SageMaker Studio makes model building, monitoring easier
New SageMaker tools from AWS to open up machine learning models by making it easier for non-expert developers make, run, monitor and debug machine learning models.
LAS VEGAS -- AWS launched a host of new tools and capabilities for Amazon SageMaker, AWS' cloud platform for creating and deploying machine learning models; drawing the most notice was Amazon SageMaker Studio, a web-based integrated development platform.
In addition to SageMaker Studio, the IDE for platform for building, using and monitoring machine learning models, the other new AWS products aim to make it easier for non-expert developers to create models and to make them more explainable.
During a keynote presentation at the AWS re:Invent 2019 conference here Tuesday, AWS CEO Andy Jassy described five other new SageMaker tools: Experiments, Model Monitor, Autopilot, Notebooks and Debugger.
"SageMaker Studio along with SageMaker Experiments, SageMaker Model Monitor, SageMaker Autopilot and Sagemaker Debugger collectively add lots more lifecycle capabilities for the full ML [machine learning] lifecycle and to support teams," said Mike Gualtieri, an analyst at Forrester.
SageMaker Studio, Jassy claimed, is a "fully-integrated development environment for machine learning." The new platform pulls together all of SageMaker's capabilities, along with code, notebooks and data sets, into one environment. AWS intends the platform to simplify SageMaker, enabling users to create, deploy, monitor, debug and manage models in one environment.
Google and Microsoft have similar machine learning IDEs, Gualtieri noted, adding that Google plans for its IDE to be based on DataFusion, its cloud-native data integration service, and to be connected to other Google services.
SageMaker Notebooks aims to make it easier to create and manage open source Jupyter notebooks. With elastic compute, users can create one-click notebooks, Jassy said. The new tool also enables users to more easily adjust compute power for their notebooks and transfer the content of a notebook.
Meanwhile, SageMaker Experiments automatically captures input parameters, configuration and results of developers' machine learning models to make it simpler for developers to track different iterations of models, according to AWS. Experiments keeps all that information in one place and introduces a search function to comb through current and past model iterations.
"It is a much, much easier way to find, search for and collect your experiments when building a model," Jassy said.
As the name suggests, SageMaker Debugger enables users to debug and profile their models more effectively. The tool collects and monitors key metrics from popular frameworks, and provides real-time metrics about accuracy and performance, potentially giving developers deeper insights into their own models. It is designed to make models more explainable for non-data scientists.
SageMaker Model Monitor also tries to make models more explainable by helping developers detect and fix concept drift, which refers to the evolution of data and data relationships over time. Unless models are updated in near real time, concept drift can drastically skew the accuracy of their outputs. Model Monitor constantly scans the data and model outputs to detect concept drift, alerting developers when it detects it and helping them identify the cause.
Automating model building
With Amazon SageMaker Autopilot, developers can automatically build models without, according to Jassy, sacrificing explainability.
Autopilot is "AutoML with full control and visibility," he asserted. AutoML essentially is the process of automating machine learning modeling and development tools.
The new Autopilot module automatically selects the correct algorithm based on the available data and use case and then trains 50 unique models. Those models are then ranked by accuracy.
"AutoML is the future of ML development. I predict that within two years, 90 percent of all ML models will be created using AutoML by data scientists, developers and business analysts," Gualtieri said.
Mike GualtieriAnalyst, Forrester
"SageMaker Autopilot is a must-have for AWS, but it probably will help" other vendors also, including such AWS competitors as DataRobot because the AWS move further legitimizes the automated machine learning approach, he continued.
Other AWS rivals, including Google Cloud Platform, Microsoft Azure, IBM, SAS, RapidMiner, Aible and H2O.ai, also have automated machine learning capabilities, Gualtieri noted.
However, according to Nick McQuire, vice president at advisory firm CCS Insight, some of the new AWS capabilities are innovative.
"Studio is a great complement to the other products as the single pane of glass developers and data scientists need and its incorporation of the new features, especially Model Monitor and Debugger, are among the first in the market," he said.
"Although AWS may appear late to the game with Studio, what they are showing is pretty unique, especially the positioning of the IDE as similar to traditional software development with … Experiments, Debugger and Model Monitor being integrated into Studio," McQuire said. "These are big jumps in the SageMaker capability on what's out there in the market."
Google also recently released several new tools aimed at delivering explainable AI, plus a new product suite, Google Cloud Explainable AI.