Can you see what the computer sees?

In a recent study expert ophthalmologists were asked to determine the sex of a patient only looking at a picture of a retina. They were correct roughly 50% of the time, which is about as good as guessing. They then used an artificial intelligence (AI) algorithm to determine the sex and it was correct over 97% over the time. The doctors could not explain how the analysis was so accurate. It appears the algorithm detected something the doctors could not see. While this seems like a great case study for validating the use of AI for many applications (computers are better than humans), there can be a downside. What should be understood is whether your AI algorithm is giving you good results.

The pitfalls of AI

One of the classic examples of the danger of using machine learning and artificial intelligence is when they are used to predict future criminal behavior. The results have been questionable at best. This is due to the inherent bias introduced in the algorithm. Because the data contains parameters that highly correlate with criminal activity (race, ZIP code, income, etc.), the output produced results that “predicted” criminal activity. There were many other factors that should have been considered. Interestingly enough, the predictions were not very good when long term results were analyzed.

The point I am making is when using artificial intelligence you should always perform a reality check on the results. It may provide an answer that leads to a bad decision. Some drivers recently experienced this phenomenon while using a map application on their phones. There was an accident on a highway, so the application rerouted people onto a dirt road. Because of the recent rain, about 100 drivers ended up in a muddy field rather than arriving at their destinations. It would have been better to stay on a current route or consider an alternative route. In short, is the data giving you a good answer?

The benefits of AI

While there are pitfalls to artificial intelligence, there are many benefits to it. One of my favorite use cases is with a contract manufacturer that wanted to better predict finished product quality. With the help of a data scientist intern they analyzed a significant amount of process data. It should be noted they did not preclude any process variable from the data set. They soon discovered that the process variables people thought had the highest correlation to quality, was not all that significant. Moreover, the process variables that were thought to have little impact, appeared to be significant. They results were meet with skepticism at first (reality check), but after some additional analysis, they soon understood why this was the case.

Next steps

As you progress along your IIoT journey and are ready for artificial intelligence, be sure to keep the risks and benefits in mind when developing your algorithm and selecting data. Is their bias in the algorithm? Does this result make sense? Is this really a good decision? Answering these questions will help ensure you have resolved manufacturing inefficiencies, and not created an even larger problem.

 

Leave a Reply

%d bloggers like this: