CEO Andy Jassy announced a barrage of new machine learning capabilities for AWS SageMaker during his Re:Invent keynote on Tuesday.
SageMaker is Amazon’s big machine learning hub that aims to remove most of the heavy lifting for developers and let them use ML more expansively. Launched in 2017, there have been numerous features and capabilities introduced over the years, with more than 50 added to it in 2019 alone.
Of the SageMaker announcements made at the company’s annual conference in Las Vegas, the biggest was AWS SageMaker Studio, an IDE that allows developers and data scientists to build, code, develop, train and tune machine learning workflows all in a single interface. Within it information can be viewed, stored, collected and used to collaborate with others through the studio.
In addition to SageMaker Studio, the company announced a further five new capabilities: Notebooks, Experiment Management, Autopilot, Debugger and Model Monitor.
AWS SageMaker Studio interface
The first of these is described as a ‘one-click’ notebook with elastic compute.
“In the past, Notebooks is frequently where data scientists would work and it was associated with a single EC2 instance,” explained Larry Pizette, the global head of ML solutions Lab. “If a developer or data scientist wanted to switch capabilities, so they wanted more compute capacity, for instance, they had to shut that down and instantiate a whole new notebook.
“This can now be done dynamically, in just seconds, so they can get more compute or GPU capability for doing training or inference, so its a huge improvement over what was done before.”
All of the updates to SageMaker have a specific purpose to simplify the machine learning workflows, like Experiment Management, which enables developers to visualise and compare ML model iterations, training parameters, and outcomes.
Autopilot lets developers submit simple data in CSV files and have ML models automatically generated. SageMaker Debugger provides real-time monitoring for ML models to improve predictive accuracy, reduce training times.
And finally, Amazon SageMaker Model Monitor detects concept drift to discover when the performance of a model running in production begins to deviate from the original trained model.
“We recognised that models get used over time and there can be changes to the underlying assumptions that the models were built with – such as housing prices which inflate,” said Pizette. “If interest rates change it will affect the prediction of whether a person will by a home or not.”
“When the model is initially built to keep statistics, it will notice what we call ‘Concept Drift’ if that concept drift is happening, and the model gets out of sync with the current conditions, it will identify where that’s happening and provide the developer or data scientist with the information to help them retrain and retool that model.”