Evaluating Insight, Not Just Computation

What I Actually Want to Measure When I give an exam, my primary concern is not whether a student can carry out a long computation without mistakes. In machine learning, for example, nobody in industry will be asked to manually differentiate a loss function or expand a matrix derivation on a whiteboard, except perhaps during an interview. What matters, years later, is something more durable: whether they recognize when a problem is about optimization, when it reduces to linear algebra, when regularization is implicitly shaping the hypothesis space, or when a probabilistic interpretation clarifies what a model is actually assuming. The students need to see how optimization, regularization, and probabilistic modeling relate to one another. In computer graphics, students must understand how linear transformations structure space. In hardware architecture, they must grasp how memory hierarchies and parallelism constraints shape performance. In each case, I am less interested in whether a…