week 5

Seite 1

cost function for neural networks:

input layer : hidden layer : output unit

L = total number of layers in network

sl = no of units in layer l;

binary classification : 0 or 1; Sl = 1 ; K =1 ;

multi-class classification : K classes ; Sl = K ; Kl >=3

[1 0 0 0] ' [0 1 0 0]' [0 0 1 0]' [0 0 0 1]'

car pedestrian truck bus

J(theta)

Seite 2

back propagation algoritm is used in neural networks to minimize the cost function J(theta) in neural networks;

(equivalent of gradient descent in logistic/linear regression)

ie., find theta such that the cost function J is a minimum;