2 / 8 Perceptron & SVM

The Widest Gap

For a linear classifier with decision boundary w⋅x+b=0w \cdot x + b = 0, how is the geometric margin of a correctly classified training point xix_i with label yi∈{−1,+1}y_i \in {-1, +1} defined?

A

∥w∥/yi(w⋅xi+b)|w| / y_i(w \cdot x_i + b)

B

1/∥w∥21 / |w|^2

C

yi(w⋅xi+b)/∥w∥y_i(w \cdot x_i + b) / |w|

D

yi(w⋅xi+b)y_i(w \cdot x_i + b)

E

(w⋅xi+b)2(w \cdot x_i + b)^2