Credit: Dreamstime
Myth: AI will make medical decisions and diagnoses. Reality: Yes, but AI won’t have the last word.
Today, radiologists are experts in the evaluation of X-rays, MRIs, CAT scans, and other medical imagery. One of the major efforts of AI is teaching image classifiers to recognize abnormalities like tumors. AI has the ability to scan millions of images to learn to interpret scans faster and more thoroughly than any human could ever achieve.
However, a doctor or radiologist will still have the final call in determining a diagnosis. It’s just that a diagnosis may come in minutes instead of days or weeks.
Myth: I have no idea what the AI is doing and if I can trust it. Reality: AI is much more transparent now.
Early on, AIOps was perceived as being a “black box,” i.e. a mysterious system that generated output without providing insights into what the underlying algorithm did and why. However, over time we are seeing these solutions mature, and more “white box” approaches that are gaining trust and adoption.
“While some systems don’t provide transparency, increasingly software vendors and AI systems are providing more visibility into why they did what they did,” said OpsRamp’s Byrne. “The tricky thing is to provide appropriate transparency, to not overwhelm the user, to gain their trust and understanding,” he said.
Myth: I need a data lake to train my AI. Reality: Drain the swamp.
Unstructured data is worse than structured data because it takes up space. To get rid of it you have to use up resources to sift through it all. For that reason, says QTS’s McCall, unstructured data for the sake of unstructured data can be worse than useless.
“What the world is working on now is how do I structure and organize data to mine it and how do I build historical algorithms and paradigms,” McCall said. “A little unstructured data is okay but when we open floodgates on data points, you absolutely have to have a data lake with the ability to organize and structure it later.”
Myth: Modeling determines outcome. Reality: You can’t be certain of that.
All AI initiatives begin as test projects. You may get excellent results during the testing phase, but find that your model is far less accurate when you deploy it into production. That’s because AI and machine learning models must be trained on data, and that training data must be representative of the real data, or results will suffer.
Note too that the training of your AI model is never complete. As soon as you put your model to use in the real world, its accuracy will begin to degrade. The speed of decline will depend on how fast the real world data changes (and customer preferences can change quickly), but sooner or later your model will have to be retrained with new data that represents the new state of the world.
“It’s a delicate task of defining your training data set. Your training data has to be the same as your production data,” said Marwaha. “That is the key to making your programs successful.” It’s a key you will need to turn to again and again throughout the lifetime of your AI model.
