J. A. Bilmes. Dynamic bayesian multinets. In Proceedings of UAI, 2000.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....as follows. We feed the system using the observed data (the Xo [t] Then, the system determines the structure S (i.e. the dependencies) and the parameters which best represent the data. This strategy is known as structural learning in the BNs literature. A related work has been proposed in [5], where a structural learning algorithm using Bayesian Multinets is performed. In [5] each element of the observation vector is considered as a separate variable, then time cross dependencies among these variables are learned . The time window may include variables from the past and also from the ....

....determines the structure S (i.e. the dependencies) and the parameters which best represent the data. This strategy is known as structural learning in the BNs literature. A related work has been proposed in [5] where a structural learning algorithm using Bayesian Multinets is performed. In [5], each element of the observation vector is considered as a separate variable, then time cross dependencies among these variables are learned . The time window may include variables from the past and also from the future. The dependencies are chosen so as to reduce the entropy and to ensure that ....

[Article contains additional citation context not shown here]

Bilmes, J. A., "Dynamic Bayesian multinets", Proc. UAI2000.

.... of the conditional probabilities of the variables given their parents. Indeed, the JPD can be expressed in a factored way as P (X) Q n P (X i j i ) where i denotes the parents of X i in S. The use of DBNs in speech recognition has gained a lot of interest in the last few years [2, 3, 4]. In this paper, we use the exibility of this framework and instead of xing a priori the structure of the acoustic models (as is done with HMMs) we build an intelligent system which works as follows. We feed the system using the observed data. Then, the system determines the structure S ....

.... ] Xh [t 1] 3) The observation variable at time t is independent of all other variables given the hidden variables in the time window [t p ; t f ] for some positive integers p and f , P (Xo [t]jX 1 n fXo [t]g) P (Xo [t]jXh [t p ] Xh [t f ] 4) Xo [1] Xo [2] Xo [3] Xo [4] X h [1] X h [2] X h [3] X h [4] Figure 2: DBN structure with ( p ; f ) 2; 1; 1) T = 4 Hence, the search class of allowed DBN structures is de ned by the triples ( p ; f ) for, 1 max ; 0 p pmax ; 0 f fmax , where ( max ; pmax ; fmax ) is an upper ....

[Article contains additional citation context not shown here]

J. A. Bilmes, \Dynamic bayesian multinets," in UAI, 2000.

....especially when recognizing with the entire vocabulary. PhoneBook has typically been used for investigations either into tasks where the training and test vocabularies are different [11] or into new types of acoustic modeling and representation (e.g. modeling additional dependencies in [12] and [13], and articulatory state models in [14] We use the same breakdown of the database into training, development, and test sets as defined in [11] There is no overlap between speakers or words in the different sets. Two training sets are defined; the small training set contains about 20,000 ....

....with context dependent duration mod2 Training set # params 600 wd ER 8,000 wd ER 20k 627k 3.6 13.6 80k 1.55M 2.3 9.9 Table 1: Error rates (ER, in ) of the baseline recognizer on the PhoneBook test set. Reference Description, # params ER Dupont et al. 11] hybrid HMM ANN, 166k 5. 3 Bilmes [13] HMM or Dynamic Bayesian 5.6 Multinet, p 200k Richardson et al. 14] HMM Hidden 4.17 Articulator MM, 458k Table 2: Published test set error rates (ER, in ) on the 600 word PhoneBook task using the 20k training set. els. The pronunciation rules and diphone models, however, remain the ....

J. A. Bilmes, "Dynamic Bayesian Multinets," in Proc. 16RTS Conf. on Uncertainty in Artificial Intelligence, Stanford, California, 2000.

No context found.

J. A. Bilmes. Dynamic bayesian multinets. In Proceedings of UAI, 2000.

No context found.

J.A. Bilmes. Dynamic Bayesian Multinets. In Proceedings of the 16th conf. on Uncertainty in Artificial Intelligence. Morgan Kaufmann, 2000.

No context found.

J.A. Bilmes. Dynamic Bayesian Multinets. In Proceedings of the 16th conf. on Uncertainty in Artificial Intelligence. Morgan Kaufmann, 2000.

No context found.

J.A. Bilmes. Dynamic Bayesian Multinets. In Proceedings of the 16th conf. on Uncertainty in Artificial Intelligence. Morgan Kaufmann, 2000.

....pattern classification, and requires statistical models to discriminate between different speech utterances. Apart from discriminatively learned model parameters (such as means, variances, or transition matrices) graphical models are ideally suited for experimenting with discriminative structures [8, 19]. 2.2. Computation Probabilistic inference, such as evaluating (or computing the most likely value of) a conditional distribution, is the foundation behind all statistical computing. Graphical models have an associated set of algorithms which perform inference as efficiently as possible. Apart ....

....HMM states correspond to the same phoneme aa but in different contexts. The explicit approach is useful when modeling the detailed and intricate structures of ASR. It is our belief, moreover, that such an approach will yield improved results when combined with a discriminative structure [6, 8, 19], because it directly exposes events such as word endings and phone transitions for use as switching parents (see Section 3.4) The implicit approach is further useful in tempering computational and or memory requirements. In any case, GMTK supports both extremes and everything in between a ....

[Article contains additional citation context not shown here]

J.A. Bilmes. Dynamic Bayesian Multinets. In Proceedings of the 16th conf. on Uncertainty in Artificial Intelligence. Morgan Kaufmann, 2000.

....derived from nor different spectral sub bands of the same signal are conditionally independent given the class [2] On the other hand, producing low entropy distributions over HMM states from a product of sometimes incorrect classifiers might outweigh this inaccuracy. Alternatively, as argued in [4], an assumption that is incorrect for predictive accuracy does not ensure discriminative inaccuracy. In previous work [21] we additionally investigated other rules such as the max rule: P (cjx 1 ; xN ) maxnP (cjx n ) P K c=1 maxn P (cjx n ) 3) and the min rule: P (cjx 1 ; xN ) ....

J.A. Bilmes. Dynamic Bayesian Multinets. In Proceedings of the 16th conf. on Uncertainty in Artificial Intelligence. Morgan Kaufmann, 2000.

....it is sufficient to combine classifiers with low bias. When working with probabilistic decision making systems, it is usually advantageous to explicitly state the assumed underlying statistical model. For example, a hidden Markov model, easily defined by its conditional independent properties [3], is often used to represent speech for ASR. In a mixture model, it is assumed that a hidden and unknown cause selects each mixture component. A model may also be used to represent classifier combination. Explicating the models leading to a given combination rule could provide insight about when ....

....n A h . representations derived from nor different spectral sub bands of the same signal are CI given the class [1] On the other hand, producing low entropy distributions over HMM states from a product of sometimes incorrect classifiers might outweigh this inaccuracy. Alternatively, as argued in [3], an assumption that is incorrect for predictive accuracy does not ensure discriminative inaccuracy. 4. NEW DGM COMBINATION MODELS AND THEIR RULES It is possible to model classifier combination using distinct hidden random variables for each classifier output and the target. A specific form of ....

J.A. Bilmes. Dynamic Bayesian Multinets. In Proceedings of the 16th conf. on Uncertainty in Artificial Intelligence. Morgan Kaufmann, 2000.

No context found.

J. Bilmes. Dynamic bayesian multinets. In Proceedings of UAI, pages 38--45, 2000.

No context found.

J.A.Bilmes, "Dynamic Bayesian mu ltinets," in Proc. 16 Conf. on Unce tainty in Artificial Inteal"ete , 2000.

No context found.

J. A. Bilmes, "Dynamic Bayesian Multinets," in Proc. 16RTS Conf. on Uncertainty in Artificial Intelligence, Stanford, California, 2000.

No context found.

J.A. Bilmes. Dynamic Bayesian Multinets. In Proceedings of the 16th conf. on Uncertainty in Artificial Intelligence. Morgan Kaufmann, 2000.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC