How Machine Learning on AWS Can Transform Your Education Business AWS Webinar

Transcript

Hi everyone, I'm very happy to talk to you today. My name is Julien, I'm a principal evangelist focusing on AI and machine learning at AWS. Today, I'm going to talk to you about machine learning on AWS and how it can help you improve and transform your education business. We have a lot to cover, and we'll have some time at the end for Q&A, so let's get started. Just before we start, I want to point out that we have a series of webinars. This is the first in the series, and I'll be happy to run another one next week, a more technical one where I will show you how to actually use the services and we'll go through some demos. If you want to see how to build these things, stick with me and let's discuss this next week. We'll have two more webinars. One will focus on Amazon Polly, our text-to-speech service. If you have voice applications, this is one you don't want to miss on April 24th. Lastly, on May 10th, there will be another webinar on Amazon SageMaker, one of our newer services that lets you build end-to-end machine learning applications. It's a pretty technical one. You can register at the same URL you used to register for this one. If you want to learn more, just go to that URL, and we'll be happy to deliver those additional webinars. If you want to get in touch with us after this webinar because you have some projects you want to get going, this is the email address to write to: edtechteam@amazon.com. Please make sure to use that subject. It will help us find you quickly. Please give us as much information as possible about your project, and we'll get back to you quickly and put you in touch with the right people to help you start your successful machine learning project. edtechteam.amazon.com is the place to go. So, what are we going to cover today? First, we'll talk for a few minutes about why AI and machine learning matter and why you should really be looking into these technologies. Then, we'll discuss machine learning at Amazon to give you some examples of how we use these technologies to build advanced applications. The bulk of the presentation will focus on use cases for our different services, and I'll try to give you some examples and ideas of machine learning and AI applications in the educational space. Hopefully, this will get your creative juices flowing, and you can start building stuff as well. We'll also share some links and resources to help you get started successfully. AWS has been working in the public sector for a long time now. We have a large number of government agencies, educational institutions, and NGOs working on AWS. We also have a dedicated AWS region called GovCloud for federal agencies. So, we're not new to this. When it comes to customers, we're very proud to have all these fine institutions and companies as public customers. Of course, I can only talk about the public references. We have very cool startups like Coursera, and universities like Harvard, the University of Chicago, and Berkeley, among many others, as well as research labs. A mix of ed techs, universities, and research labs have been using AWS for years, not just for general infrastructure purposes, but increasingly for AI and machine learning. If I asked you how important it is for your organization to invest in machine learning and AI, most of you might answer that you're not a machine learning company or organization, and these things are not key to your business. You have more important topics to address, and that's perfectly fine. Most of the companies and organizations I talk to usually answer this way. But if I asked you how important it is to deliver a better learning outcome for students, to help students be more successful and learn better, most of you, and probably all of you, would say that this is your number one priority. We help people learn and be successful. But if you ask me, this is pretty much the same question, and the purpose of this webinar is to show you why this is really the same question. Amazon is famous for its flywheel, its continuous improvement, and its virtuous circle. You could design a similar flywheel for data. More data in general will help you build better analytics, understand your company better, and hopefully, using those insights, you can build better products and improve customer satisfaction or the user experience, optimize costs, or address whatever your top business problem might be. If you do a better job at this, it's likely that you will have more users. Your product will be more successful, more engaging, and easier to use. Having more users should bring you more data. This is where machine learning fits. Machine learning is about predicting the future using past data to make your product smarter. We'll see some examples during this webinar. Before we dive deeper, I want to clarify some definitions because these terms are often thrown around and not always used correctly. AI is a science that has been around since the mid-50s, and its purpose is to build software applications that exhibit human-like behavior. This can mean speaking, understanding speech, understanding natural language in spoken or text form, reasoning, and making educated guesses about future events. Within AI, there's a subdomain called machine learning, which is probably the most successful. Machine learning aims to teach machines to learn behaviors without being explicitly programmed. Instead of writing code that gives explicit orders to machines, we use standard algorithms that learn from data sets, which can be text, video, images, or anything else. Machine learning builds a model that helps a system predict what the future will look like and predict new samples. Within machine learning, there's another subset called deep learning, which uses neural networks. Neural networks are capable of understanding very complex, unstructured data like text, video, images, music, and voice. They can automatically learn the features and patterns present in data sets, something that machine learning cannot do without human intervention, known as feature engineering. With deep learning, you don't have to do that. Neural networks are smart enough to find the patterns, which is why this technology has been super successful in handling complex data like images and videos. In this scope, we'll focus mostly on machine learning, as it probably accounts for 99% of applications today. A good example is what Amazon has built over the years. On the amazon.com website, you'll see personalized content and product recommendations. If two people surf to the same page at the same time, they'll see different things based on their interests and Amazon's testing of different pages. Personalization, recommendation, and altering the layout of the website are important. We also have the orange robots in our fulfillment centers, which move products and shelves to the picking teams autonomously, avoiding collisions. The Echo family of devices uses AI and machine learning for natural language processing, speech-to-text, and text-to-speech. Amazon Go is a grocery store where you can walk in, pick items, and walk out without standing in line or dealing with cashiers. You receive your bill within a few minutes of leaving the store. This works because cameras inside the store use image recognition to see what you pick. The technology is hard to fool, and it's quite spectacular. Drone deliveries are still being tested but are expected to be available in the future. These examples are practical machine learning use cases built by Amazon for its own use, and we've been doing this for 20 years. When it comes to AWS, our mission is to bring these same capabilities to everyone else. More AI and machine learning is built and runs on AWS than anywhere else. Our AI and machine learning references include large companies like Netflix, Capital One, and the NFL, as well as smaller startups. These services are not just for the big players; we want every organization, even individual developers, to be able to use them. Our machine learning stack today looks like this. At the high level, we have application services, which are API services for vision, speech, and text processing. These are based on machine learning and deep learning, but you don't need to know anything about those to use them. They're just one API call away, and any junior developer can use them in 15 minutes. For more advanced users who want to train on their own data sets, build their own models, and tweak their own algorithms, we have platform services like Amazon SageMaker, which I'll discuss in next week's webinar. At the lowest level, you can use AWS instances to run your own training jobs and manage everything yourself, just like you would manage your own infrastructure. Let's look at a few use cases now, grouped by topic. In next week's webinar, we'll talk more about the services and give you some demos, but here I want to focus on actual use cases. How can image and video analysis help improve your education business? We've thought about this and have some examples. Please share your own ideas and examples during the Q&A. One example is ensuring that people taking exams are the right people. This is called a virtual proctor, where you compare images to actual faces. Another example is analyzing classroom reactions during a class, whether online or in person, to detect emotions on students' faces. This can provide real-time feedback to the professor, indicating whether students are confused, happy, or bored. You can also use this technology for school and campus safety. The main service we recommend here is Amazon Rekognition, which can do live real-time analysis on images and videos, including video streams. It can detect objects and scenes, perform facial analysis, and compare faces for authorization and identification. Here's a real-life example. This is a picture I took in India during a SageMaker workshop. The participants seemed pretty happy at the end, so we took a big selfie. Running this through Rekognition, we detected all the faces, found the positions of the eyes, nose, and mouth, and got extra information like whether each person has a beard, mustache, or glasses, and whether they look happy or sad. This is a visual representation of the information Rekognition sends back, but in reality, you get this information in the API response and can process it in your app. Another example is the city of Orlando Police Department, which uses Rekognition to monitor video streams from cameras deployed all over the city. They can detect missing persons, persons of interest, accidents, fires, and other dangerous situations in seconds, allowing them to dispatch the proper response. This technology can also be used for school safety and campus monitoring. A more advanced example is Onfido, a UK startup that built a face recognition model using TensorFlow and AWS GPU instances. Their main application is verifying identity by comparing a face to a photo ID. This is useful for virtual proctoring in online courses, where verifying the identity of participants is challenging. Let's talk about voice for a moment. Our text-to-speech service is called Amazon Polly. It supports 25 languages and 52 voices, including male and female voices for different languages. You can pass a text string, select a voice, and call an Amazon Polly API to receive a sound file in real time. This is fast enough for real-time conversation or can be archived and played later during course delivery. Polly has a free tier, and the Alexa voices on Echo devices are generated by Polly. You could use Polly to add voice to all your content, such as slides, web pages, and written content. This can work in archive mode, generating podcasts and audio content from historical classes, or in real time, generating audio content for slides and broadcasting it to listeners. This is especially helpful for students with visual impairments or disabilities. An example is TenMarks, an ed tech company that uses Polly to build voice instructions for math practice, making content more accessible to young children and those with disabilities. A personal example is Troy, a cloud architect based in the U.S., who built a system called PolyXy using AWS services like Polly. PolyXy is based on a Raspberry Pi and deployed across his home to help him and his wife interact with their son, who has severe autism. Using visual and voice signals, they can ensure he's okay and get what he needs. This system is a clever and efficient way to communicate and engage with special needs children. You can read more about this in a blog post and watch a YouTube video for a better explanation. Now, let's talk about natural language processing. You could analyze digital content to improve search and recommendation for students, organize and categorize vast archives of courses and content automatically, detect plagiarism by comparing paper topics, and analyze sentiment in chat rooms, surveys, and student comments. We have a new service called Amazon Comprehend, announced at re:Invent 2017, which can handle large collections of documents. It automatically extracts key phrases, entities, performs sentiment analysis, detects language, and builds topics. The Washington Post uses Comprehend to process their archives, extract keywords and entities, and personalize the reader experience, recommend content, and optimize search and ad tech. Let's talk about translation and transcription. Translation is an obvious use case for delivering content in multiple languages. You could translate content manually or outsource it, but it's time-consuming and expensive. Using an automated solution, you can affordably and automatically translate everything. You could build multi-language captions for video courses or MOOCs, translate user-generated content, and interact with foreign students using just-in-time translation. We have a new service called Amazon Translate, announced at re:Invent 2017, which is still in preview. It supports 12 language pairs, with more to come. It automatically detects the source language and translates it to the target language. For transcription, we have Amazon Transcribe, also announced at re:Invent 2017 and in preview. It converts speech to text, supports English and Spanish, and provides punctuation, formatting, and timestamps, which are important for video captions. It supports high and low-resolution audio, making it useful for call centers and phone call transcription. We have a large customer whose name will be revealed in the next few days. This company is capturing video lectures across the world, in 30 countries, for millions of students. Their goal is to transcribe all of it, providing text transcriptions of video lectures to their customers, such as educational organizations and universities. Once you have a text transcription of the video, you can run natural language processing, using Amazon Comprehend to extract entities, concepts, and use these for analytics, indexing content, and more. Imagine a professor giving a speech, and in near real-time, you have a transcription to text. This could be translated automatically and shared with non-English speakers using Amazon Translate. You could even generate a podcast in multiple languages in near real-time. Using Comprehend, you could extract information from the text, index it, and publish it to your website. This content could be available in multiple languages, in text form, as a podcast, and indexed on your website within 30 minutes of the class ending. The reach of this content and the number of students you could address would be significantly larger. Let's move on to chatbots. Chatbots are popular and not only for education. What can we do with them? We can use them to answer questions from prospective students and parents, providing 24/7 support for a school or university. This could even be an Alexa skill deployed on Echo devices, used on campus or within the school to simplify admin tasks and everyday life. Instead of creating tickets or calling people, you could talk to a chatbot to book appointments or request admin tasks. This would be a simple and intuitive communication channel between professors and students. They could ask questions even if the professor is not in the office, and the professor could answer later or book appointments. This fluidifies the relationship between professors, students, and administration. We have a service called Amazon Lex, which is easy to use. You can define the conversation between the user and the bot, with the purpose of extracting pieces of information called slots. Once all slots are filled, the bot will automatically invoke a piece of code running in the cloud to book an appointment or answer a question. An example of a customer using Lex is Freshdesk, a help desk company. They use it to centralize admin tasks and provide better customer support, simplifying ticket management and allowing admin staff to focus on higher-value tasks. Another customer is Liberty Mutual, an insurance company that uses Lex to let customers get car insurance quotes using a bot. You describe the car, and within a few sentences, you get a quote. This is an example of an application bot with many potential uses. Now, let's talk about advanced machine learning use cases. We'll look at more advanced ways to use your data to improve your education business. You could use a machine learning model to find students who should join a university or receive aid. With thousands of applicants, it's not easy to select them manually. Machine learning can help build personalized content and learning paths based on students' levels, grades, and interests, recommending classes and content to help them learn better and faster. You could also predict if a student will be successful and detect early if they need personalized help. Ivy Tech Community College uses their data to build models that identify successful and unsuccessful student behaviors. They can predict with over 80% accuracy which students are likely to fail within the first two weeks of the term. Early detection allows for timely intervention. Another example is Cirigo, which has built natural language processing on millions of articles, creating a knowledge graph to connect concepts and build personalized learning experiences for students. Edcast, known as the Netflix of learning, offers video content and learning paths with recommendations, similar to Netflix, but for serious learning. For STEM promotion, Bluespurse builds IoT kits and held a workshop with young students at re:Invent, integrating the kit with services like Polly and Lex to build fun projects. I built a Raspberry Pi robot called Johnny Pie, integrated with an Amazon Echo and several AI services, which can take pictures, recognize faces, and more. Instructions are on my blog on Medium, and I'm happy to help you build it for your students. We also announced DeepLens at re:Invent, a deep learning camera that lets you deploy image recognition models. It's a developer education tool, pre-equipped with sample projects, and you can pre-order it from Amazon. I've used it to educate my kids on deep learning, and it's a fun way to learn about image recognition technology. When getting started with these services, data security is crucial. AWS is compliant with numerous security and privacy certifications and regulations. For education, FERPA and COPPA are important. We have a white paper on FERPA compliance on AWS, and all mentioned services can be used for COPPA-subject applications. For more information, visit our website. To start, visit the URLs for the high-level machine learning page and the AI blog for technical examples. You can also get free training online at AWS Training, which includes introductions to all the services described. Registration is open for upcoming webinars, where we'll dive deeper into the services with demos and detailed discussions on Polly and SageMaker. If you want to get started, my colleagues are available at edtechteam@amazon.com. Follow me on Twitter at Jules Simon, and check my Medium blog for technical articles on machine learning and deep learning. Thanks for listening, and I hope to see you next week for the second webinar.