Enterprise adoption of AI/ML has exploded in recent years, but most practitioners would agree that thus far results have not quite kept pace with early hopes - and with hype. Some now question whether the stage is set for another period of disillusionment, similar to those experienced in the past when expectations outran reality.While there might be some truth to these fears, there are important reasons to believe that a few relatively simple changes can fix a lot of the problems afflicting current AI/ML adoption.This talk aims to outline these simple practical fixes:
The need to differentiate between the lifecycle of a model and the lifecycle of software, for instance.
In addition, we cover four problems affecting AI solutions - overfitting, training-serving skew, concept drift, concerted adversaries - and discuss how to mitigate these problems.
Solution techniques include ensemble learning and model stacking to reduce overfitting, handling batch and streaming data correctly to eliminate training-serving skew, real-time monitoring of model performance to address concept drift, and re-establishing the link between machine learning and manual learning to deal with concerted adversaries.