Amazon Web Services Inc. has become a powerhouse provider of cloud-based tools for enterprises to develop and deploy sophisticated artificial intelligence applications.

As we look ahead to the seventh annual re:Invent conference late this month, AWS continues to impress with its deep investments in building a sophisticated AI ecosystem:

• Optimization of core infrastructure-as-a-service cloud for AI workloads: The original AWS EC2 compute service now has dozens of instance types from general purpose to compute-optimized, memory-optimized, graphics processing unit-accelerated and field-programmable gate array-accelerated.
• Implementation of sophisticated AI accelerator hardware in the AWS cloud: AWS’ cloud provides on-demand access to the very latest GPUs to support machine learning and other AI workloads. The EC2 P3 instances offer the latest Nvidia Corp. Tesla V100 GPUs with up to eight GPUs per node. AWS claims this high density is 1.8 to two times better in price/performance and six times better absolute performance than GPU instances from any of the other cloud vendors.
• Improved AI development abstractions: Over the past several years, AWS has moved to progressively higher abstraction layers for AI, ML and other core compute services. All of these capabilities, plus the underlying cloud databases, are now available with cloud-native APIs for containerized deployment as microservices over the container orchestrator Kubernetes, as well as serverless deployment as AWS Lambda functions-as-a service. Most of these capabilities are programmable through Gluon abstraction layer, which enables visual development of the most sophisticated AI, ML and deep learning applications in the AWS-developed Apache MXNet, Microsoft Cognitive Toolkit and other open-source data-science workbenches.
• Sophisticated AI DevOps toolchains and libraries: With each passing year, AWS deepens its solutions to help enterprise developers address new AI, ML and deep learning challenges. In 2015, it launched its cloud-based ML service. At re:Invent 2016, it introduced pretrained models for text-to-speech with Amazon Polly, image classification and recognition with Amazon Rekognition, and conversational bots with Amazon Lex. At that time, it also launched Deep Learning AMIs, which are EC2 instances pre-installed with a variety of deep learning frameworks. At re:Invent 2017, AWS introduced Amazon SageMaker, which enables data scientists and developers to build, train and deploy machine learning models. It builds on the earlier efforts to further simplify the AI and machine learning process. This fully managed service enables developers to pull data from their S3 data lake, leverage a library of preoptimized algorithms, build and train models at scale, optimize them through ML-driven hyperparameter optimization capabilities, and deploy them in real-time into production EC2 cloud instances.
• Driving AI to the edge for embedding in mobile and Internet of Things devices: At re:Invent 2017, AWS launched enhancements to Greengrass for more sophisticated edge deployments, with the new AWS Greengrass ML Inference enables ML models to be deployed directly to mobile and other edge devices — such as Amazon Alexa — where they can drive local inferencing whether or not a device is currently connected to the cloud. It launched AWS IoT Analytics, a new service that supports easy analysis of IoT device data through AWS’ QuickSight solution as well as through AI models built in AWS SageMaker. And it released AWS DeepLens, a fully programmable video camera that developers can use – along with SageMaker, prebuilt models and code examples – to build and train video analytics for streaming in the AWS cloud.

Going into re:Invent 2018, Wikibon expects AWS to focus on several key themes for boosting the productivity of AI developers in its public cloud:

• Automating AI development
• Driving the democratization of AI
• Speeding AI access, development and operationalization
• Improving AI deliverables through a more sophisticated cloud data platform
• Leveraging deep data to improve conversational AI apps
• Sustaining a truly open AI development ecosystem

A year ago, Swami Sivasubramanian (pictured), AWS vice president of AI, discussed these developer ecosystem imperatives in his interview on theCUBE at re:Invent 2017. Here are some highlights:

• Automating AI development: “Our goal is to actually put machine learning capabilities in the hands of all developers and data scientists. That’s where SageMaker is an end-to-end platform that lets people build, click, train and deploy these models in a one-click fashion. It supports all popular deep learning frameworks. It can be TensorFlow, MXNet or PyTorch. We also not only help train but automatically tune where we use machine learning to build these things. It’s very powerful. The other thing we’re excited about is… API services… the new abstraction layer where app developers who do not want to know anything about machine learning, but they want to transcribe their audio to convert from speech to text, or translate it or understand the text, or analyze videos.”
• Driving the democratization of AI: “When I started getting into deep learning… I realized there’s a transformative capability of these technologies. It was too hard for developers to use it on a day to day fashion, because these models are too hard to build and train. That’s why we actually think of it as in a multilayered strategy where we cater to expert practitioners and data scientists. For them we have SageMaker. Then for app developers who do not want to know anything about machine learning they say, “I’ll give you an audio file, transcribe it for me,” or “I’ll give you text, get me insights or translate it.” For them we actually we actually provide simple to use API services, so that they can actually get going without having to know anything about what is TensorFlow or PyTorch.”
• Speeding AI access, development, and operationalization: “In Amazon we have been using machine learning for like 20 years. We have been leveraging machine learning for recommendation engine, in fulfillment center where we use robots to pick packages and then Alexa of course and Amazon Go. One of the things we actually hear is while frameworks like TensorFlow or PyTorch or MXNet are cool, it is just too hard for developers to make use of it. We actually don’t mind our users using Caffe or TensorFlow. We want them to be successful where they take from idea to product shell. And when we talk to developers, this process took anywhere from six to 18 months and it should not be this hard. We wanted to do what AWS did to the IT industry for compute storage and databases. We want to do the same for machine learning by making it really easy to get started.”
• Improving AI deliverables through a more sophisticated cloud data platform: “The data that goes into machine learning is going to be the determining factor for how good it is in terms of how well you train it and so forth. This is where data scientists keep saying the breath of storage and database and analytics offerings that exist really matter for them to build highly accurate models. When you talk about not just the data, actually the underlying database technology and storage technology really is important. S3 is the world’s most powerful data lackthat exists that is highly secure, reliable, scalable and cost effective.”
• Leveraging deep data to improve conversational AI apps: “Think of [Lex] as automatic speech recognition and natural language understanding [that is] pretrained on our data. But when to customize it for your own chatbots or voice applications, you can actually add your own intents and several things and customize the underlying deep learning model specifically to your data. You’re leveraging the amount of data that we have trained in addition to specifically tuning for yours. It’s only going to get better and better.”
• Sustaining a truly open AI development ecosystem: “In AWS we have tens of thousands of partners of course, right from ISV [and] SIs. The software industry is an amazing industry where it’s not [a] winner-take-all market. For example, in the document management space, even though we have S3 and WorkDocs, it doesn’t mean Dropbox and Box are not successful either. What we provide in AWS is the same infrastructure for any startup or for my team, even though I build probably many of the underlying infrastructure [services]. Nowadays for my AI team, it’s literally like a startup except I probably stay in an AWS building, but otherwise I don’t get any internal APIs. It’s the same API. It’s a level playing field.”

Since re:Invent 2017, AWS has addressed these requirements with a steady stream of strategic product releases, which have been covered extensively in SiliconANGLE. We’ve also covered them deeply in Wikibon’s recent big data market survey and other research studies.

To catch what AWS executives, partners and customers are saying now, get drill-downs on their forthcoming announcements and receive compelling glimpses into their roadmap going forward, tune into theCUBE live Tuesday through Thursday, Nov. 27-29.

Image: TK