Search

20 Feb 2018: 1 and a (uniform) ran-dom action otherwise. The obtained transition is then used tolearn the new estimate Q̂i1 (and σ̂i1). ... The parameters used in the two new explorationschemes were set as follows: α = 12, β = 1000 and β0 = 100.These were

20 Feb 2018: This allows asimple trade off between the emphasis placed on new informationcompared to old accumulated beliefs. ... This paper has introduced a new Bayesian Update of Dialogue Stateapproach to maintaining beliefs in dialogue systems.

13 Mar 2018: A New Look at Filtering Techniques for Illumination Invariance in AutomaticFace Recognition.

15 Jun 2018: The performance of the bi-RNNLMs is evaluatedon three speech recognition tasks: broadcast news; meetingtran-scription (AMI); and low-resource systems (Babel data). ... This projected, word vectorxt1 is then combined with thehistory vectorht2, which

20 Feb 2018: In previous experiments the user simulator would change its goaland require a venue matching the new goal but this is difficult to test ina real environment.

20 Feb 2018: The results show thatusing a pool of non-native adaptation data collected in advance,can substantially improve the performance on new non-nativespeakers. ... Speaker depen-dent adaptation might be expected to provide further improve-ment, however as

20 Feb 2018: New slots Old slots Joint. meanUnadapted 0.866 0.877 0.552. Adapted 0.869 0.890 0.566. ... Ac-curacy is reported on the Old slots and New slots, i.e.

20 Feb 2018: Inadifferentapproach to improving thesyntheticspeechquality,Toda and Tokuda [5] have introduced a new criterion on a higher-order moment called the global variance (GV), which is the vari-ance of the static ... 6. CONCLUSIONSThis paper has described a

20 Feb 2018: To train a policy using multi-objective GP-SARSA, a new weight configuration is sampledrandomly for each training dialogue.

13 Mar 2018: A new approach. An engineering viewpoint. This is part of a more ambitious project on understanding vision and making machines see.