The need to stress test AI

Often AI is a high compute environment. At prediction time, the need for computational horsepower might be high, specially for unsupervised pipelines where large amounts of data might be processed at run time. We often spend a lot of our time testing AI for accuracy. And that is quite understandable. But consider this. If you are processing thousands or millions of records; and there are multiple users doing the same at multiple points of consumption, you might end up slowing down the system, or in worst case bringing it down. Hence the need to stress test your AI while simulating real life usage patterns, with added factor of safety. What do you do with the results of stress testing. These results have the potential of making you change your use case or algorithms used or the way your AI pipeline is designed or the schedule of usage or compromise between data churned and accuracy, and/or many other potential fixes. My point is not to miss this crucial step of stress testing your AI. Do this and you will be one step closer to practical AI. #abhayPracticalAI #ArtificialIntelligence #AI