IBM Seeks Transparency into AI Decision Making
Among the stumbling blocks to AI adoption is lack of transparency, understanding, and therefore trust in how machines actually make decisions. How and why AI arrived at an answer or recommendation is something that, as of today, machine learning models can’t provide, researchers say.
IBM (NYSE: IBM) announced this week it is addressing that and other gaps in AI development, gaps that contribute to bias in AI applications and undermine trust in critical decision making (such as in medical diagnostics), with the release of a technology platform called AI OpenScale. The platform also attempts to address the AI skills shortfall while making it easier to use the growing number of AI tools from different vendors.
IBM said Friday (Dec. 14) its AI platform advances the machine learning concept known as “explainability” so that developers and users can better understand how and why AI recommendations are made “in everyday business terms.” The inability to explain the reasoning behind recommendations has stoked worries about AI bias along with other ethical issues.
To that end, IBM said its open approach addresses bias in AI via an automated “de-biasing technology” that continuously monitors AI applications. The idea is to instill trust in AI applications by making them “transparent and easy to manage,” said David Kenny, senior vice president of IBM’s Cognitive Solutions unit.
The platform “can detect and address bias across the spectrum of AI applications, as those applications are being run,” IBM claimed.
Transparency also is tackled through an auditing feature that logs each prediction in every version of a machine learning model while tracking all the data used to train models. The audit trail also can be used to comply with data governance rules such as the EU’s General Data Protection Regulation.
As the name indicates, AI OpenScale is touted by IBM as a way to deploy AI applications in a “vendor-agnostic way.” That feature would enable users to manage AI applications regardless of source and where they are being run.
The platform also incorporates IBM’s Neural Network Synthesis Engine as a way to “use AI to build AI.” The engine is designed to automate the building of neural networks “from scratch.” IBM said its neural net builder will initially be available in beta form on its AI OpenScale platform.
In addition, the platform seeks to integrate the growing number of machine and deep learning models such as Keras, SparkML and TensorFlow. It supports AI applications and models running on IBM Watson along with non-IBM platforms such as Amazon Web Services’ (NASDAQ: AMZN) SageMaker and Microsoft’s (NASDAQ: MSFT) AzureML.
The AI platform is part of a larger IBM initiative to develop greater trust in AI recommendations while using automation to assist humans inundated by data. For example, a parallel research effort called Project Debater seeks nothing less than moving humans and machines away from “superficial thinking” by building machines that can “debate humans on complex topics.”