NIPS2009_0138_slide - ExperimentalResults...

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
Related Work ~ The Twenty-Third Annual Conference on Neural Information Processing Systems (NIPS2009) initialize for do Receive new instance Predict Receive label if then end if end for ALGORITHM: Perceptron A novel approach to online learning, which not only updates the weight of the newly added support vector, but also adjusts the weight of one existing support vector that seriously conflicts with the new support vector; Compared with a number of competing algorithms, the mistake bound can be significantly reduced by the proposed DUOL; Future work: DUOL for multi class online learning and budget online learning. Framework of Online Learning Initialize prediction function for do Receive instance Predict a label Receive the true label If then algorithm suffer a mistake if condition satisfied then Update prediction function to end if end for n t R x g f = 0 T t ,..., 2 , 1 = } 1 , 1 { )) ( ( ˆ 1 + = t t t x f sign y } 1 , 1 { + t y 1 t f t f t t y y ˆ c can be any reasonable function If , we can simply assign can be any reasonable condition, for example, g 0 )) ( ( 1 = t t x f sign 1 ˆ + = t y c t t y y ˆ Kernel function measures the similarity between the two instances x and y. R R R y x n n × : ) , ( κ Kernel Based Perceptron T t ,..., 2 , 1 = )) ( ( ˆ 1 t t t x f sign y = t t y y ˆ ) , ( ) ( ) ( 1 x x y x f x f t t t t α + = 1 , 0 0 = = weight f Limitation of Perceptron: the weights assigned to the misclassified examples(or support vectors) remain unchanged during the entire learning process. Similar with Perceptron, most online
Background image of page 1
This is the end of the preview. Sign up to access the rest of the document.

This note was uploaded on 02/12/2010 for the course COMPUTER S 10586 taught by Professor Jilinwang during the Fall '09 term at Zhejiang University.

Ask a homework question - tutors are online