- 1 current stop: learning
- 2 upcoming
- 3 upcoming
- 4 upcoming
- 5 upcoming
Finish after stop 5
Stop 1 of 5 · Bike shop: two ways to fix a wheel
Trick to keep in mind+3 XP
A formula is a lucky shortcut for the easy case; the gradient works for nearly everything else.
A straight line with squared error happens to have a neat closed-form solution. Add many features, bend the model or change the loss, and a formula is usually unavailable. Gradient descent only needs to compute the slope of the loss at the current point, which is almost always possible, so it scales to large models.
Read it once, then a question tests whether the trick stuck.