AWS Summit Singapore 2019: Day 2
An engineer’s review of the Machine Learning track
Tech conferences are always great fun, and AWS Summit Singapore 2019 was no exception. Held at the Singapore Expo Convention & Exhibition Centre on the 10th & 11th of April, the AWS team has put up a great developer focused conference for us devs in Singapore.
One of the fun things I got to check out this year was the DeepRacer track they had built into the event space. A queue of hopefuls were lined up, waiting for their chance to compete for a spot at the grand finals at Las Vegas, Nevada at re: Invent at the end of the year.
Some competitors got their racers through the track smoothly but others weren’t so lucky. I witnessed lots of hand-wringing and embarrassed laughter, but it was all in good fun. The top times are pretty impressive though, you can check them out here.
Amazon has stepped up their machine learning game again this year.
With the Machine Learning breakout track this year, AWS showed what a broad range of solutions they have on offer in the space. They’ve built services that span what a modern machine learning application pipeline consists of — from data ingestion/munging to model training and serving.
There was definitely a slight sales-y vibe to some of the presentations, but overall it was informative and relevant. Moments like this that make me think of Zach Kanter’s piece on “Why amazon is eating the world”. Really crushing the whole red+blue ocean (does anybody call it purple?) strategy here.
I was excited for Barnam Bora’s talk on “Building Business Outcomes with Machine Learning on AWS”, hoping for some insights to amazon’s business-driven approach to data science and machine learning. Unfortunately, it felt more like an introduction of the AWS platform’s capabilities to business people rather than a technical audience.
Intro slides like the one below inevitably cause me to raise my eyebrow(s).
In hindsight, Barnam’s background should have clued me in (along with a literal interpretation of the session’s title, d’oh). The saving grace was ViSenze CTO Li Guangda’s sharing of their service during the last ten minutes, which went into a bit more depth on the actual deployment and use case.
Guangda shared how industry use-cases require more than just a single detection or embedding model to be viable in production, which closely mirrors my own experience. Peak throughput of 5M images per hour is definitely impressive — and no doubt costs a pretty penny! EC2 instances are notoriously expensive, and their GPU offerings are no different. A quick web search will return you tons of results on price comparisons, but here’s a great rundown by Jeff Hale titled “maximize your gpu dollars” written quite recently.
Of course, the reality of using a separate provider if one is already entrenched in the AWS world is an increase in the complexity in your deployments. There might also be some pushback from infra/platform hardliners and advocates. You have to weigh these tradeoffs as every company’s situation is different.
ML in the Physical World with IoT
This was my favorite talk for the day. IoT tends to have a lot of inroads in the consumer space, and we can all relate to it on a personal level. Our office Alexa has become a part of the team with her reminders and bad jokes. Although I don’t have a chance to work with IoT on a daily basis, seeing the possibilities of edge deployments made me excited for the future of the whole space. It also shows just how much more room there is to grow for machine learning applications.
Tim Cruse gave a nice introduction to AWS IoT Greengrass, and briefly touched on fender’s wood matching architecture. He also (very) smartly included architecture diagrams in his presentation 💯 👏
Andreas Ruland from Smove talked about how IoT works with Smove’s vehicle fleet. After sharing lots of great ideas of how they wanted to use the data, he admitted they were slightly behind on implementation and ended his portion with a hiring pitch. Kudos for the honesty though.
Gary Han from Murata shared on predicting machine failures in an industrial manufacturing environment. I wasn’t familiar with the name Murata, but a quick search left me with eye-watering numbers — 75,326 employees, 1.2 Trillion Yen revenue and a close to 4T Yen market cap.
What they did was to build a wirelessly connected vibration sensor, which you can then magnetically attach onto industrial machinery. The data flows into AWS and all the software magic happens there.
Closing thoughts
A good effort and a well run event by the AWS Singapore team.
I personally wish some of the talks could have dived deeper into technical details, since they marketed it as a developer conference. But the reality is that businesses need to have their branding & marketing targets met, and sponsors need to have their space to do their PR and awareness stuff.
Not to be outdone, Google has made some announcements of their own for their end-to-end AI platform. Check out their beta here: https://cloud.google.com/ai-platform/
If you’re interested in working in Machine Learning, Computer Vision or NLP, OPEN8 is hiring engineers from all levels & experiences! Connect with me at michaely@open8.com
Edit: The TechFest on Day 1 did have a machine learning workshop, but it’s really an intro for beginners who want to get their feet wet in a safe space. Not for pseudo-grizzled (hah!) vets like me.
Machine Learning in Action, Building Predictive Models
Speaker: Pedro Paez, WWCS Solutions Architects, AWS
In this workshop, you will learn how to train a machine learning model using Amazon ML technologies. By the end of this workshop we will build an end-to-end pipeline that serves real-time predictions on streaming data. As data exponentially grows in organizations, there is an increasing need to use machine learning (ML) to gather insights from this data at scale and to use those insights to perform real-time predictions on incoming data.
