Observations from AWS re:Invent 2017: AI, Voice, ML Frameworks, and Video Processing

Driving 16 hours to Las Vegas, the author attended this year’s AWS re:Invent and shares his impressions.

AI has been a major hotspot for the past two years, and AWS naturally follows with a large number of related sessions; limited time meant only a selection could be covered, and readers are invited to add any missed highlights.

Voice is the new Mobile

This year’s Re:Invent prominently introduced the AWS Alexa + Lex + Polly combination, together with the previously launched serverless Lambda, allowing developers to more easily provide voice‑based services.

Lex enables developers to build relatively complex and more human‑like dialogue systems; Polly’s speech‑synthesis quality is impressive, never failing in demos and sounding very natural; AWS Lambda further reduces the cost of experimenting with ideas.

For many small businesses or startups, using Alexa + Lex + Polly + Lambda + DynamoDB + S3 can create a voice‑first service fully managed by AWS, and the author predicts an explosive growth of such systems in the coming years.

A drawback is that Lex still has a long way to go; the author’s hands‑on experience showed mechanical conversation‑state management and basic keyword‑based semantic recognition, hoping AI‑driven improvements will address these issues.

MXNet vs Tensorflow

As a company with a large ecosystem, AWS lags behind Tensorflow in machine‑learning tools by about half a step, but is now trying to catch up.

AWS dedicated sessions introduced Gluon and Sockeye, aiming to attract more developers to the MXNet platform.

From the author’s personal perspective (potentially biased toward his former employer), catching up is difficult for MXNet because Tensorflow already dominates academic publications, yet AWS’s advantage lies in its massive enterprise customer base, so seamless integration of MXNet with user data could still present a significant opportunity.

Video is Hot

After images and voice have been partially conquered by neural networks, video processing emerges as the new hotspot, offering vast creative possibilities, albeit with higher storage and compute demands that AWS can satisfy.

At the Deep Learning Summit on Thursday, leading researchers presented the latest advances, including generating realistic video clips from 360° camera footage and automatically linking video objects to spoken words; although not yet mature, rapid progress and cloud resources like AWS suggest practical applications may appear within a few years.

Past reviews:

AWS re:Invent 2017 – Compute Track

AWS re:Invent 2017 – Storage and Database Track

AWS re:Invent 2017 – Chatbot Track