dq do k 1 q r k p1 qt q x for t

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

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: ward algorithm 1 2 3 4 5 6 forwards pass: for q = 1 . . . |DQ | do → − (k ) α1 (q ) ← R k (p(¯1 |Qt = q )) ; x + for t = 2 . . . T do for q = 1 . . . |DQ | do (k ) αt (q ) ← p(¯t |Qt = q ) x (k) 9 10 11 12 p(Qt = q |Qt−1 = r)αt−1 (r) ; (k ) αt (q ) ∈ arg⊕r p(Qt = q |Qt−1 = r)αt−1 (r) ; ˇ 7 8 (k) r (k) (k) αT ∈ arg⊕r αT (r) ; ˇ backwards pass: (k ) ∗ ∗ ∗ ∗ ∗ ∗ ˇ Identify (qT1 , kT1 ), (qT2 , kT2 ), . . . , (qTk , kTk ) = αT ; for t = T . . . 2 do ∗1 ∗1 ∗2 ∗2 ∗k ∗k Set (qt−1 , kt−1 ), (qt−1 , kt−1 ), . . . , (qt−1 , kt−1 ) ← (k ) (k) (k) ∗ ∗ ∗ ∗ ∗ ∗ αt (qt 1 )(kt 1 ), αt (qt 2 )(kt 2 ), . . . αt (qt k )(kt k ) ; ˇ ˇ ˇ Prof. Jeff Bilmes EE596A/Winter 2013/DGMs – Lecture 9 - Feb 6th, 2013 page 9-10 (of 180) k-best without the penalty Island Summary Scratch Computing Viterbi from max-margin - non-unique case Proposition 9.3.1 ∗ Let qt (1) be the state value of any Viterbi path of an HMM at time t. ∗...
View Full Document

This document was uploaded on 04/05/2014.

Ask a homework question - tutors are online