Since Amazon launched its AWS machine learning platform SageMaker four years ago, the company has banked on the idea that more and more businesses will want to develop ML-based systems without having to get their hands too dirty. In its quest to create tools that make ML easier to build and deploy — even for business users who aren’t engineers or data scientists — Amazon unveiled a slew of new features in its SageMaker suite during this week’s annual re:Invent show.
“We need to lower the bar for adopting of ML,” said Swami Sivasubramanian, vice president of Amazon AI, during a keynote talk Wednesday morning.
The new products and features intend to make that happen in a variety of ways such as through a visual interface that facilitates machine learning development without coding, as well as tools that simplify or speed up some of the in-the-weeds data work required to develop ML models such as labeling data and compiling training code.
The company said its no-code SageMaker Canvas tool, which uses a drag-and-drop interface to walk people through an ML workflow, is now widely available. The tool promises to help even data novices choose training data and pick the most appropriate ML models without writing any code. The system also provides reports and explanations of how the models work.
Another service aimed at putting machine learning in the hands of non-practitioners and students is SageMaker Studio Lab, a tool that enables free access to AWS compute resources so people can experiment with ML models and store projects to come back to later.
Amazon also aims to ease and speed up some of the tedious backend work required to prepare data used to train machine learning algorithms. But unlike its other automated offerings, its Ground Truth Plus service does require human intervention in a hybrid process. The company’s human “expert labelers” categorize information in datasets, then at some point as its ML systems learn from that labeled data, their automated processes take over to “pre-label” data like images. Several services have provided often low-paid data labeling labor for years.
In addition, Amazon pushed out new features that automatically compile training code to maximize compute and memory usage, help ML engineers choose the most appropriate instances to test models and enable unification of data processing and machine learning workloads for easier collaboration among data engineers and data scientists.