Amazon has unveiled a boatload of new machine learning services for AWS. The new offerings include AI services for developers, models and algorithms for Amazon SageMaker, automatic data labeling and reinforcement learning services, and AWS-optimized versions of TensorFlow and other familiar machine learning libraries.
The new AWS-Optimized TensorFlow offering, now generally available, allows TensorFlow jobs to be automatically scaled and balanced across multiple GPU-powered EC2 nodes. Amazon claims that its improvements to TensorFlow scaling increase operational efficiency across dozens of GPUs, yielding faster model training than by setting up TensorFlow manually.
Another new AWS offering speeds up serving inferences, or predictions, from generated models at scale. Amazon Elastic Inference takes models from “all popular frameworks” (TensorFlow, Apache MXNet, etc.) and uses them to serve predictions, but does so from a relatively modest EC2 instance with GPU performance that can be dialed up or down as necessary. The customer pays only for GPU used. The idea is to limit GPU costs to only what’s needed, instead of over-provisioning the EC2 instance with a dedicated GPU that will go mostly unused.
Other new offerings are additions to Amazon SageMaker, an AWS managed service that handles machine learning workflows.
Many machine learning models need data that has been labeled, or preclassified. Unfortunately, labeling data is a job that’s often time-consuming, because it generally has to be done by hand. Amazon SageMaker Ground Truth learns the labels for a data set in real time as they’re applied by humans. Once trained on a subset of the data, it may be used to apply labels automatically.
Another machine learning challenge Amazon is addressing in SageMaker relates to reinforcement learning systems, where the model is continuously refined based on real-world feedback. Amazon SageMaker RL lets developers “build, train, and deploy with reinforcement learning through managed reinforcement learning algorithms,” bundled with many of the common ingredients needed for a reinforcement learning stack.
Yet another new Amazon SageMaker offering, SageMaker Neo, optimizes machine learning models to run faster and use less resources. It’s akin to how TensorFlow models are deployed on low-end hardware. Right now SageMaker Neo is restricted to deploying models to “Amazon EC2 instances, Amazon SageMaker endpoints, and devices managed by AWS IoT Greengrass.”
Prepackaged machine learning models for common business tasks—demand forecasting, data preparation, natural language processing—can now be purchased in AWS Marketplace and deployed to Amazon SageMaker.
Finally, Amazon unveiled a number of new AI services that allow developers to add intelligence to their applications. Amazon Textract uses machine learning to extract data from documents or forms. Amazon Comprehend Medical applies natural-language processing to medical documents. Amazon Personalize is a real-time personalization and recommendation service. And Amazon Forecast is a service that generates custom machine learning models from historical data to create time-series forecasts.
