What is gradient descent?
Whether you can explain the mechanism and its failure modes rather than recite the update rule.
Machine learningWarm-up
What a strong answer does
Gradient descent is an iterative method for choosing parameters that make a loss function as small as possible. The important reason it exists is that, for many models, you cannot just rearrange the equations and solve for the best parameters directly. In linear regression with squared error there is a closed-form solution, but in neural networks and many other models the loss is a complicated function of millions or billions of parameters.