Amazon’s AI Leadership Advances at re:Invent 2019

Impressive pace of innovation on show in Las Vegas

Amazon Web Services (AWS) was true to form at re:Invent 2019, as more than 65,000 customers, partners and analysts descended on Las Vegas for the company’s annual cloud computing event.

As has been the case in previous years, AWS announced an onslaught of over 100 new capabilities spanning many areas of cloud services. CEO Andy Jassy’s signature three-hour marathon keynote slot was once more a whirlwind of customer case studies and digs at the competition, backed by an ’80s-fuelled house band. Van Halen, Billy Joel and other blasts from the past provided the soundtrack to this year’s core themes of business transformation and how companies can get value from the AWS cloud.

Artificial intelligence (AI) and machine learning once again stole the spotlight, with more than 218 sessions covering the topics this year. Major announcements dealt with the most important enterprise challenges facing the technology: a lack of machine learning-related skills, technology complexity and finding the right applications.

Below, I look at some of the most impactful moves and assess what they mean for AWS’ strategy in AI.

AWS Expands Customer Set in 2019

“Tens of thousands” of customers are now deploying AWS’ machine learning tools, which the company claims is double the number of its nearest rival, including flagship customers such as NASCAR, Intuit and most recently, the Seattle Seahawks. Before the event kicked off, the NFL club announced it had selected AWS as its preferred cloud supplier, saying that the company’s AI is enabling better tracking of players, performance analytics and video analysis.

This highlights an important trend we’re seeing in cloud computing. As the market for data analytics and AI grows, the dominance of multicloud strategies is changing. According to CCS Insight’s recent annual survey of IT decision-makers, 57% of US and European businesses that are deploying AI now favour either a single cloud or a preferred cloud strategy when it comes to their data and machine learning needs (see IT Decision-Maker Workplace Technology Survey 2019).

AWS breaks down its capabilities into three domains: frameworks and infrastructure, machine learning services including Amazon SageMaker, and its suite of off-the-shelf models, developer APIs and services. It made more than a dozen announcements in machine learning spanning these important areas.

SageMaker Pushes into Machine Learning Life Cycle Management and Explainability

Amazon SageMaker, the company’s fully managed platform to build, train and deploy machine learning models and AI services, has become one of AWS’ most important products. At re:Invent, the company introduced six new capabilities, including SageMaker Autopilot, AWS’ answer to automated machine learning, giving greater visibility and control into the process. Additionally, SageMaker Notebooks allows developers to automate the process of sharing notebooks, and SageMaker Experiments helps them visualize and compare machine learning model iterations, trainings and outcomes in one place.

AWS also unveiled SageMaker Studio, an integrated development environment for machine learning. This includes the new SageMaker Debugger, a fully managed debugging service for real-time monitoring of models that warns and provides remediation advice when it finds problems, and SageMaker Model Monitor, a service that detects concept drift in models and alerts developers when the performance of a model running in production begins to deviate from the original trained model.

I spoke with one of AWS’ key machine learning customers, Vueling Airlines, owned by the International Airlines Group, about its experience using the tools. As part of the airline’s transformation and shift toward a data-driven culture, Vueling has hired more than 20 data scientists, many of whom are working with multiple teams and business areas. Already an extensive user of AWS services, Vueling believes the new SageMaker releases are crucial. Its head of data and analytics told me that the advancements will enable Vueling’s vision of having more machine learning models used by its business units and governed by its data centre of excellence.

As the market shifts from experimental projects to making machine learning operational in business processes, SageMaker is evolving quickly to meet this change, especially in the hot areas of machine learning life cycle management, explainability and governance (see also Taking AI from the Lab to Real Life). According to our survey of IT decision-makers, transparency into how systems work and are trained is now one of the most important requirements when investing in AI and machine learning, cited by 40% of respondents. The survey also found that 43% of respondents list tools that support AI operations and life cycle management as being the biggest current gap in the market for suppliers of AI platforms.

During his keynote presentation, Matt Wood, vice president of AI at AWS, also revealed some solid improvements in machine learning explainability in SageMaker based on the Shapley additive explanations technique, as seen in the photo below. AWS has clearly been working hard to advance this critical area.

AWS machine learning explainability

AWS Enters Search Market with Amazon Kendra

AWS also continued its expansion into business and vertical-market applications in what we call the fields of applied AI. They include: Amazon Rekognition Custom Labels, which lets organizations build custom image-recognition capabilities based on machine learning to identify objects or images specific to their business; Amazon Fraud Detector, which looks for online identity and payment fraud in customers’ system activity based on technology from Amazon’s consumer business; Amazon Transcribe Medical, which allows developers to add speech-to-text capabilities to medical applications so that doctors can dictate clinical notes into patients’ electronic health records; and CodeGuru, which automates processes for reviewing code using models pretrained by Amazon’s own code-review projects.

One of the most intriguing announcements of all was Amazon Kendra. This new enterprise search offering uses natural language processing to make searching for information easier through connectors to data stored in SharePoint Online, Java Database Connectivity and Amazon S3 repositories. Search is an area that customers frequently list as being broken in their organizations.

AWS doesn’t have a wide range of software-as-a-service applications that generate a body of information that its AI can improve for search, so it’s an interesting move and part of its strategy in helping businesses customize AI to specific industry and business challenges. It will be fascinating to watch AWS take on this big opportunity in a market currently lacking a clear winner.

Amazon’s Speed Sets It Apart

A year ago, I argued that customers need more from AWS in the fields of machine learning explainability and governance, and this year, the company ticked these important boxes. So how do the moves stack up in terms of advancing AWS’ differentiation in the increasingly crowded AI market?

I see several important areas where Amazon is now moving ahead.

  • Cloud. The AWS cloud is being rapidly engineered to support a range of machine learning workloads. The company claims that more machine learning happens on its cloud than anywhere else, supported by running 85% of TensorFlow and Apache MXNet workloads. Additionally, innovations in custom silicon such as AWS Inferentia, announced in 2018, and Inf1 Instances for Amazon EC2, announced earlier in 2019, are addressing infrastructure needs for machine learning inference, specifically where 90% of the cost resides.
  • Edge. Another important area is in hybrid and edge computing, where AWS has an early lead through products like Outposts, IoT Greengrass, AWS IoT and dedicated edge machine learning solutions such as SageMaker Neo and AWS DeepLens. I expect the company to extend its lead with the introduction of Local Zones and Wavelength, two of its major cloud announcements this year. As 5G and mobile edge computing take hold over the next five years, developers will increasingly look to the best cloud that supports low-latency applications and for development of hybrid and edge applications.
  • Robotics. AWS also stands out in robotics. RoboMaker, its platform to enable companies to build and deploy robots, such as unmanned ground vehicles, robotic arms and drones, didn’t receive any big announcements this year, but the area has an important opportunity beyond the 200,000 robots Amazon uses in its fulfilment centres. At the event, NASA’s Jet Propulsion Laboratory shared how it uses the platform along with reinforcement learning in SageMaker for an open-source, build-it-yourself Rover project it’s running. Combining AWS’ edge, AI and robotics capabilities will make a potent mix, particularly for industrial scenarios.
  • Speed. Most importantly, all this points to the fact that few companies (if any) are outpacing AWS in machine learning in 2019. SageMaker alone has received a staggering 150 updates since the beginning of 2018. This pace can be bewildering for customers, but AWS’ stated ambition of being “the broadest and deepest cloud platform” doesn’t look unrealistic when it comes to this area.

A good way to understand the pace of change is by comparing the two images below that I took from CEO Andy Jassy’s keynote sessions at the event in 2018 and 2019. The side-by-side comparison highlights the rapid expansion of AWS’ portfolio in just 12 months (see also re:Invent Shows Amazon’s Accelerating Efforts in AI).

Expansion of AWS portfolio from re:Invent 2018 to re:Invent 2019

What’s Next?

AWS is doing a good job at helping developers build, scale and apply machine learning in their organizations. But the company can’t rest on its laurels. I believe the following three important areas of expansion will come into focus in 2020.

  • Machine learning security. The Wall Street Journal reported in August that in 2018, a voice-based fake of a major UK company’s CEO emerged and tricked a senior employee into wiring more than $240,000 to a criminal bank account. It was the world’s first case of voice fraud and an astonishing reminder that needs for dedicated machine learning security will escalate in the future. I expect AWS will beef up its platform and algorithmic solutions in this area to help protect businesses, particularly in regulated industries, from security threats posed to machine learning, such as model inversions, trojans, spoofing and adversarial attacks.
  • More on explainability. AWS has made some good steps into explainability this year, but it’s still early days here. I expect the company to become more proactive in publishing and incorporating research in this area, and more prescriptive to customers about the uses where requirements are high and about the trade-offs between greater performance and explainability.
  • Business audiences. Finally, AWS needs to grow its mindshare with business audiences. Much of its storytelling focusses on developers and its vast array of tactical and technical features that make them more productive. However, AI is also a topic of concern for those in business and C-suite positions; one that influences strategy, company culture, operations and governance. AWS will need to raise its positioning and develop a constant cadence of communication to capture the attention of business decision-makers as well as technical developers. For this, the expansion of its Amazon ML Solutions Lab and similar offerings will also be crucial.

The re:Invent conference is a vital portal into Amazon’s future direction and investments. This year it revealed some of the most important moves we’ve seen in machine learning in 2019 and signalled the company’s growing differentiation. But Amazon will need to do more for customers in the realms of machine learning security, explainability and business strategy in the future if it’s to continue its leadership into 2020 and beyond.