Amazon Web Services (AWS) announced updates to their Deep Learning virtual machine image, as well as improvements to their AI services SageMaker Ground Truth and Rekognition.
Amazon Elastic Compute Cloud (EC2) is one of the oldest AWS services, but it still provides the most flexibility for constructing a machine learning solution. EC2 virtual machines are defined by an Amazon Machine Image (AMI), which is a "snapshot" of the machine's hard drive. This gives developers a way to quickly start up a computer, or a cluster of computers, with an operating system and suite of software already installed.
Announced in late 2016, the AWS Deep Learning AMI is Amazon's pre-packaged setup for "do-it-yourself" machine learning developers. It includes the Anaconda distribution of Python as well as several popular deep-learning frameworks, including TensorFlow, Apache MXNet, PyTorch, and Chainer. The newest version of the AMI includes an upgrade to PyTorch 1.1 and to Chainer 5.4. A full list of changes is available in the release notes.
Many machine learning training tasks require a dataset that includes known "answers" or labels. For example, training a system to recognize objects in an image requires a collection of images as well as the list of objects that are in them. This usually requires "natural intelligence"; that is, a person must examine each image and write down the objects in it.
Amazon SageMaker Ground Truth lets users upload unlabelled datasets and define labelling jobs for them, and then distribute the work to a team of human workers. SageMaker Ground Truth supports the use of a "public workforce" (via Mechanical Turk), 3rd party vendor workforce, or the user's own private workforce. The new release now allows users to configure the private workforce such that new incoming jobs send an email to the workers. Previously, the workers would have to poll the system for available jobs.
Google Cloud Platform has a similar labelling solution, but it is a "pre-release" feature, and only supports Google's own workforce and 3rd-party vendor workers; it does not support setting up a dedicated private workforce. The AWS private workforce solution, like most things AWS, is designed to work in conjunction with other AWS services. For example, the workers log in to the system using Amazon's Cognito identity service, and email notifications are sent via the Simple Notification Service (SNS).
For those who don't care to "roll their own" machine learning, AWS offers a variety of AI services. Amazon Rekognition is AWS's fully-managed video and image analysis solution. Given an image or video, Rekognition "can identify the objects, people, text, scenes, and activities, as well as detect any inappropriate content." However, like most of the newer services, it is not available in every AWS region. With this new release, Rekognition is now available in four more regions: EU (Frankfurt), EU (London), US West (N. California) and Asia Pacific (Singapore).
The other major cloud players have services similar to Rekognition. Microsoft Azure's Computer Vision service offers a comparable set of features. Like Rekognition, it is not available in every region. Google's Vision API is available globally, but only works images, not on full video.
