# 1 2 4 3 2 all but or prof je bilmes 1 all

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

This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: “(” “)” EE596A/Winter 2013/DGMs – Lecture 5 - Jan 25th, 2013 page 5-83 (of 232) HMMs Trellis Other HMM queries MPE Sampling What HMMs can do Summary Scratch Mealy vs. Moore HMMs Two ways to show the pronunciation model for the word “and” ε æ ε ε ε n ε ε d ε n:ε ε q0 æ:ε :ε q1 ε n:ε q2 d:ε q4 page 5-84 (of 232) n:ε :ε Left: Moore machine, Right: Mealy machine. Prof. Jeﬀ Bilmes EE596A/Winter 2013/DGMs – Lecture 5 - Jan 25th, 2013 HMMs Trellis Other HMM queries MPE Sampling What HMMs can do Summary Scratch Mealy vs. Moore HMMs Two ways to show the pronunciation model for the word “and” ε æ ε ε ε n ε ε d ε n:ε ε q0 æ:ε :ε q1 ε n:ε q2 d:ε q4 n:ε :ε Left: Moore machine, Right: Mealy machine. The output label on each Mealy arcs is shown as input:output and means “null” Prof. Jeﬀ Bilmes EE596A/Winter 2013/DGMs – Lecture 5 - Jan 25th, 2013 page 5-84 (of 232) HMMs Trellis Other HMM queries MPE Sampling What HMMs can do Summary Scratch Mealy vs. Moore HMMs Graphical model for Moore-HMM: as we have seen it (Qt is a parent of Xt for each t. y1 y2 y3 y4 y5 y1 y2 y3 y4 y5 x1 x2 x3 x4 x5 x1 x2 x3 x4 x5 Prof. Jeﬀ Bilmes EE596A/Winter 2013/DGMs – Lecture 5 - Jan 25th, 2013 page 5-85 (of 232) HMMs Trellis Other HMM queries MPE Sampling What HMMs can do Summary Scratch Mealy vs. Moore HMMs Graphical model for Moore-HMM: as we have seen it (Qt is a parent of Xt for each t. y1 y2 y3 y4 y5 y1 y2 y3 y4 y5 x1 x2 x3 x4 x5 x1 x2 x3 x4 x5 Graphical model for Mealy-HMM: Xt has two parents, either (Qt−1 , Qt ) or (Qt , Qt+1 ). y1 y2 y3 y4 y5 y1 y2 y3 y4 y5 x1 x2 x3 x4 x5 x1 x2 x3 x4 x5 Prof. Jeﬀ Bilmes EE596A/Winter 2013/DGMs – Lecture 5 - Jan 25th, 2013 page 5-85 (of 232) HMMs Trellis Other HMM queries MPE Sampling What HMMs can do Summary Scratch Mealy vs. Moore HMMs Mealy and Moore HMMs are equivalent in that they correspond to the same class of probability distributions. Prof. Jeﬀ Bilmes EE596A/Winter 2013/DGMs – Lecture 5 - Jan 25th, 2013 page 5-86 (of 232) HMMs Trellis Other HMM queries MPE Sampling What HMMs can do Summary Scratch Mealy vs. Moore HMMs Mealy and Moore HMMs are equivalent in that they correspond to the same class of probability distributions. Mealy → Moore conversion: cluster adjacent variables in the graphical model. Prof. Jeﬀ Bilmes EE596A/Winter 2013/DGMs – Lecture 5 - Jan 25th, 2013 page 5-86 (of 232) HMMs Trellis Other HMM queries MPE Sampling What HMMs can do Summary Scratch Mealy vs. Moore HMMs Mealy and Moore HMMs are equivalent in that they correspond to the same class of probability distributions. Mealy → Moore conversion: cluster adjacent variables in the graphical model. More → Mealy conversion: Use a separate arc for each possible output with observation randomness applied to the arc. Prof. Jeﬀ Bilmes EE596A/Winter 2013/DGMs – Lecture 5 - Jan 25th, 2013 page 5-86 (of 232) HMMs Trellis Other HMM queries MPE Sampling What HMMs can do Summary Scratch Mealy vs. Moore HMMs Mealy and Moore HMMs are equivalent in that they correspond to the same class of probability distributions. Mealy → Moore conversion: cluster adjacent variables in the graphical model. More → Mealy conversion: Use a separate arc for each possible output with observation randomness applied to the arc. It is sometimes more eﬃcient, in terms of number of states, to specify a language in terms of a Mealy HMM (see parenthesis matching example in the readings). Prof. Jeﬀ Bilmes EE596A/Winter 2013/DGMs – Lecture 5 - Jan 25th, 2013 page 5-86 (of 232) HMMs Trellis Other HMM queries MPE Sampling What HMMs can do Summary Scratch Sources for Today’s Lecture Probability and Random Processes, Grimmett/Stirzaker, 3rd Ed, 2001. “doc.pdf” sections 8.1 - 8.3 Prof. Jeﬀ Bilmes EE596A/Winter 2013/DGMs – Lecture 5 - Jan 25th, 2013 page 5-87 (of 232) HMMs Trellis Other HMM queries MPE Sampling What HMMs can do Summary Scratch Scratch Paper Prof. Jeﬀ Bilmes EE596A/Winter 2013/DGMs – Lecture 5 - Jan 25th, 2013 page 5-88 (of 232) HMMs Trellis Other HMM queries MPE Sampling What HMMs can do Summary Scratch Scratch Paper Prof. Jeﬀ Bilmes EE596A/Winter 2013/DGMs – Lecture 5 - Jan 25th, 2013 page 5-89 (of 232) HMMs Trellis Other HMM queries MPE Sampling What HMMs can do Summary Scratch Scratch Paper Prof. Jeﬀ Bilmes EE596A/Winter 2013/DGMs – Lecture 5 - Jan 25th, 2013 page 5-90 (of 232)...
View Full Document

## This document was uploaded on 04/05/2014.

Ask a homework question - tutors are online