Created
August 25, 2012 19:30
-
-
Save oskarth/3469830 to your computer and use it in GitHub Desktop.
Intro to gradient descent, and why feature scaling leads to better convergence
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
ml gives computers ability to learn without being explicitly programmed. | |
# Linear Regression | |
linear regression is a simple model to find best for for data. | |
http://upload.wikimedia.org/wikipedia/commons/thumb/3/3a/Linear_regression.svg/400px-Linear_regression.svg.png | |
view above x and y axis as features, generalize to n-dimensional. | |
# Example | |
You have a data set with commute and sleep time, salary and happiness (1-10) | |
for the three features, commute, sleep and salary you want to predict happy | |
# Cost function and finding a good fit | |
How well does this line predict the data? concept of a cost function, once it converges we know we have a good fit. How find it? gradient descent. | |
# Contour plot | |
Here is a contour plot, like hills in a landscape: | |
http://upload.wikimedia.org/wikipedia/commons/f/fa/Cntr-map-1.jpg | |
# Gradient/Steepest descent | |
imagine we are walking up a hill by taking the steepest path. Thats gradient (ascent) descent, but in n-dimensions of course. | |
# Proportionality and feature scaling | |
Commute time on scale of 10m-2h, salary in 5-6 figure salary, obv not proportional. | |
A way of normalizing our values. Using x' = (x - min) / (max - min). | |
x=15 with min=0, max=120 gives x' = (15 - 0) / (120 - 0) = 15/120 = 0.125 | |
# Convergence speed | |
Feature scaling helps with convergence speed, without proportionality we get uneven descent. | |
good: http://upload.wikimedia.org/wikipedia/commons/thumb/7/79/Gradient_descent.png/350px-Gradient_descent.png | |
bad: http://komarix.org/ac/papers/thesis/thesis_html/img30.png | |
3d: http://upload.wikimedia.org/wikipedia/commons/6/68/Gradient_ascent_%28surface%29.png |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment