# Je bilmes ee596awinter 2013dgms lecture 4 jan 23rd

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: M, other recursions It is possible to derive a temporal recursion for quantities other than α and β . E.g., here is a γt (j ) = p(Qt = j |x1:T ) backwards recursion. p(qt , qt+1 |x1:T ) = γ ( qt ) = qt+1 p(qt |qt+1 , x1:T )p(qt+1 |x1:T ) qt+1 p(qt |qt+1 , x1:T )γ (qt+1 ) = = qt+1 p(qt |qt+1 , x1:t )γ (qt+1 ) qt+1 qt+1 p(qt , qt+1 , x1:t ) γ (qt+1 ) qt p(qt , qt+1 , x1:t ) qt+1 p(qt+1 |qt )p(qt , x1:t ) γ (qt+1 ) qt p(qt+1 |qt )p(qt , x1:t ) = = Prof. Jeﬀ Bilmes EE596A/Winter 2013/DGMs – Lecture 4 - Jan 23rd, 2013 page 4-48 (of 239) HMMs HMMs as GMs Other HMM queries What HMMs can do MPE Summ HMM, other recursions It is possible to derive a temporal recursion for quantities other than α and β . E.g., here is a γt (j ) = p(Qt = j |x1:T ) backwards recursion. p(qt , qt+1 |x1:T ) = γ ( qt ) = qt+1 p(qt |qt+1 , x1:T )p(qt+1 |x1:T ) qt+1 p(qt |qt+1 , x1:T )γ (qt+1 ) = = qt+1 p(qt |qt+1 , x1:t )γ (qt+1 ) qt+1 qt+1 p(qt , qt+1 , x1:t ) γ (qt+1 ) qt p(qt , qt+1 , x1:t ) qt+1 p(qt+1 |qt )p(qt , x1:t ) γ (qt+1 ) qt p(qt+1 |qt )p(qt , x1:t ) qt+1 p(qt+1 |qt )αt (qt ) γ (qt+1 ) qt p(qt+1 |qt )αt (qt ) = = = Prof. Jeﬀ Bilmes EE596A/Winter 2013/DGMs – Lecture 4 - Jan 23rd, 2013 page 4-48 (of 239) HMMs HMMs as GMs Other HMM queries What HMMs can do MPE Summ HMM, other recursions It is possible to derive a temporal recursion for quantities other than α and β . E.g., here is a γt (j ) = p(Qt = j |x1:T ) backwards recursion. p(qt , qt+1 |x1:T ) = γ ( qt ) = qt+1 p(qt |qt+1 , x1:T )p(qt+1 |x1:T ) qt+1 p(qt |qt+1 , x1:T )γ (qt+1 ) = = qt+1 p(qt |qt+1 , x1:t )γ (qt+1 ) qt+1 qt+1 p(qt , qt+1 , x1:t ) γ (qt+1 ) qt p(qt , qt+1 , x1:t ) qt+1 p(qt+1 |qt )p(qt , x1:t ) γ (qt+1 ) qt p(qt+1 |qt )p(qt , x1:t ) qt+1 p(qt+1 |qt )αt (qt ) γ (qt+1 ) qt p(qt+1 |qt )αt (qt ) = = = Therefore, there is a backwards pass recursion using just the α’s without directly touching the observations again (better memory). Prof. Jeﬀ Bilmes EE596A/Winter 2013/DGMs – Lecture 4 - Jan 23rd, 2013 page 4-48 (of 239) HMMs HMMs as GMs Other HMM queries What HMMs can do MPE Summ Filtering Filtering query is p(qt |x1:t ) If we need this for one particular t, this is identical to one of the posteriors we already have p(qT |x1:T ) More likely, we need recursion from t to t + 1. This is obtained immediately from α-recursion, since αq (t) = p(x1:t , Qt = q ), we have p(qt |x1:t ) = αqt (t) q αqt (t) (4.33) t So, normalized αs are just the ﬁltering operation. Prof. Jeﬀ Bilmes EE596A/Winter 2013/DGMs – Lecture 4 - Jan 23rd, 2013 page 4-49 (of 239) HMMs HMMs as GMs Other HMM queries What HMMs can do MPE Summ Prediction p(qt |x1:s ), with t > s. This is also easily obtained from the αs for we have p(qt , x1:s ) = p(qt , qt−1 , . . . , qs+1 , qs , x1:s ) (4.34) qs ,qs+1 ,...,qt−1 p(qt |qt−1 )p(qt−1 |qt−2 ) . . . p(qs+1 |qs )αqs (s) = qs ,qs+1 ,...,qt−1 (4.35) p(qt |qt−1 ) = qt−1 p(qs+1 |qs )αqs (s) (qt−1 |qt−2 ) · · · qt−2 qs (4.36) Prof. Jeﬀ Bilmes EE596A/Win...
View Full Document

Ask a homework question - tutors are online