With businesses increasingly dependent on machine learning models for core operations, incorporating safeguards against bias is now essential to prevent inaccurate or deceptive outcomes. Today at AWS re:Invent, Amazon unveiled Amazon SageMaker Clarify, a new tool designed to mitigate bias within machine learning models.

“We’re announcing Amazon SageMaker Clarify, which provides visibility into your data and models throughout the entire machine learning process,” stated Bratin Saha, Amazon VP and general manager of machine learning, in an interview with TechCrunch.

He explained that the tool is intended to examine data for potential biases prior to the data preparation phase, enabling users to identify and address these issues before model development even begins.

“After I have my training dataset, I can [examine aspects such as] whether I have an equitable distribution of different classes – for instance, an equal representation of men and women, or a balanced number of other categories – and we provide a range of metrics for statistical analysis to gain valuable insights into dataset balance,” Saha clarified.

Following model construction, SageMaker Clarify can be utilized again to detect similar biases that may have emerged during the building process. “You begin by conducting statistical bias analysis on your data, and then, after training, you can perform a similar analysis on the model itself,” he added.

Various forms of bias can infiltrate a model due to the backgrounds of the data scientists involved, the inherent characteristics of the data, and the interpretations those data scientists apply to the data through their model design. This can be generally problematic, and can also result in algorithms perpetuating racial stereotypes. For instance, facial recognition systems have demonstrated high accuracy with white faces, but significantly lower accuracy when identifying individuals of color.

Identifying these biases through software can be challenging, as they often relate to team composition and other factors beyond the scope of a software analysis tool. However, Saha indicated that they are striving to make the software solution as comprehensive as possible.

“SageMaker Clarify offers data bias analysis, model bias analysis, model explainability, per-inference explainability, and global explainability,” Saha outlined.

Saha affirmed that Amazon recognizes the issue of bias and developed this tool to assist in addressing it, while acknowledging that the tool alone will not resolve all potential bias issues in machine learning models. They also provide additional resources to support users.

“We are also collaborating with our customers in several ways, offering documentation, best practices, and guidance on how to architect and interact with their systems to achieve the desired outcomes,” he said.

SageMaker Clarify is available now in multiple AWS regions.