A bakery installs a vision system on its production line that inspects an image of each pastry and returns a probability that it is defective. Engineers configure the reject arm to activate only when that probability exceeds 90 percent, so that pastries with minor visual noise, like a bit of extra flour dust, are not removed unnecessarily. What is being configured here?
Select an answer to reveal the explanation.
Short Explanation
Imagine the vision system as an inspector whispering a percentage next to every pastry: 'I'm 40 percent sure this one's bad,' or '95 percent sure.' The bakery isn't touching how that inspector learned the job, how the historical photos were split up for training, or which visual details the inspector pays attention to - all of that happened earlier, during development. What they're doing now is deciding how sure the inspector has to be before anyone actually acts on the opinion. Set the bar too low and you toss good pastries over a dusting of flour; set it too high and real defects slip through. That dial, the minimum confidence required before triggering an action, is exactly what's being adjusted here, and it's a knob you turn after the model is already built and running.
Full Explanation
The correct answer is C. The vision model outputs a probability, or confidence score, for each pastry, and the engineers are deciding how high that score must be before the system treats the prediction as actionable, in this case rejecting the pastry. That cutoff is the confidence score threshold, and tuning it up or down trades off false rejections against missed defects. Option A is incorrect because the learning rate controls how much the model's internal weights shift during training, a step that happens before the model is deployed, not a setting applied to its live predictions on the production line. Option B is incorrect because the training and test split ratio determines how historical data is divided to build and evaluate the model during development; it has no bearing on how a already-trained model's real-time output is acted upon. Option D is incorrect because feature engineering concerns choosing and transforming the raw inputs, such as pixel data or color channels, that the model consumes, not the decision rule applied to its output probability. Only the threshold setting matches what the engineers changed.