News

IBM's AI OpenScale Unlocks 'Black Box of AI'

• By John K. Waters
• 10/22/2018

Business leaders know they need to adopt AI and machine learning technologies to remain competitive, but adoption has been hindered by several challenges, including the well-known scarcity of in-house talent, but also management-level concerns over trust and compliance.

Those two issues topped the list of adoption-hindering worries cited by 5,000 C-level executives in a recent IBM survey. Big Blue is addressing those concerns, among others, with a newly announced platform designed to enable companies to manage their AI implementations through the full AI lifecycle, regardless of the environments in which they were built or on which they run.

Called AI OpenScale, the platform provides a layer of transparency between the applications and the machine learning models that deliver predictions, providing businesses with the ability to "unlock the black box of AI," explained Ritika Gunnar, IBM's VP of Watson and AI.

"When you look at AI today," she told Pure AI, "it's fairly isolated to the builders of the world -- the data scientists and the developers -- through API services. But we're finding that the businesses don't really trust how that AI is operating. That's why AI OpenScale provides a dashboard geared more toward the business and operations people, so they can really understand how their AI is performing. By breaking open that black box and providing visibility into how the AI is running at the point where predictions are actually being made, and within the context the applications infused with AI, we're making it possible for organizations to trust how their AI is working in mission critical workloads, which removes a key impediment to adoption."

AI OpenScale automates explainability, Gunnar said, mitigates bias, and provides auditability throughout the lifecycle of AI in a vendor-agnostic way.

"When we talk about trust and transparency, this is really about helping businesses understand how their AI applications are reaching decisions while those decisions are being made," she said.

AI OpenScale is also, as the name implies, an open platform, which means it allows organizations to operate and automate AI at scale, wherever it resides, Gunnar said, with transparent, explainable outcomes. It can work with open source machine learning or deep learning models, such as Tensorflow, Scikitlearn, Keras, and SparkML. It can also handle applications and models trained and hosted on common environments, including IBM Watson, IBM PowerAI, Seldon, AWS SageMaker, AzureML, and other non-IBM engines, she said.

"This is about harnessing a business's AI, regardless of where it was built, how it was built, or where it runs," she said.

With this new platform, IBM aims to help organizations:

• Understand how AI applications reach decisions by explaining how AI recommendations are made in everyday business terms.
• Address bias in AI applications automatically by continually monitoring AI applications and preventing bias through a unique, automated de-biasing technology.
• Ensure AI applications are auditable, by logging every prediction, every model version, and all the training data used, along with all metrics to help businesses comply with regulations such as GDPR.
• Use AI to build AI with IBM's Neural Network Synthesis Engine (NeuNetS), which addresses the short supply of human AI engineers by throwing AI at the problem. The NueNetS allows businesses to auto-generate deep neural networks rapidly from scratch.

"We think NueNetS is a game changer for how AI is used for innovation," Gunnar said.

IBM is expected to release the AI OpenScale platform later this year for IBM Cloud and IBM Cloud Private.