How Amazon thinks about AI

Following on to yesterday's thoughts on Amazon's strategy, this is a succinct description of Amazon's work in Artificial Intelligence shared by Swami Sivasubramanian, vp of Amazon AI:

There are three layers. Top-layer applications like Lex, Polly and Rekognition are pretrained deep learning models, application programming interfaces catering to application developers who do not want to know anything about deep learning but want to build intelligent applications that can hear, speak or see.

The next layer is API platform services like Amazon Machine Learning and also various parts like EMR [Elastic MapReduce, for analyzing massive amounts of data], catering toward those who want to build their own machine learning models sitting on top of the data in Redshift [AWS’s data warehouse] or relational databases.

And the next layer my team does is around deep learning frameworks and machine learning algorithms. A bunch of scientists on my team are working on core deep learning frameworks. At AWS we are very open about supporting all deep learning frameworks like from Apache MXNet to TensorFlow to Caffe to Theano and more.

This makes sense, these three layers: at the top are special-purpose services for vision, speech, and speech recognition; in the middle are services that let you build your own models on your own data; and the base underpinning these are Amazon's contributions to the libraries and frameworks that make it all possible. 

Swami also shared two areas where machine learning will continue to improve is the ability to operate on edge nodes: such as cars, phones, and cheap computing devices; and models will become increasingly proficient with less data. Swami continues:

My daughter [Swami's daughter] is two years old, and around one or so, she recognized what a tomato is after seeing two tomatoes. She didn’t need a thousand tomatoes to be displayed. That’s exactly why I think deep learning is in its infancy. There are actually techniques that exist today where you can improve the accuracy of the deep learning model with very limited data.