Demonstrating excellent predictive power in the lab after days or weeks of data gathering and model tuning is exciting. However, we must temper that excitement with recognition that at that point we have just built a proof of concept. It demonstrates the potential locked in our data.
What comes next is the hard part – deploying a robust system that delivers sustained real-world value. To accomplish that we have to think hard about:
- Variability we will see in real-world
- Evolution of that reality over time
- Foreseeing failure
- Impact of correlation drift
- Dealing with data drift
- Distinguishing outliers vs. errors
- Getting the right answers for the wrong reasons
Robustness is all about reducing the risk of future failures, explainable systems make that easier, black boxes make that harder. How can you project and avoid system failure if you can’t articulate how the system works?