Machine Learning 4771

Transcription

1 Machine Learning 4771 Instructor: Tony Jebara

2 Topic 3 Additive Models and Linear Regression Sinusoids and Radial Basis Functions Classification Logistic Regression Gradient Descent

3 Polynomial Basis Functions To fit a P th order polynomial function to multivariate data: concatenate columns of all monomials up to power P E.g. dimensional data and nd order polynomial (quadratic) x 1 ( 1) x 1 ( ) x i 1 ( ) x i ( ) x N 1 ( ) x N ( ) 1 x 1 ( 1) x 1 ( ) x 1 1 ( )x 1 1 ( ) x 1 1 ( )x 1 ( ) x 1 ( )x 1 1 x i ( 1) x i ( ) x i ( 1)x i ( 1) x i ( 1)x i ( ) x i ( )x i ( ) ( ) 1 x N ( 1) x N ( ) x N ( 1)x N ( 1) x N ( 1)x N ( ) x N ( )x N ( )

4 Sinusoidal Basis Functions More generally, we don t just have to deal with polynomials, use any set of basis fn s: ( ) P = θ p φ ( p x) f x;θ + θ 0 p=1 These are generally called Additive Models Regression adds linear combinations of the basis fn s For example: Fourier (sinusoidal) basis φ k ( x ) i = sin kx i ( ) φ ( x ) = cos( kx ) k+1 i i Note, don t have to be a basis per se, usually subset θ 1 + θ + θ 3

5 Radial Basis Functions Can act as prototypes of the data itself f ( x;θ) = θ k exp 1 N x x k=1 σ k Parameter σ = standard deviation σ = covariance controls how wide bumps are what happens if too big/small? Also works in multi-dimensions Called RBF for short

6 Radial Basis Functions Each training point leads to a bump function f ( x;θ) = θ k exp 1 N x x k=1 σ k Reuse solution from linear regression: θ * = X T X Can view the data instead as X, a big matrix of size N x N X = exp 1 σ x 1 x 1 exp 1 σ x x 1 exp 1 σ x 3 x 1 exp 1 σ x 1 x exp 1 σ x x exp 1 σ x 3 x ( ) 1 X T y For RBFs, X is square and symmetric, so solution is just exp 1 σ x 1 x 3 exp 1 σ x x 3 exp 1 σ x 3 x 3 θ R = 0 X T Xθ = X T y Xθ = y θ * = X 1 y

7 Evaluating Our Learned Function We minimized empirical risk to get θ* How well does f(x;θ*) perform on future data? It should Generalize and have low True Risk: R true ( θ) = P ( x,y)l( y, f ( x;θ) )dx dy Can t compute true risk, instead use Testing Empirical Risk We randomly split data into training and testing portions {( x 1,y 1 ),,( x N,y )} ( x,y N N +1 N +1),, x N +M,y N +M { ( )} Find θ* with training data: Evaluate it with testing data: R train R test ( θ) = 1 N ( θ) = 1 M N i=1 N +M i=n +1 L( y i, f ( x i ;θ)) L( y i, f ( x i ;θ))

8 Crossvalidation Try fitting with different sigma radial basis function widths Select sigma which gives lowest R test (θ*) Loss R ( test θ * ) underfitting overfitting R ( train θ * ) 1/σ Best sigma Think of sigma as a measure of the simplicity of the model Thinner RBFs are more flexible and complex

9 Regularized Risk Minimization Empirical Risk Minimization gave overfitting & underfitting We want to add a penalty for using too many theta values This gives us the Regularized Risk R regularized ( θ) = R ( empirical θ) + Penalty ( θ) = 1 N N i=1 L( y i, f ( x i ;θ)) + λ N θ Solution for Regularized Risk with Least Squares Loss: 1 θ R regularized = 0 θ N θ * = ( X T X + λi ) 1 X T y y Xθ + λ N θ = 0

10 Regularized Risk Minimization Have D=16 features (or P=15 throughout) Try minimizing R regularized (θ) to get θ* with different λ Note that λ=0 give back Empirical Risk Minimization

11 Crossvalidation Try fitting with different lambda regularization levels Select lambda which gives lowest R test (θ*) Risk R ( test θ * ) underfitting overfitting R ( train θ * ) 1/λ Best lambda Lambda measures simplicity of the model Models with low lambda are more flexible

12 From Regression To Classification Classification is another important learning problem Regression Classification { ( )} x RD y R 1 X = ( x 1,y 1 ),( x,y ),, x N,y N { ( )} x RD y { 0,1} X = ( x 1,y 1 ),( x,y ),, x N,y N E.g. Given x = [tumor size, tumor density] Predict y in {benign,malignant} Should we solve this as a least squares regression problem? x x x x x x x x x x O O O O O O O O O

13 Classification vs. Regression a) Classification needs binary answers like {0,1} b) Least squares is an unfair measure of risk here e.g. Why penalize a correct but large positive y answer? e.g. Why penalize a correct but large negative y answer? Example: not good to use regression output for a decision f(x)>0.5 Class 1 f(x)<0.5 Class 0 if f(x)=-3.8 & correct class=0, squared error penalizes it f(x) o o x x x x x x x We pay a hefty squared error loss here even if we got the correct classification result. The thick solid line model makes two mistakes while the dashed model is perfect

14 Classification vs. Regression We will consider the following four steps to improve from naïve regression to get better classification learning: 1) Fix functions f(x) to give binary output (logistic neuron) ) Fix our definition of the Risk we will minimize so that we get good classification accuracy (logistic loss) and later on 3) Make an even better fix on f(x) to binarize (perceptron) 4) Make an even better risk (perceptron loss)

15 Logistic Neuron (McCullough-Pitts) To output binary, use squashing function g(). f ( x;θ) = θ T x ( ) = g θ T x ( ) = 1 + exp( z ) f x;θ g z ( ) ( ) 1 Linear neuron Logistic Neuron This squashing is called sigmoid or logistic function

16 Logistic Regression Given a classification problem with binary outputs X = ( x 1,y 1 ),( x,y ),,( x N,y ) N Use this function and output 1 if f(x)>0.5 and 0 otherwise ( ) = 1 + exp θ T x f x;θ { } x R D y 0,1 ( ( )) 1 { }

17 Short hand for Linear Functions What happened to adding the intercept? f x;θ ( ) = θ T x + θ 0 = θ(1) θ() θ(d) T x(1) x() x(d) + θ 0 = θ 0 θ(1) θ() θ(d) T 1 x(1) x() x(d) = θ T x

18 Logistic Regression Given a classification problem with binary outputs X = ( x 1,y 1 ),( x,y ),,( x N,y ) N Fix#1: use f(x) below, output 1 if f(x)>0.5 and 0 otherwise ( ) = 1 + exp θ T x f x;θ { } x R D y 0,1 ( ( )) 1 { }

19 Logistic Regression Given a classification problem with binary outputs X = ( x 1,y 1 ),( x,y ),,( x N,y ) N Fix#1: use f(x) below, output 1 if f(x)>0.5 and 0 otherwise ( ) = 1 + exp θ T x f x;θ ( ( )) 1 Fix#: instead of squared loss, use Logistic Loss L log ( y, f ( x;θ) ) = y i 1 ( )log( 1 f ( x;θ) ) y i log f x;θ ( ( )) This method is called Logistic Regression. But Empirical Risk Minimization has no closed-form sol n: R emp { } x R D y 0,1 ( θ) = 1 N N ( y i 1)log( 1 f ( x i ;θ)) y i log f x i ;θ i=1 { } ( ( ))

20 Logistic Regression With logistic squashing function, minimizing R(θ) is harder R emp ( θ) = 1 N N ( y i 1)log( 1 f ( x i ;θ)) y i log f x i ;θ i=1 1 y θ R = 1 i N 1 f ( x i ;θ) y N i f ( x i ;θ) f ' ( x ;θ ) = 0??? i=1 i Can t minimize risk and find best theta analytically! Let s try finding best theta numerically. Use the following to compute gradient ( ) = 1 + exp θ T x f x;θ ( ( )) 1 = g ( θ T x) Here, g() is the logistic squashing function ( ) = 1 + exp( z ) g z ( ) 1 g ' z ( ) ( ) = g ( z) 1 g ( z) ( ( ))

21 Gradient Descent Useful when we can t get minimum solution in closed form Gradient points in direction of fastest increase Take step in the opposite direction! Gradient Descent Algorithm choose scalar step size η, & tolerance ε initialize θ 0 = small random vector θ 1 = θ 0 η θ R emp θ 0, t = 1 while θ t θ t 1 { θ t+1 = θ t η θ R emp, t = t + 1 θ t } For appropriate η, this will converge to local minimum

22 Logistic Regression Logistic regression gives better classification performance Its empirical risk is N R emp ( y i 1)log( 1 f ( x i ;θ)) y i log( f ( x i ;θ)) This R(θ) is convex so gradient descent always converges to the same solution Make predictions using ( ) = 1 + exp θ T x f x;θ ( θ) = 1 N i=1 ( ( )) 1 Output 1 if f > 0.5 Output 0 otherwise