LAS VEGAS – Doing a story on the machine learning announcements at the AWS re;Invent here is an extremely difficult task this year. One of the key themes of the event is that machine learning has reached the point – thanks to the cloud – where it is becoming ubiquitous, and will not peak and fall off, as it has done multiple times over the last three decades plus. So in that sense, almost everything at re:Invent this year is about machine learning. Some of the innovations announced can be more explicitly defined as such, however.
“The AI boom has come around before, every five to six years, but has never really stuck,” said Dr. Matt Wood, GM of Artificial Intelligence at AWS. “Scale has been the obstacle. Producing the modules, doing the training of them, and then putting these modules into production has been a challenge.”
“It’s different this time because of the scale that the cloud brings,” said Jeff Barr, AWS Chief Evangelist. “You aren’t limited to small data sets.”
Beyond this technical improvement in the ability to produce machine learning modules through the cloud, a remaining barrier to adoption has been its complexity, something AWS CEO Andy Jassy addressed in his Wednesday keynote.
“The hype and hope around machine learning has been tremendous,” Jassy said. “It is the loudest buzzword I have heard in eleven plus years at AWS. The issue is how do we turn it into something many more people can take advantage of. Builders want it to be much easier to engage with than it has been. Machine learning is still too complicated for everyday developers. There have been a lot of challenges and developers throw up their hands in frustration, because there has been too much heavy lifting.”
Jassy offered up Amazon SageMaker, a new fully managed service for developers and data scientists, as a response to the problem.
“Amazon SageMaker is an easy way to build, train, deploy and machine machine learning models for everyday developers,” he said. “We have taken top the 10 commonly-used algorithms, installed all the drivers and configured the framework for you. We have them optimized them, so eight of them run 10x faster than anywhere else – and the other two run 3x faster. You can also bring your own algorithms if you want.”
Once you have picked an algorithm, the training is much easier.
“In one click, it spins up an isolated cluster with its own software-defined network – and when it is done, it tears down the cluster,” Jassy said. “You don’t need a team to manage the training. You now have just two main choices to make when tuning: do I change the data being ingested; and how do I tune the parameters. That has been random and takes a lot of time and error. Now Hyper-Parameter Optimization [HPO] spins up multiple copies of models and uses machine learning to tune the parameters. It means that machine learning model builders don’t have to tune. It’s a huge weight off builders’ backs. And when you deploy to production, it’s one click. There is no other solution out there that is this easy. This is a big deal for everyday developers and scientists.
“This should make machine learning much more accessible to customers,” Jassy continued. “We are really excited about this.”
“SageMaker takes away most of the muck of machine learning,” Wood said, describing an example of how it was used to build a music recommendation service.
Another high-profile machine learning announcement was AWS DeepLens, a deep learning-enabled wireless video camera that runs real-time computer vision models, and was designed to give developers hands-on experience with machine learning.
“DeepLens is a high-definition camera optimized for machine learning,” Jassy said. “It integrates with SageMaker and Lambda, and has Greengrass in it. There are lots of tutorials and prebuilt models. You will be doing work within ten minutes from unboxing the camera.”
“We’ve shipped lots of devices before, like Alexa and Kindle Fire,” Barr said. “So this is similar to things we have already done. To me, DeepLens is a learning vehicle, where customers can put together pieces easily and work with them. I expect these to show up in schools before too long.”
While DeepLens is being sold on the Amazon side, a plan for it going to market through channel partners as well hasn’t been designed yet, but Jassy suggested that it is likely.
“It’s very early to see how that happens,” he said. “Our partners come to us with interesting proposals. We enter new fields and new markets with very little in the way of preconceived ideas.”
For the top level of machine learning, application services, Jassy announced four new offerings: Amazon Rekognition Video; Amazon Transcribe; Amazon Translate; and Amazon Comprehend.
Amazon Rekognition Video, which is available now, does real-time batch video analytics. It is designed to track people, objects, faces and inappropriate content in Amazon S3, using computer vision models that are trained to accurately detect thousands of objects and activities, and extract motion-based context.
“It can handle millions of videos,” Jassy said. “It allows you to do things as simple as automatically have a garage door open when it recognizes a license plate.”
Amazon Transcribe, which is available in preview, will not be good news for providers of transcription services. It converts speech from audio files stored in Amazon S3 into accurate, fully punctuated text.
“It is very hard to search audio well, so it’s usually converted to text,” Jassy said. “The problem is that it is expensive to transcribe manually, so typically only the parts considered to be the most important and transcribed.”
Amazon Transcribe has been trained to handle even low fidelity audio like contact centre recordings, with a high degree of accuracy. It can generate a time stamp for every word so that developers can precisely align the text with the source file. Out of the gate, Amazon Transcribe will support English and Spanish, but Jassy promised that many more languages will follow.
Amazon Translate, another service now available in preview, uses neural machine translation techniques to provide highly accurate translation of text from one language to another. It presently supports translation between English and six other languages (Arabic, French, German, Portuguese, Simplified Chinese, and Spanish), with many more to come in 2018.
“It is great for use cases requiring real time translation,” Jassy said. “Like all our other services, you will also find this to be very cost effective.”
“Translate has very low latency, so it can build effective near-real time language pairs,” Barr said. “It’s a neat black box.”
Finally, Amazon Comprehend, which is available now, is a fully managed natural language processing service that can understand natural language text from documents, social network posts, articles, or any other textual data stored in AWS. It uses deep learning techniques to identify text entities the language the text is written in, key phrases with concepts and adjectives, and whether a text is positive, negative or neutral.
“This is idea to assess data stored in a data lake, using natural language processing to give highly accurate information about what it contains,” Jassy said. He cited an example from Hotels.com, a property of Expedia, a new AWS client announced at the event.
“Before, their customer views and comments was just a big blob of data,” he said. “They use this to see unique characteristics that people like, and don’t like, about each hotel. This allows them to make better recommendations to users about other hotels they might like.”
In health care, Jassy noted that Comprehend will allow users to group all biased on diagnosis or symptoms.