IBM rolls out new platform to avoid prejudices in AI

IBM is trying to reduce the problem of prejudice in Artificial Intelligence (AI) based decision making tools. It does so with a new platform that can inform people about how AI models come to their conclusions, reports Silicon Angle.

According to the company, the platform, AI OpenScale, is needed because a significant proportion of the companies AI is not familiar enough when it comes to making the most important decisions. The biggest problem is that companies don’t know if their AI models have prejudices. In addition, companies want to better understand how AI comes to its conclusions and how it sets up its logic, according to IBM.

Prejudice is a by-product of the non-transparent nature of AI applications. The apps do not always show the logic behind a decision. Prejudice can arise through training data, but also through the logic of machine learning. Removing these prejudices can be a difficult exercise.

AI OpenScale

Jeff Welser, vice president and lab director at IBM Research’s Almaden Research Center, states that prejudices are not always easy to recognize. “If you have very large datasets, you may not even realize that the data have a small preference for one gender or whatever you are analyzing.”

The AI OpenScale platform is designed to solve the problem using algorithms that can detect and correct prejudices and preferences across the full spectrum of AI applications. It works by constantly monitoring the apps and applying an automated de-biasing technology to report prejudices. The software logs every forecast, model version and all training data to help companies understand how AI apps make decisions through explanations in normal English.

“So now we have systems that analyze datasets and see if they are overrepresented in specific sets of characteristics”, says Welser “It is possible that you have trained the model too much on those characteristics”.

Other functions

In addition, AI OpenScale helps make it easier to deploy AI models by running applications created with an open source machine learning or deep learning model on any common environment. These include IBM’s own Watson and PowerAI platforms, the Seldon open source framework, AWS’ SageMaker and Microsoft’s AzureML.

The platform also includes the Neural Network Synthesis Engine, which according to IBM enables companies to quickly and automatically build neural networks to run the AI.