Artificial intelligence is only as good as the people who create it. This is true not only for the algorithm, but the inspection data that helps train it.

Sewer inspection AI training requires inspection media to be thoroughly and accurately labeled. Developers use this labeled media to teach AI what to look for during automatic defect coding. But what happens when a pipe defect is too ambiguous for even experienced sewer pros to code with 100% accuracy? Do the professionals correcting the AI introduce subjectivity?

Subjectivity affects us when decisions are influenced by personal feelings, tastes, or opinions. The absence of subjectivity is often cited as a benefit of technology: Computers make their "decisions" based on pure facts.

In reality, however, developers accidentally introduce bias and subjectivity into many computer systems,  including while training AI. Even if the majority of the defects labeled in training media is accurate, the incorrect labels that sneak through the QA/QC process can taint the AI’s accuracy. 

AI developers work hard to identify and prevent these biases so that their systems see more accurate, reliable results. And in the case of Sewermatics, successfully countering bias requires tapping years of experience from a full team of wastewater professionals.

Pinpointing Corner Cases

When inspecting sewer systems, even the best operators will eventually encounter defects they just can’t make out. These unclear defects that occur outside of the average are called “corner cases''. In sewer inspection AI, corner cases are defects that cause debate among coders. They don’t generally appear in the majority of sewer data, so when they do, it’s because of qualities that set them outside of the average bell curve.

A corner case might be a large crack that borders on a fracture, or a dark spot that looks like a crack to one operator but a root to another. These discrepancies need to be resolved before the labeled inspection media is fed to the AI during the training process. If mistakes are repeatedly fed to the AI, machine learning tells us that those errors could spoil the accuracy of future results. However, we can correct these biases before they affect the AI engine by changing how we label our inspection media.

Overcoming Subjectivity and Bias in AI With Collective Knowledge

To overcome these biases, many AI developers are refocusing efforts from training methods that prioritize quantity of training media to bias-correcting practices that prioritize quality. By bringing together a full team of certified sewer professionals, as WinCan has done to support Sewermatics, AI teams are using their collective knowledge to prevent trainer bias from impacting data. This cuts off AI bias at the source, allowing the AI to more accurately process corner cases that show up in sewer inspections.