What is Amazon's anti-AI bias tool SageMaker Clarify?

Amazon SageMaker Clarify is a newly-launched service that allows AWS customers to detect bias in machine learning models, and raise transparency by explaining model behaviour to customers.

Concerns over bias in tech have heightened during 2020, with the Black Lives Matter protests forcing companies into pulling their AI facial recognition technologies. Organisations including Red Hat, Cisco and IBM have also come together to work on stripping out certain problematic terms from development, such as master/slave.

Meanwhile, although AI capabilities are becoming more and more sophisticated and widely adopted, significant concerns remain over the potential for biases to be embedded into algorithms.

In this context, Amazon has launched its SageMaker Clarify bias-detection tool that offers customers increased transparency when running machine learning models. It aims to address one of the biggest challenges with AI, in that it’s often difficult to understand why a model has come to any particular prediction.

What can SageMaker Clarify do?

Clarify serves as an additional set of capabilities for the broader SageMaker fully-managed machine learning service. These tools will be integrated with the web-based development environment SageMaker Studio, as well as other services such as SageMaker Data Wrangler, SageMaker Experiments, and SageMaker Model Monitor.

There are several capabilities that will stand out for customers, chief among them being that Amazon claims the service will help data scientists detect bias in datasets both before and after training models. Beyond this, users will be able to measure bias using a host of metrics, as well as to detect bias drift over time.

Detecting bias would be the first step, and the firm claims that SageMaker can simply examine your dataset and produce a set of bias metrics. This information can then be used to add your own bias reduction techniques to the data processing pipeline. Once the AI model is trained, however, Clarify can also run a bias analysis (including automatic deployment) and compute another set of bias metrics. These metrics, specifically, include the difference in positive proportions in labels, the difference in positive proportions in predicted labels, accuracy difference, and the counterfactuals-flip test.

One of the biggest features, however, is explaining AI model predictions, which SageMaker Clarify claims to do via support for a popular technique called SHapely Additive exPlanations (SHAP). SHAP analyses the individual contribution of feature values to the predicted output for each data instance and represents them as a positive or negative value. The example Amazon provides is of a credit scoring system, with an overall score influenced by variables in positive and negative ways, such as employment status or income level.

When will SageMaker Clarify be released?

SageMaker Clarify can be used today, having launched at AWS re:Invent 2020. Customers can start using the AI anti-bias tools in all regions where SageMaker is available at no additional cost to their regular subscriptions.