Data/ML/AI
Gradient Descent
Also written as Backpropagation, Loss Function, Optimizer
The method nearly all model training uses. Measure how wrong the model is, nudge its internal numbers in the direction that reduces the error, and repeat millions of times.
Think of it like
Walking downhill in thick fog by always stepping in the steepest downward direction you can feel.
Junior or senior?
Theory most candidates can recite, so recitation proves little.
Senior sounds like
Has watched training fail and can describe a run that would not converge and what they changed.
Ask them
“Tell me about a training run that wasn't converging. What did you change?”