feature

3307.07 -> that lets you automate your entire machine learning workflow,

3311.63 -> from processing data to building models

3314.82 -> to retraining and training models and deploying models,

3319.77 -> and not just that, SageMaker Pipelines lets you do CI/CD,

3324.77 -> continuous integration/continuous deployment,

3327.53 -> for your machine learning workflows just like you do with software,

3332.39 -> which makes your machine learning a lot more robust.

3337.05 -> SageMaker Pipelines also enhances your governance

3339.95 -> and audit posture,

3341.28 -> because it tracks your data, your models, your ML artifacts

3345.03 -> at every step of the machine learning workflow,

3347.84 -> and lets you do source and version control.

3351.18 -> Finally, I am excited to announce

3353.75 -> that SageMaker Pipelines now integrates bias detection

3357.95 -> into every step of your machine learning workflow,

3361.57 -> so you can detect bias in your data or your models,

3365.19 -> you can do model explainability, and you can also monitor your models

3370.63 -> so that you get alerted if there any quality degradations.

3378.76 -> One common theme you may have noticed

3382.8 -> as I was talking about the key drivers of machine

3385.26 -> learning innovation

3387.35 -> is that Amazon SageMaker has played a major role

3392.23 -> in establishing these key drivers of innovation

3395.19 -> and making them a reality.

3397.71 -> In fact, Amazon SageMaker is today

3400.79 -> one of the fastest growing services in AWS history,

3405.99 -> and we have been amazed and grateful

3408.88 -> to our customers for the adoption that we have seen.

3413.44 -> Just in the last year,

3414.87 -> we launched more than 60 new features and functionalities,

3418.7 -> that along with everything we launched at re:Invent,

3421.76 -> makes Amazon SageMaker the most comprehensive machine

3425.46 -> learning service that lets ML scientists, ML engineers,

3430.57 -> ML ops engineers, and analysts build, train,

3435.43 -> and deploy machine learning models.

3438.65 -> but it isn’t just SageMaker where we have been investing.

3442.13 -> We are continuously innovating and delivering new capabilities

3445.79 -> in our language AI services, in our vision AI services,

3449.61 -> in our health and industrial AI services,

3452.47 -> in our AI DevOps services,

3454.58 -> in our personalization and forecasting services,

3458.21 -> so together our AI Services, along with SageMaker,

3461.69 -> along with ML frameworks and infrastructure,

3465.21 -> are enabling hundreds of thousands of customers

3469.51 -> to transform their business with machine learning.

3474.71 -> Let me conclude with the main takeaway from this session,

3478.91 -> and bring you back to the lesson that I talked about

3481.28 -> at the beginning of this talk,

3483.89 -> and that is machine learning is no longer the future,

3488.59 -> machine learning is the present and here and now,

3493.26 -> and based on Amazon’s own experience

3495.58 -> and the experience of more than 100,000 of our customers,

3499.89 -> you can also transform your business through machine

3503.14 -> learning by using the four key drivers machine learning innovation.

3509.58 -> First, use machine learning to solve problems

3513 -> that you could not solve before;

3515.31 -> second, the processing of massive amounts of multi-modal data;

3521.97 -> third, empower more employees to build machine learning;

3528.09 -> and finally, industrialize machine learning across your business

3532.01 -> so you scale machine learning deployments.

3536.94 -> To learn more about the features that we are launching

3539.43 -> at re:Invent, please visit the AWS Machine Learning website,

3543.7 -> and don’t miss the deep dive AI/ML sessions,

3546.38 -> chalk talks, and workshops at the rest of this re:Invent.

3550.13 -> Thank you for coming, and enjoy the rest of re:Invent.

3552.621 -> [applause]

3553.807 -> [music playing]

Source: https://www.youtube.com/watch?v=egD6qsRapNY