Je bilmes ee596awinter 2013dgms lecture 9 feb 6th

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: (we used b = 2 in figure). This can be quite a significant reduction in memory costs, but added computation is nominal Prof. Jeff Bilmes EE596A/Winter 2013/DGMs – Lecture 9 - Feb 6th, 2013 page 9-39 (of 180) k-best without the penalty Island Summary Scratch Island Analysis Standard forward/backward (linear) inference is O(T ) memory and O(T ) compute (looking only at time for the moment, all costs are really multiplied by the average state cost, O(|DQ |2 ) in the case of time, and O(|DQ |) in the case of space) island case, pay O(b ∗ logb (T )) memory and O(T logb (T )) compute, where b, the base of the logarithm, is the number of subsegments a segment is divided into at each recursion level (we used b = 2 in figure). This can be quite a significant reduction in memory costs, but added computation is nominal (although log T can be large, say T = 1024, and b = 2, then we pay an extra time cost of a factor of 10,e.g., 1 day to 10 days)! Prof. Jeff Bilmes EE596A/Winter 2013/DGMs – Lecture 9 - Feb 6th, 2013 page 9-39 (of 180) k-best without the penalty Island Summary Scratch Analysis Detail We are given , the “linear section threshold” Prof. Jeff Bilmes EE596A/Winter 2013/DGMs – Lecture 9 - Feb 6th, 2013 page 9-40 (of 180) k-best without the p...
View Full Document

This document was uploaded on 04/05/2014.

Ask a homework question - tutors are online