At AWS re:Invent, Amazon Web Services, Inc. (AWS), an Amazon.com company, announced five new artificial intelligence (AI) services designed to put machine learning in the hands of more application developers and end users – with no machine learning experience required. AWS introduced new services that use AI to allow more developers to apply machine learning to create better end user experiences, including new machine learning-powered enterprise search, code reviews and profiling, fraud detection, medical transcription, and human review of AI predictions.
Machine learning continues to grow at a rapid clip, and today there are tens of thousands of customers doing machine learning on AWS, including many customers that opt to use AWS’s fully managed AI Services.
In the past year, AWS has introduced several new fully managed AI Services like Amazon Personalize and Amazon Forecast that allow customers to benefit from the same personalization and forecasting machine learning technology used by Amazon’s consumer business to power its award-winning customer experiences. AWS customers are interested in learning from Amazon’s vast experience using machine learning at scale to improve operations and deliver better customer experiences, without needing to train, tune, and deploy their own custom machine learning models. Today, AWS is announcing five new AI services that build upon Amazon’s rich experience with machine learning, and allow organizations of all sizes across all industries to adopt machine learning in their enterprises – with no machine learning experience required. Here are the four that can potentially help in the hospitality industry:
Despite many attempts over many years, internal search remains a vexing problem for today’s enterprises, and most employees still frequently struggle to find the information they need. Organizations have vast amounts of unstructured text data, much of it incredibly useful if it can be discovered, stored in many formats and spread across different data sources (e.g. Sharepoint, Intranet, Amazon S3, and on-premises file storage systems). Even with common web-based search tools widely available, organizations still find internal search difficult because none of the available tools do a good job indexing across existing data silos, don’t provide natural language queries, and can’t deliver accurate results. When employees have questions, they are required to use keywords that may appear in multiple documents in different contexts, and these searches typically generate long lists of random links that employees then have to sift through to find the information they seek – if they find it at all.
Amazon Kendra aims to reinvent enterprise search by allowing employees to search across multiple silos of data using real questions (not just keywords) and deploys AI technology behind the scenes to deliver the precise answers they seek (instead of a random list of links). Employees can run their searches using natural language (keywords still work, but most users prefer natural language searches). As an example, an employee can ask a specific question like 'when does the IT help desk open?', and Amazon Kendra will give them a specific answer like ‘the IT help desk opens at 9:30 AM’, along with links back to the IT ticketing portal and other relevant sites. Customers can use Amazon Kendra across their applications, portals, and wikis. With a few clicks in the AWS Management Console, customers point Amazon Kendra at their various document repositories and the service aggregates petabytes of data to build a centralized index. Kendra helps to ensure that search results adhere to existing document access policies by scanning permissions on documents so that search results only contain documents for which the user has permission to access. Additionally, Amazon Kendra actively retrains its machine learning model on a customer specific basis to improve accuracy using click-through data, user location, and feedback to deliver better answers over time.
Just like Amazon, AWS customers write a lot of code. Software development is a well understood process. Developers write code, review it, compile the code and deploy the application, measure the performance of the application, and use that data to improve the code. Then, they rinse and repeat. And, yet, all of this process doesn’t matter if the code is incorrect to begin with, which is why teams perform code reviews to check the logic, syntax, and style before new code is added to an existing application code base. Even for a large organization like Amazon, it’s challenging to have enough experienced developers with enough free time to do code reviews given the amount of code that gets written every day. And even the most experienced reviewers miss problems before they get into customer-facing applications, resulting in bugs and performance issues.
Amazon CodeGuru is a new machine learning service that automates code reviews and finds an application’s most expensive lines of code. There are two components of Amazon CodeGuru – code reviews and application profiling. For code reviews, developers commit their code as usual (support for GitHub and CodeCommit exist today, with more repositories coming over time) and add Amazon CodeGuru as one of the code reviewers, with no other changes to the normal process or software to install. Amazon CodeGuru receives a pull request and automatically starts evaluating the code using pre-trained models that have been trained on several decades of code reviews at Amazon and the top ten thousand open-source projects on GitHub. Amazon CodeGuru will review the code changes for quality, and it if discovers an issue, it will add a human-readable comment to the pull request that identifies the line of code, specific issue, and recommended remediation, including example code and links to relevant documentation.
Amazon CodeGuru also contains a machine-learning powered application profiler that helps customers find their most expensive lines of code. To get started, customer install a small, low-profile agent in their application, so Amazon CodeGuru can observe the application run time and profile the application code every five minutes. This code profile includes details on the latency and CPU utilization, linking directly back to specific lines of code. Amazon CodeGuru can help operators find the most expensive line of code in an application, and it produces a flame graph that helps visually identify other lines of code that are creating performance bottlenecks. Amazon’s internal teams have used Amazon CodeGuru to profile code on more than 80,000 applications over the years. From 2017 to 2018, the extensive use of an internal version of Amazon CodeGuru helped the Amazon Prime Day team at Amazon’s consumer business increase its application efficiency, with a 325% increase in CPU utilization, a reduction in the number of instances needed to manage Prime Day, and 39% lower costs overall.
Amazon Fraud Detector
Tens of billions of dollars are lost to fraud every year by organizations around the world. Today, many AWS customers invest in large, expensive fraud management systems. These systems are often based on hand-coded rules that are time consuming, expensive to customize, and difficult to keep up-to-date as fraud patterns change – all of which results in systems that have lower than desired accuracy. This leads organizations to reject good customers as fraudsters, conduct more costly fraud reviews, and miss opportunities to drive down fraud rates. Amazon has been using sophisticated technology including machine learning to detect fraudulent transactions for more than 20 years, and understands it is a constant cat-and-mouse game with fraudulent actors that requires significant resources to build defenses and to keep evolving them. AWS customers have asked if AWS could share its expertise and experience.
Amazon Fraud Detector provides a fully managed service for detecting potential online identity and payment fraud in real time, based on the same technology used by Amazon’s consumer business – with no machine learning experience required. Amazon Fraud Detector uses historical data of both fraudulent and legitimate transactions to build, train, and deploy machine learning models that provide real-time, low-latency fraud risk predictions. To get started, customers upload transaction data to Amazon Simple Storage Service (S3) to customize the model’s training. Customers only need to provide the email address and IP address associated with a transaction, and can optionally add other data (e.g. billing address, or phone number). Based upon the type of fraud customers want to predict (new account or online payment fraud), Amazon Fraud Detector will pre-process the data, select an algorithm, and train a model – using the decades of experience running fraud detection risk analysis at scale at Amazon. Amazon Fraud Detector also uses machine learning-based data detectors that were trained on data from Amazon. These data detectors help identify patterns commonly associated with fraudulent activity at Amazon (e.g. anomalous email naming conventions) to help boost the accuracy of the trained model even if the number of fraudulent examples provided by a customer to Amazon Fraud Detector are low. Amazon Fraud Detector trains and deploys a model to a fully managed, private application programming interface (API) end point. Customers can send new activity (e.g. signups or new purchases) to the API and receive a fraud report, which includes a fraud risk score. Based on this report, the application can determine the right course of action (e.g. accept a purchase, or pass it to a human for review). With Amazon Fraud Detector, customers can detect fraud quicker, easier, and more accurately.
Vacasa is the largest full-service vacation rental management company in North America, with more than 23,000 vacation homes in 17 countries serving over 2 million guests per year. “Since our founding, we have used technology to enable our local teams to focus on caring for homes and guests, while maximizing revenue for vacation home owners,” says Eric Breon, Founder and CEO, Vacasa. “We’re excited about the introduction of Amazon Fraud Detector because it means we can more easily use advanced machine learning techniques to accurately detect fraudulent reservations. Protecting our ‘front door’ from potential harm enables us to focus on making the vacation rental experience seamless and worry-free.”
Amazon Augmented Artificial Intelligence (A2I)
Machine learning can provide highly accurate predictions for a variety of use cases, including identifying objects in images, extracting text from scanned documents, or transcribing and understanding spoken language. In each case, machine learning models provide a prediction and also a confidence score that expresses how certain the model is in its prediction. The higher the confidence number, the more the result can be trusted. For many use cases, when developers receive a high confidence result, they can trust that the results are likely to be accurate, and they can process them automatically (e.g. automatically moderating user-generated content on a social network, or adding subtitles to a video). However, in situations where confidence is lower than desired, due to some ambiguity in the prediction result, results may require a human review to resolve this ambiguity. This interplay between machine learning and human reviewer is critical to the success of machine learning systems, but human reviews are challenging and expensive to build and operate at scale, often involving multiple workflow steps, custom software to manage human review tasks and results, and recruiting and managing large groups of reviewers. As a result, developers sometimes spend more time managing the human review process than building their intended application, or they have to forego having a human review, which leads to less confidence and utility in many predictions.
Amazon Augmented Artificial Intelligence (A2I) is a new service that makes it easier to build and manage human reviews for machine learning applications. Amazon A2I provides pre-built human review workflows for common machine learning tasks (e.g. object detection in images, transcription of speech, and content moderation) that allow machine learning predictions from Amazon Rekognition and Amazon Textract to be human-reviewed more easily. Developers choose a confidence threshold for their specific application and all predictions with a confidence score below the threshold are automatically sent to human reviewers for validation. Developers can choose to have their reviews performed by Amazon Mechanical Turk’s 500,000 global workers, third-party organizations with pre-authorized workers (including Startek, iVision, CapeStart Inc., Cogito, and iMerit), or their own private reviewers. The results are stored in Amazon S3, and developers receive a notification when review is complete so they can take action based on the trusted results from human reviewers. Amazon A2I brings human review to all developers, removing the undifferentiated heavy lifting associated with building and managing custom reviewing pipelines or recruiting large numbers of human reviewers.