The Perceptron Algorithm, Margins

Transcription

1 The Perceptron Algorithm, Margins MariaFlorina Balcan 08/29/2018

2 The Perceptron Algorithm Simple learning algorithm for supervised classification analyzed via geometric margins in the 50 s [Rosenblatt 57]. Originally introduced in the online learning scenario. Online Learning Model Perceptron Algorithm Its Guarantees under large margins

3 The Online Learning Model Example arrive sequentially. We need to make a prediction. Afterwards observe the outcome. For i=1, 2,, : Example x i Phase i: Online Algorithm Prediction h(x i ) Observe true label (x i ) Mistake bound model Goal: Minimize the number of mistakes. Analysis wise, make no distributional assumptions, consider a worst sequence of examples.

4 The Online Learning Model. Motivation classification (distribution of both spam and regular mail changes over time, but the target function stays fixed last year's spam still looks like spam). Recommendation systems. Recommending movies, etc. Predicting whether a user will be interested in a new news article or not. Add placement in a new market.

5 Linear Separators Feature space X = R d Hypothesis class of linear decision surfaces in R d. h x = w x w 0, if h x 0, then label x as, otherwise label it as X X X X X X X X X X O O O O O O O O w Trick: Without loss of generality w 0 = 0. Proof: Can simulate a nonzero threshold with a dummy input feature x 0 that is always set up to 1. x = x 1,, x d x = x 1,, x d, 1 w x w 0 0 iff w 1,, w d, w 0 x 0 where w = w 1,, w d

6 Linear Separators: Perceptron Algorithm Set t=1, start with the all zero vector w 1. Given example x, predict positive iff w t x 0 On a mistake, update as follows: Mistake on positive, then update w t1 w t x Mistake on negative, then update w t1 w t x Natural greedy procedure: If true label of x is 1 and w t incorrect on x we have w t x < 0, w t1 x w t x x x = w t x x 2, so more chance w t1 classifies x correctly. Similarly for mistakes on negative examples.

7 Linear Separators: Perceptron Algorithm Set t=1, start with the all zero vector w 1. Given example x, predict positive iff w t x 0 On a mistake, update as follows: Mistake on positive, then update w t1 w t x Mistake on negative, then update w t1 w t x Note: w t is weighted sum of incorrectly classified examples w t = a i1 x i1 a ik x ik w t x = a i1 x i1 x a ik x ik x Important when we talk about kernels.

8 Perceptron Algorithm: Example Example: 1,2 1,0 1,1 1,0 1, 2 1, 1 X X X Algorithm: Set t=1, start with allzeroes weight vector w 1. Given example x, predict positive iff w t x 0. On a mistake, update as follows: Mistake on positive, update w t1 w t x Mistake on negative, update w t1 w t x w 1 = (0,0) w 2 = w 1 1,2 = (1, 2) w 3 = w 2 1,1 = (2, 1) w 4 = w 3 1, 2 = (3,1)

9 Geometric Margin Definition: The margin of example x w.r.t. a linear sep. w is the distance from x to the plane w x = 0 Margin of positive example x 1 x 1 w Margin of negative example x 2 x 2

10 Geometric Margin Definition: The margin of example x w.r.t. a linear sep. w is the distance from x to the plane w x = 0 Definition: The margin w of a set of examples S wrt a linear separator w is the smallest margin over points x S. w w w

11 Geometric Margin Definition: The margin of example x w.r.t. a linear sep. w is the distance from x to the plane w x = 0 Definition: The margin w of a set of examples S wrt a linear separator w is the smallest margin over points x S. Definition: The margin of a set of examples S is the maximum w over all linear separators w. w

12 Perceptron: Mistake Bound Guarantee: If data has margin and all points inside a ball of radius R, then Perceptron makes R/ 2 mistakes. (Normalized margin: multiplying all points by 100, or dividing all points by 100, doesn t change the number of mistakes; algo is invariant to scaling.) R w*

13 Perceptron Algorithm: Analysis Guarantee: If data has margin and all points inside a ball of radius R, then Perceptron makes R/ 2 mistakes. Update rule: Mistake on positive: w t1 w t x Mistake on negative: w t1 w t x Proof: Idea: analyze w t w and w t, where w is the maxmargin sep, w = 1. Claim 1: w t1 w w t w. Claim 2: w 2 t1 w 2 t R 2. After M mistakes: w M1 w M (by Claim 1) w M1 R M (by Claim 2) (because l x x w ) (by Pythagorean Theorem) w t1 w t x w M1 w w M1 (since w is unit length) So, M R M, so M R 2.

14 Finding a Consistent Separator Can use Perceptron in a batch setting too, to find a consistent linear separator given a set S of labeled examples that is linearly separable by margin. We repeatedly feed the whole set S of labeled examples into the Perceptron algorithm up to R 2 1 rounds, until we get to a point where the current hypothesis is consistent/correct with the whole set S. Note we are guaranteed to reach such a point. The runnning time is then polynomial in R 2 and S.

15 What if there is no perfect separator? Perceptron can be adapted to the case where there is no perfect separator as long as the so called hinge loss (i.e., the total distance needed to move the points to classify them correctly large margin) is small. Let TD = total distance needed to move the points to classify them correctly large margin. After M mistakes: w M1 w M TD (new Claim 1) w M1 R M (by Claim 2) So, M TD R M, so M R 2 2 TD.

16 Margin Perceptron for Approx Maximizing the Margins Set t=1, w 1 = l x x, where x is the first example. Given example x, predict positive if w t x 2 predict negative if w t x 2 margin mistake if w t x (, ) 2 2 On a mistake, update: w t1 w t l(x)x Guarantee: If data has margin and all points inside a ball of radius R, then Margin Perceptron makes 8 R 2 R 4 mistakes.

17 Perceptron Discussion Simple, but very useful in applications like branch prediction; it also has interesting extensions to structured prediction. Can be kernelized to handle nonlinear decision boundaries! See future lecture!!!

18 Margin Important Theme in ML If large margin, # mistakes Peceptron makes is small (independent on the dim of the ambient space)! Large margin can help prevent overfitting. If large margin and if alg. produces a large margin classifier, then amount of data needed depends only on R/ [Bartlett & ShaweTaylor 99]. w Why not directly search for a large margin classifier? Support Vector Machines (SVMs).

19 Geometric Margin WLOG homogeneous linear separators [w 0 = 0]. Definition: The margin of example x w.r.t. a linear sep. w is the distance from x to the plane w x = 0. Margin of example x 1 If w = 1, margin of x w.r.t. w is x w. x 1 w Margin of example x 2 x 2

20 Geometric Margin Definition: The margin of example x w.r.t. a linear sep. w is the distance from x to the plane w x = 0. Definition: The margin w of a set of examples S wrt a linear separator w is the smallest margin over points x S. Definition: The margin of a set of examples S is the maximum w over all linear separators w. w

21 Support Vector Machines (SVMs) Directly optimize for the maximum margin separator: SVMs First, assume we know a lower bound on the margin Input:, S={(x 1, y 1 ),,(x m, y m )}; Find: some w where: w 2 = 1 For all i, y i w x i w Output: w, a separator of margin over S The case where the data is truly linearly separable by margin

22 Support Vector Machines (SVMs) Directly optimize for the maximum margin separator: SVMs E.g., search for the best possible Input: S={(x 1, y 1 ),,(x m, y m )}; Find: some w and maximum where: w 2 = 1 For all i, y i w x i w Output: maximum margin separator over S

23 Support Vector Machines (SVMs) Directly optimize for the maximum margin separator: SVMs Input: S={(x 1, y 1 ),,(x m, y m )}; Maximize under the constraint: w 2 = 1 For all i, y i w x i w

24 Support Vector Machines (SVMs) Directly optimize for the maximum margin separator: SVMs Input: S={(x 1, y 1 ),,(x m, y m )}; Maximize under the constraint: w 2 = 1 For all i, y i w x i This is a constrained optimization problem. objective function constraints Famous example of constrained optimization: linear programming, where objective fn is linear, constraints are linear (in)equalities

25 Support Vector Machines (SVMs) Directly optimize for the maximum margin separator: SVMs Input: S={(x 1, y 1 ),,(x m, y m )}; Maximize under the constraint: w 2 = 1 For all i, y i w x i w This constraint is nonlinear. In fact, it s even nonconvex w 1 w 2 2 w 1 w 2

26 Support Vector Machines (SVMs) Directly optimize for the maximum margin separator: SVMs Input: S={(x 1, y 1 ),,(x m, y m )}; Maximize under the constraint: w 2 = 1 For all i, y i w x i w = w/, then max is equiv. to minimizing w 2 (since w 2 = 1/ 2 ). So, dividing both sides by and writing in terms of w we get: w Input: S={(x 1, y 1 ),,(x m, y m )}; 2 Minimize w under the constraint: For all i, y i w x i 1 w x = 1 w w x = 1

27 Support Vector Machines (SVMs) Directly optimize for the maximum margin separator: SVMs Input: S={(x 1, y 1 ),,(x m, y m )}; argmin w w 2 s.t.: For all i, y i w x i 1 This is a constrained optimization problem. The objective is convex (quadratic) All constraints are linear Can solve efficiently (in poly time) using standard quadratic programing (QP) software

28 Support Vector Machines (SVMs) Question: what if data isn t perfectly linearly separable? Issue 1: now have two objectives maximize margin minimize # of misclassifications. Ans 1: Let s optimize their sum: minimize w 2 C(# misclassifications) where C is some tradeoff constant. w x = 1 Issue 2: This is computationally very hard (NPhard). [even if didn t care about margin and minimized # mistakes] w w x = 1

29 Support Vector Machines (SVMs) Question: what if data isn t perfectly linearly separable? Replace # mistakes with upper bound called hinge loss Input: S={(x 1, y 1 ),,(x m, y m )}; 2 Minimize w under the constraint: For all i, y i w x i 1 Input: S={(x 1, y 1 ),,(x m, y m )}; Find argmin w,ξ1,,ξ m w 2 C σ i ξ i s.t.: For all i, y i w x i 1 ξ i ξ i 0 ξ i are slack variables w x = 1 w w x = 1 w w x = 1 w x = 1

30 Support Vector Machines (SVMs) Question: what if data isn t perfectly linearly separable? Replace # mistakes with upper bound called hinge loss Input: S={(x 1, y 1 ),,(x m, y m )}; Find argmin w,ξ1,,ξ m w 2 C σ i ξ i s.t.: For all i, y i w x i 1 ξ i ξ i 0 ξ i are slack variables C controls the relative weighting between the twin goals of making the w 2 small (margin is large) and ensuring that most examples have functional margin 1. w x = 1 w w x = 1 l w, x, y = max(0,1 y w x)

31 What you should know Perceptron simple online algo for learning linear separators with good guarantees when data has large geometric margin. The importance of margins in machine learning. The Support Vector Machines (SVM) algorithm.