AWS Helps ML Devs Streamline Human Reviews with Amazon A2I
A new managed service called Amazon Augmented AI, or Amazon A2I, is now available from Amazon Web Services (AWS) to help developers improve the accuracy of their machine learning models with the help of human oversight.
Amazon A21 provides developers with pre-configured workflows for instances when their machine learning predictions need an extra level of human verification. Certain machine learning applications — for example, those used to scan financial documents or perform facial recognition for law enforcement — require especially high levels of accuracy. To achieve that accuracy, developers are tasked with roping in other humans, from either a third-party auditor or in-house, to validate their machine learning predictions.
However, the task of building these so-called “human review systems” can be expensive to build and time-consuming to maintain.
“[H]uman reviews are challenging and expensive to build and operate at scale, often involving multiple workflow steps, operating custom software to manage human review tasks and results, and recruiting and managing large groups of reviewers,” AWS explained in its announcement of Amazon A2I last week. “As a result, developers sometimes spend more time managing the human review process than building the intended application, or they have to forego having human reviews, which leads to less confidence in deploying applications that utilize machine learning.”
Amazon A2I does away with this heavy lifting, instead automating the human review process for AWS solutions like Rekognition (image recognition) and Textract (character recognition), as well as the SageMaker machine learning platform. It has 60 prebuilt workflows in its arsenal for common tasks like speech transcription, image analysis and content moderation.
Developers can initiate a human review process for any machine learning prediction that falls below a certain confidence threshold. They can set the number of reviewers for each prediction and choose how to source those reviewers, whether they’re in-house, outsourced from specialized AWS partners or crowd-sourced via Amazon Mechanical Turk. Amazon A2I also lets developers post instructions to reviewers to ensure consistency, and to adjust those instructions as needed to address recurring reviewer errors.
More information on Amazon A2I is available here.
Gladys Rama is the senior site producer for Redmondmag.com, RCPmag.com and MCPmag.com.