Perceptrons and Optimal HyperplanesExample: Majority-Vote Function • Definition: Majority-Vote Function fmajority – N binary attributes, i.e. x  {0,1}N – If more than N/2 attributes in x are true, then fmajority(x)=1, else fmajority(x)=-1. • How can we represent this function as a decision tree? – Huge and awkward tree! • Is there an “easier” representation of fmajority?Example: Spam Filtering • Instance Space X: – Feature vector of word occurrences => binary features – N features (N typically > 50000) • Target Concept c: – Spam (+1) / Ham (-1) • Type of function to learn: – Set of Spam words S, Set of Ham words H – Classify as Spam (+1), if more Spam words than Ham words in example. viagra learning the dating lottery spam?Example: Spam Filtering • Use weight vector w=(+1, -1, 0, +1, +1) – Compute sign(wx) • More generally, we can use real valued weights to express “spamminess” of word. • w=(+10,-1,-0.3,+1,+5) • Which vector is most likely to be spam with this weighting? A=x1, B=x2, C=x3 viagra learning the dating lottery spam?Linear Classification Rules • Hypotheses of the form – unbiased: – biased: – Parameter vector w, scalar b • Hypothesis space H – – • Notation – – –Geometry of Hyperplane Classifiers • Linear Classifiers divide instance space as hyperplane • One side positive, other side negativeHomogeneous Coordinates X = (x1, x2) W = (w1, w2, b) X = (x1, x2, 1) W = (w1, w2, w3)1 0(Batch) Perceptron Algorithm Training EpochExample: Perceptron Training Data: Updates to weight vector: 3•Init: w=0, =1 •(w0 x1) = 0  incorrect w1 = w0 +  y1 x1 = 0+ 1*1*(1,2) = (1,2)  hw1x1 = (w0+1*1*x1) * x1 = hw0(x1)+ 1 * 1 * (x1*x1) = 0 + 5 •(w1x2) = (1,2) (3,1) = 5  correct •(w1  x3) = (1,2) (-1,-1) = -3  correct •(w1  x4) = (1,2) (-1,1) = 1  incorrect •w2 = (1,2) +  y4 x4 = (1,2) - (-1,1) = (2,1)  hw2 x4 = (w1+1*-1*x4) * x4 = hw1(x4) + 1 * -1 * (x4 * x4) = -1Example: Reuters Text Classification “optimal hyperplane”Optimal Hyperplanes Assumption: Training examples are linearly separable.Hard-Margin Separation Goal: Find hyperplane with the largest distance to the closest training examples. Support Vectors: Examples with minimal distance (i.e. margin). Optimization Problem (Primal): d d dWhy min ½w·w? • Maximizing δ and constraining w is equivalent to constraining δ and minimizing w – We want maximum margin δ, • we don’t care about w • But because δ=wx, just requiring maximum δ will yield large w… – So we ask for maximum δ but constrain w • This is equivalent to constraining δ and minimizing wNon-Separable Training Data Limitations of hard-margin formulation – For some training data, there is no separating hyperplane. – Complete separation (i.e. zero training error) can lead to suboptimal prediction error.SlackSoft-Margin Separation Idea: Maximize margin and minimize training error. Soft-Margin OP (Primal): Hard-Margin OP (Primal): • Slack variable ξi measures by how much (xi,yi) fails to achieve margin δ • Σξi is upper bound on number of training errors • C is a parameter that controls trade-off between margin and training error.Soft-Margin OP (Primal): A B Which of these two classifiers was produced using a larger value of C?Controlling Soft-Margin Separation •Σξi is upper bound on number of training errors •C is a parameter that controls trade-off between margin and training error. Soft-Margin OP (Primal):Example Reuters “acq”: Varying CExample: Margin in High-Dimension x1 x2 x3 x4 x5 x6 x7 y Example 1 1 0 0 1 0 0 0 1 Example 2 1 0 0 0 1 0 0 1 Example 3 0 1 0 0 0 1 0 -1 Example 4 0 1 0 0 0 0 1 -1 w1 w2 w3 w4 w5 w6 w7 b Hyperplane 1 1 1 0 0 0 0 0 2 Hyperplane 2 0 0 0 1 1 -1 -1 0 Hyperplane 3 1 -1 1 0 0 0 0 0 Hyperplane 4 1 -1 0 0 0 0 0 0 Hyperplane 5 0.95 -0.95 0 0.05 0.05 -0.05 -0.05