The Real Reason Feature Scaling Matters

追加: 2026-05-28

[00:00:00]Wait a sec. Why does the left say unscaled and the right say scaled? The numbers are completely different, too.

[00:00:08]What's going on again?

[00:00:10]The left wants just the raw data.

[00:00:12]>> [music] >> The right one uses feature scaling, basically squeezing everything into a similar range, usually between 0 and 1.

[00:00:19]So, gradient descent learns way faster.

[00:00:22]The numbers look different, but the actual patterns in the data stay the same.

[00:00:27]If the scales are way off, the loss surface gets super stretched out and gradient descent starts zigzagging instead of heading straight to the minimum.

[00:00:35]And honestly, for visuals like this, scaling's pretty much mandatory.

[00:00:39]>> [music] >> Otherwise, the paraboloid gets so warped, it just looks like some weird skinny valley.

[00:00:46]Ah, so since the intercept is huge, but doesn't change much compared to the slope, you're putting them on the same playing field while keeping their original characteristics. Plus, it makes it way easier to visualize.

[00:00:59]Was this helpful? There's even more exciting content waiting for you in our full lesson. Click the link below now and let's take the next step in our growth together.