. Why does an overall accuracy of just over 50% on this lesson's 4-class CIFAR-10 task hide something important?
A single overall-accuracy number averages away per-class differences; here it hides that the truck class (starved to 90 training images) is rarely predicted correctly.
. In a confusion matrix where row i, column j counts examples of true class i predicted as class j, what does a perfect classifier's matrix look like?
A correct prediction means predicted class equals true class, i.e. j = i, which is exactly the diagonal.
. What does recall measure for a given class?
Recall = TP / (TP + FN): low recall means the model misses that class often, regardless of what it says about other classes.
. Why does the starved truck class end up with high precision but very low recall?
This is the standard signature of class imbalance: the model rarely guesses the rare class, but when it does, it's confident and usually correct.