AWS Looks to ‘Demystify’ Machine Learning

(iQoncept/Shutterstock)

Amazon Web Services used a big data conference in the backyard of some of its largest government customers to showcase its AI and machine learning tools that are helping to funnel ever-larger volumes of data into its storage and computing infrastructure.

Making a pitch for better data management tools like metadata systems, AWS executives addressing a big data conference in Tysons Corner, Va., said the the public cloud giant aims to go beyond democratizing big data to “demystify” AI and machine learning.

The combination of organized data and analytics will accelerate the building and deployment of machine learning models, many that currently never make it to production. Those that are deployed often require up to 18 months to roll out, noted Ben Snively, a solution architect at AWS (NASDAQ: AMZN).

Open source tools for model development often advance a generation or two in the time it takes many enterprises to develop, train and launch a machine learning model, he added.

Snively asserted that the combination of big data and analytics along with AI and machine learning creates a “flywheel effect” in which organized, accessible data leads to faster insights, better products and—completing the cycle—more data.

(Hence, the cloud vendor forecasts as much as 180 zettabytes of widely varied and fast-moving data by 2025.)

As it seeks to demystify machine automation technologies and move beyond the current technology “hype phase,” AWS executives note that deployment of machine learning models and, eventually, full-blown platforms, remains hard. Among the reasons are “dirty” data that must be cleansed to foster access. The company estimates that 80 percent of data lakes currently lack metadata management systems that help determine data sources, formats and other attributes needed to wrangle big data.

That makes the heavy investments in data lakes “inefficient,” stressed Alan Halamachi, a senior manager for AWS solution architectures. “If data is not in a format where it can be widely consumed and accessible,” Halamachi stressed, machine learning developers will find themselves in “data jail.”

Once big data is wrangled and secured—“Hackers would like nothing more than to engineer a single breach with access to all of it,” Hamachi said—it can be combined with analytics on the inference side to accelerate training of machine learning models, Snively said.

Noting that most machine learning models built by enterprises never make it to production, the AWS engineers pitched several new tools including its SageMaker machine and deep learning stack introduced in November.  Described as a tool for taking the “muck” out of developing machine learning models, Snively said Sagemaker is also designed to free data scientists from IT chores like standing up a server for model development.

The cloud giant is seeing more experimentation among its customers as they seek to connect big data with machine learning development. “Voice [recognition] systems are here to stay,” Snively asserted, and developers are investigating “new ways of interacting with those systems.”

“It’s really about demystifying AI and machine learning” and getting beyond the “magic box” phase, he added.

Recent items:

AWS Takes the ‘Muck’ Out of ML with Sagemaker

How to Make Deep Learning Easy