Ethical AI – Addressing Embedded Bias

At MobiledgeX we work on several key innovations using our edge compute capabilities to support a wide range of artificial intelligence and machine learning applications. Among the most important are transportation and mobility as a service. This often involves some form of autonomous vehicle application, that contain algorithms and models to support collision detection and collision avoidance. These are key use cases required to proceed through the 5 levels of driving automation defined by NHTSA and other governing bodies.

However, as companies around the world strive to reach automation levels 3, 4 and 5, there are challenges. One that is discussed in some AI circles, but without reaching a resolution, is around bias. That would be inherent bias, placed into the automation AI systems. A recent Georgia Institute of Technology study (Predictive Inequity in Object Detection) found a 5% reduction in pedestrian (object) detection based on darker skin color, when adjusted for other factors. There are numerous articles, such as this one at Forbes, that discuss the situation of AI bias, anecdotally. Working these use cases at MobiledgeX, we need to, as others must, go beyond decrying the concerns and work diligently to address them. I’m proud that finding ethical, unbiased approaches to deploying these systems where machines, AI and human beings interact is supported and encouraged in my work.

As we move these autonomous vehicle and other AI technologies forward, we cannot and should not accept bias that impacts anyone disproportionately. The AI’s we produce won’t know the difference, it’s up to those of use constructing the systems to train them. No AI is born biased – bias is a learned behavior. Let’s train and nurture our AI systems well, so they are unbiased and ethical, as they make critical, life and death decisions for human beings. This has been, and will remain my goal.