Non-deterministic programming

Coding for machine learning. How does it differ from the code that I have always done, COBOL, Java or setting up rules to trap data outliers.

If I am coding a business rule in COBOL, that is deterministic. An if-loop which I had coded 20 years back will give me the same results today as it did give me then. A data outlier trap rule will give me the same suspense records over and over again if I subject that rule to the same input.

Data driven programs introduce non deterministic characteristics into our business logic. When I cause my logic to be derived from the ages and ages of data that I have collected over past, the program might behave differently if I give the same input over multiple iterations. It might classify a record as anomalous if the same was not a trend in the input data. However, if I repeatedly train the program over anomalous data which is streaming from a malfunctioning IOT sensor, it will over the time consider that as the new normal unless i do some rule setting.

It is this dynamism that is scary about the business logic that is driven by data driven decisioning a.k.a machine learning based logic. If I cannot look at the code and break it down into business rules just like how a business analyst can do, how am I supposed to frame a user manual for the users of the product. I am not sure if I can put a black box based machine learning structure into a user manual. It is here that explainable artificial intelligence will come to the rescue partially. Suck explanations might put at ease some of the purists who demand absolute behavioral characteristics of a system to be defined rather than acceptance of a dynamically behaving system.

We definitely need to think of explainability as an increasing number of organizations look to leverage the power of their historical data in their future business processes. An AI health scanner will be more trusted if it can tell me why it thinks I am susceptible to a sugar drop later in the day. I am sure this will be the new normal, data driven models along with an explanatory module, in the future.

#dynamic #machinelearning #explainability #XAI