Amazon SageMaker Clarify aims to mitigate bias in machine learning
AWS looks to tackle machine learning bias with a new a new tool that works with Amazon SageMaker products to help developers and users better detect and eliminate AI bias.
AWS released Amazon SageMaker Clarify, a new tool for mitigating bias in machine learning models.
Revealed at AWS re:Invent 2020 in a keynote on Dec. 8 led by vice president of Amazon AI Swami Sivasubramanian, SageMaker Clarify works within SageMaker Studio to help developers prevent bias in their models and help consumers better understand how the models work.
The product comes amid intensifying debate in society about AI ethics and the role of bias in machine learning models and whether bias can cause racial and other discrimination in facial recognition and other AI systems.
Last week, AWS rival Google was at the center of the debate as a former top Google AI researcher said she had been fired for criticizing Google's treatment of women and people of color and not removing her name from a research paper discussing concerns that a language AI tool used by Google contains bias and could lead to hate speech.
Timely release
The launch of Amazon SageMaker Clarify also is timely in that it accompanies a recent AWS push in AI, said Ritu Jyoti, program vice president of AI Research at IDC.
Integrated with many SageMaker applications, SageMaker Clarify comes as AWS works to build out its AI portfolio and many AI creators work to eliminate biases in their models.
SageMaker Clarify will help data scientists detect bias in data sets before training and in their models after training, Jyoti said.
"It will help measure bias using a variety of statistical metrics. It will help explain how feature values contribute to the predicted outcome, both for the model overall and for individual predictions, serving the needs of different stakeholders," she said.
Using SageMaker Clarify, developers can specify important model attributes, such as location or occupation, and then run a set of algorithms on their data to detect bias.
After training a model, Clarify will automatically check the model for imbalances with the attributes and provide users with a detailed report showing the importance of each attribute to the model's output.
"This can help the consumer of your machine learning model better understand why a model is making a certain prediction," said Nashlie Sephus, applied science manager at AWS Machine Learning, during a keynote session at the conference, held virtually over three weeks this month.
AWS' Nashlie Sephus talks about Amazon SageMaker Clarify at AWS re:Invent 2020.
Clarify is packaged with SageMaker Model Monitor, a SageMaker Studio tool released last year that helps developers detect and fix concept drift.
This can help the consumer of your machine learning model better understand why a model is making a certain prediction.
Nashlie SephusApplied science manager, AWS Machine Learning
Clarify "will help detect bias drift and features importance drift over time with integration with Amazon SageMaker Model Monitor," Jyoti said.
More machine learning features
During the keynote, Sivasubramanian also unveiled several other new SageMaker capabilities, including Amazon SageMaker Edge Manager, a tool for managing machine learning models at the edge, and deep profiling for Amazon SageMaker Debugger, an application that monitors machine learning training performance.
Sivasubramanian also revealed the preview of Amazon Lookout for Metrics, a new service that uses machine learning to detect anomalies in metrics. Lookout for Metrics continuously monitors connected data to automatically detect anomalies, flagging users in real time to potential abnormalities in their data.
