To The Stars


Say Hello to Amazon AI

Say Hello to Amazon AI

In the original business plan, one of the stated goals for AWS, a nascent part of Amazon at the time, was to bring technology which had traditionally been a playground for only the largest, most well funded companies, within reach of everyone. Our goal today remains the same: to put futuristic technologies which are cumbersome, complex or capital intensive into the hands of every developer.

When it comes to artificial intelligence, we started down this path in 2015 with Amazon Machine Learning - a service which put predictive analytics in the hands of developers with tons of domain knowledge (and tons of data), without requiring deep machine learning expertise. This is the same technology we use internally at Amazon for everything from fraud to counterfeit good detection, now available in a friendly console for everyone.

At AWS Re:Invent this year, we brought a new wave of AI services to developers, in the form of Amazon AI; a collection of new artificial intelligence services which, along side Amazon ML, serve to put intelligence at the heart of every app and every business, through sophisticated, high-quality artificial intelligence that is easy to use and cost effective. 

It's super early in the advent of AI; as such, we don't intend to be paternalistic on how customers should be implementing intelligence. Once again, just as with other areas of analytics, building intelligent systems is about the right tool for the job.

With Amazon AI we're offering flexibility along two axes: algorithm tuning and model training.

Tuned Algorithms & Trained Models

For developers who have limited experience in training and implementing artificial intelligence (thousand-layer networks trained with artificial neural networks), or who don't have access to the very large datasets needed for training specific intelligence domains (such as the huge image libraries needed for image analysis or high quality speech corpora required for speech generation and recognition), we'll provide models which are pre-trained using cutting edge deep learning algorithms.

Amazon Lex, a natural language understanding and automatic speech recognition service, Amazon Polly, a text to speech service, and Amazon Recognition, an image analysis service, all provide access to fully trained models using cutting-edge deep learning algorithms. Because these models are trained by AWS and run in the cloud, they continually improve without having to tweak or customize complex parameters. 

Tuned Algorithms & Custom Models

Many developers already have large amounts of their own data in the cloud, and want to apply machine learning algorithms to that data: to train custom, predictive models based on their own data in their own domain. This is where Amazon Machine Learning fits in, with a set of tools to take data in S3 or Redshift, analyze it and make recommendations for data to use in a custom machine learning model. Then train and evaluate that model, and make it available for real time or batch processing of predictions. There is an example tutorial built right into the console: predicting responses to marketing campaigns

Custom Algorithms & Custom Models

Still more developers, usually researchers or data scientists, have access to both large volumes of data, and have the skills, interest in building novel, custom algorithms and models. These algorithms and models are usually constructed and trained using deep learning programming frameworks such as MXNet and Tensorflow, which I've written about before (at the release of our P2 instance family, and in choosing a framework and benchmarking performance).

In this case, researchers are looking for as much flexibility as possible, without sacrificing scalability to train sophisticated models with tons of data. We have a Deep Learning AMI, which is pre-installed with MXNet, Tensorflow, Torch, Caffe, Theano and the CUDA drivers, available as a one-click deployment from the AWS Marketplace, and as an elastic, scale out cluster through a CloudFormation template, which customers can run on high performance, high GPU memory P2 instances. We'll also make a container image available shortly.

Getting Started With Amazon AI

Polly, Rekognition and Machine Learning are all generally available today - and all have great console experiences. Even if you've never had a hand in building intelligent systems before, they are all very approachable. Lex is available in preview; you can sign up from the detail page

The Deep Learning AMI is available in the Marketplace, and more details of the CloudFormation Template are available on the AWS Compute blog.

Looking forward to seeing what happens to businesses and applications when artificial intelligence is available to all, not just the few.

The audio reading of these two short quotes was provided by Amazon Polly. :)
From Punch Cards To Pronunciation

From Punch Cards To Pronunciation

AI Research @ Scale

AI Research @ Scale