AWS Unveils New AI and Machine Learning Offerings at re:Invent

At Amazon Web Services (AWS) annual AWS re:Invent conference, the company made several announcements around its growing portfolio in artificial intelligence. The announcements can be grouped into two major areas: applied artificial intelligence API’s and general purpose AI.

Amazon’s New Artificial Intelligence API’s

AWS announced three new API’s. These API’s are interesting in that they are accessible versions of the services that Amazon has developed internally for products like the Echo and Fire phones.

Lex is a natural language processing (NLP) API that can be used to build products like chatbots and voice command systems. With Lex, a developer can pass a string of text (such as a sentence sent to a chatbot) and the API will return a response that maps the text to one of the developer-defined “intents”. For anyone who has developed an Alexa Skill or used Alexa Voice Services, the Lex service will be very familiar. A few major differences exist, though. While an Alexa skill can only accept input from an Echo-based device, the Lex service can accept input from anything that can access the API. And while Alexa Voice Services could also accept speech from any input, the Lex service will now be able to accept speech or text without being  tied to an Alexa Skill.

Polly is a text to speech (TTS) API that helps build voice interfaces. With Polly, a developer can pass a string of text or markup, and the API will send back an audio snippet of the text. Again, users of Alexa Skills will be very familiar with this, since the console and API are similar to what is supported in a skill. The Polly API, though, offers a few advantages. The API does not require an Alexa skill, and is otherwise not tied to an Echo device in any way. In addition, the Polly service allows a developer to save and reuse the audio response without incurring an additional cost.

Rekognition is Amazon’s visual analysis API. With Rekognition, developers can pass an image, or set of images, for powerful visual search and recognition within applications The API can return the results from a variety of analyses. Rekognition looks like the evolved API version of the visual product search that was a key feature in Amazon’s short-lived Fire phone. The Rekognition API performs functions such as:

  • object detection (such as dogs, cats, landscapes)
  • facial feature detection (are the user’s eyes open; are they smiling or frowning)
  • facial comparison (how likely is it that the people in these two pictures are the same person?)
  • facial recognition (which faces does the person in this image most closely match?)

AWS’s new API’s bring them closer to parity with competitors like Microsoft, Google, and IBM, which have all had similar API features for months. AWS offers an ecosystem familiar to many developers with the compelling pricing that has made AWS a market leader. In addition, the Lex and Polly API’s leverage the services of a market leader in voice command platforms. Although AWS’s Alexa product is only a year and a half old, Amazon dwarfs Apple and Google in the number of third-party created voice commands, with over 6,000 skills now developed.

General AI Platforms

In addition to the new API’s that AWS has opened up, the company continues to bolster its general AI capabilities. At re:Invent, AWS announced several announcements related to general AI.

EC2 GPU Enhancements

Earlier this year, AWS announced support for a new instance type, P2, meant for computationally heavy tasks like deep learning. The P2 instances provide 1, 8, and 16 GPU configurations for the P2 instance type.

AWS also announced Elastic GPUs, which attaches a GPU to any EC2 instance. While a P2 instance is a better match for machine learning, the Elastic GPU might be a more cost-effective option for smaller workloads.


Shortly before re:Invent, AWS also announced official backing for MXNet, an open source deep learning library. MXNet exhibited a couple advantages for AWS. First, MXNet training time scaled better with increasing hardware. In benchmarks against an image analysis algorithm, MXNet scaled almost in line with number of GPUs. Second, MXNet has relatively compact models. This means that these models can potentially be deployed on mobile or other small scale devices.

While AWS promised increased investment in MXNet, AWS will continue to be library agnostic. In addition to MXNet, AWS’s provided Deep Learning AMI will continue to support CTNK, Theano, Torch, Caffe and TensorFlow along with Nvidia and other dependent libraries.

AWS has made a few interesting moves. The new P2 instances provide better performance and cost for machine learning tasks; with the potential to use spot instances, the cost goes down even further. In addition, AWS maintains an update-to-date, performant learning AMI with and an open mind towards libraries to be used (Google Cloud ML, in contrast, requires the user of TensorFlow).

Wrap Up

With the announcements made for re:Invent, AWS created new AI services that give customers no reason to look to other cloud providers. With their general AI updates, AWS has continued to improve the platform’s performance while lowering costs. For existing AWS users, the new services are a no-brainer to investigate. And even for those who are not currently invested in AWS infrastructure, AWS’s EC2 GPU platform, running on P2 spot instances, provides the most flexible and low-cost solution around.