Improving LVCSR system combination using neural network language model cross adaptation

State-of-the-art large vocabulary continuous speech recognition (LVCSR) systems often combine outputs from multiple subsystems developed at different sites. Cross system adaptation can be used as an alternative to direct hypothesis level combination schemes such as ROVER. The standard approach involves only cross adapting acoustic models. To fully exploit the complimentary features among sub-systems, language model (LM) cross adaptation techniques can be used. Previous research on multi-level n-gram LM cross adaptation is extended to further include the cross adaptation of neural network LMs in this paper. Using this improved LM cross adaptation framework, significant error rate gains of 4.0%-7.1% relative were obtained over acoustic model only cross adaptation when combining a range of Chinese LVCSR sub-systems used in the 2010 and 2011 DARPA GALE evaluations. Copyright © 2011 ISCA.

Paraphrastic language models and combination with neural network language models

In natural languages multiple word sequences can represent the same underlying meaning. Only modelling the observed surface word sequence can result in poor context coverage, for example, when using n-gram language models (LM). To handle this issue, paraphrastic LMs were proposed in previous research and successfully applied to a US English conversational telephone speech transcription task. In order to exploit the complementary characteristics of paraphrastic LMs and neural network LMs (NNLM), the combination between the two is investigated in this paper. To investigate paraphrastic LMs' generalization ability to other languages, experiments are conducted on a Mandarin Chinese broadcast speech transcription task. Using a paraphrastic multi-level LM modelling both word and phrase sequences, significant error rate reductions of 0.9% absolute (9% relative) and 0.5% absolute (5% relative) were obtained over the baseline n-gram and NNLM systems respectively, after a combination with word and phrase level NNLMs. © 2013 IEEE.

Forecasting Daily Chlorophyll a Concentration during the Spring Phytoplankton Bloom Period in Xiangxi Bay of the Three-Gorges Reservoir by Means of a Recurrent Artificial Neural Network

A recurrent artificial neural network was used for 0-and 7-days-ahead forecasting of daily spring phytoplankton bloom dynamics in Xiangxi Bay of Three-Gorges Reservoir with meteorological, hydrological, and limnological parameters as input variables. Daily data from the depth of 0.5 m was used to train the model, and data from the depth of 2.0 m was used to validate the calibrated model. The trained model achieved reasonable accuracy in predicting the daily dynamics of chlorophyll a both in 0-and 7-days-ahead forecasting. In 0-day-ahead forecasting, the R-2 values of observed and predicted data were 0.85 for training and 0.89 for validating. In 7-days-ahead forecasting, the R-2 values of training and validating were 0.68 and 0.66, respectively. Sensitivity analysis indicated that most ecological relationships between chlorophyll a and input environmental variables in 0-and 7-days-ahead models were reasonable. In the 0-day model, Secchi depth, water temperature, and dissolved silicate were the most important factors influencing the daily dynamics of chlorophyll a. And in 7-days-ahead predicting model, chlorophyll a was sensitive to most environmental variables except water level, DO, and NH3N.

Using the moving window incorporated neural network to forecast the population behavior of Nostocales spp. in the River Darling, Australia

The paper demonstrates the nonstationarity of algal population behaviors by analyzing the historical populations of Nostocales spp. in the River Darling, Australia. Freshwater ecosystems are more likely to be nonstationary, instead of stationary. Nonstationarity implies that only the near past behaviors could forecast the near future for the system. However, nonstionarity was not considered seriously in previous research efforts for modeling and predicting algal population behaviors. Therefore the moving window technique was incorporated with radial basis function neural network (RBFNN) approach to deal with nonstationarity when modeling and forecasting the population behaviors of Nostocales spp. in the River Darling. The results showed that the RBFNN model could predict the timing and magnitude of algal blooms of Nostocales spp. with high accuracy. Moreover, a combined model based on individual RBFNN models was implemented, which showed superiority over the individual RBFNN models. Hence, the combined model was recommended for the modeling and forecasting of the phytoplankton populations, especially for the forecasting.

Modeling phytoplankton dynamics in the River Darling (Australia) using the radial basis function neural network

A radial basis function neural network was employed to model the abundance of cyanobacteria. The trained network could predict the populations of two bloom forming algal taxa with high accuracy, Nostocales spp. and Anabaena spp., in the River Darling, Australia. To elucidate the population dynamics for both Nostocales spp. and Anabaena spp., sensitivity analysis was performed with the following results. Total Kjeldahl nitrogen had a very strong influence on the abundance of the two algal taxa, electrical conductivity had a very strong negative relationship with the population of the two algal species, and flow was identified as one dominant factor influencing algal blooms after a scatter plot revealed that high flow could significantly reduce the algal biomass for both Nostocales spp. and Anabaena spp. Other variables such as turbidity, color, and pH were less important in determining the abundance and succession of the algal blooms.

Correcting the systematic error of the density functional theory calculation: the alternate combination approach of genetic algorithm and neural network

The alternate combinational approach of genetic algorithm and neural network (AGANN) has been presented to correct the systematic error of the density functional theory (DFT) calculation. It treats the DFT as a black box and models the error through external statistical information. As a demonstration, the AGANN method has been applied in the correction of the lattice energies from the DFT calculation for 72 metal halides and hydrides. Through the AGANN correction, the mean absolute value of the relative errors of the calculated lattice energies to the experimental values decreases from 4.93% to 1.20% in the testing set. For comparison, the neural network approach reduces the mean value to 2.56%. And for the common combinational approach of genetic algorithm and neural network, the value drops to 2.15%. The multiple linear regression method almost has no correction effect here.

Determining of the delay time for a heating ventilating and air-conditioning plant using two weighted neural network approach

This paper presents an two weighted neural network approach to determine the delay time for a heating, ventilating and air-conditioning (HVAC) plan to respond to control actions. The two weighted neural network is a fully connected four-layer network. An acceleration technique was used to improve the General Delta Rule for the learning process. Experimental data for heating and cooling modes were used with both the two weighted neural network and a traditional mathematical method to determine the delay time. The results show that two weighted neural networks can be used effectively determining the delay time for AVAC systems.

Iris recognition algorithm based on point covering of high-dimensional space and neural network

In this paper, we constructed a Iris recognition algorithm based on point covering of high-dimensional space and Multi-weighted neuron of point covering of high-dimensional space, and proposed a new method for iris recognition based on point covering theory of high-dimensional space. In this method, irises are trained as "cognition" one class by one class, and it doesn't influence the original recognition knowledge for samples of the new added class. The results of experiments show the rejection rate is 98.9%, the correct cognition rate and the error rate are 95.71% and 3.5% respectively. The experimental results demonstrate that the rejection rate of test samples excluded in the training samples class is very high. It proves the proposed method for iris recognition is effective.

Method for improving classification performance of neural network based on fuzzy input and network inversion

In order to effectively improve the classification performance of neural network, first architecture of fuzzy neural network with fuzzy input was proposed. Next a cost function of fuzzy outputs and non-fuzzy targets was defined. Then a learning algorithm from the cost function for adjusting weights was derived. And then the fuzzy neural network was inversed and fuzzified inversion algorithm was proposed. Finally, computer simulations on real-world pattern classification problems examine the effectives of the proposed approach. The experiment results show that the proposed approach has the merits of high learning efficiency, high classification accuracy and high generalization capability.

The design algorithm of an associative memory neural network

A design algorithm of an associative memory neural network is proposed. The benefit of this design algorithm is to make the designed associative memory model can implement the hoped situation. On the one hand, the designed model has realized the nonlinear association of infinite value pattern from n dimension space to m dimension space. The result has improved the ones of some old associative memory neural network. On the other hand, the memory samples are in the centers of the fault-tolerant. In average significance the radius of the memory sample fault-tolerant field is maximum.

The image coding scheme based on neural network and SVM

This paper applies data coding thought, which based on the virtual information source modeling put forward by the author, to propose the image coding (compression) scheme based on neural network and SVM. This scheme is composed by "the image coding (compression) scheme based oil SVM" embedded "the lossless data compression scheme based oil neural network". The experiments show that the scheme has high compression ratio under the slightly damages condition, partly solve the contradiction which 'high fidelity' and 'high compression ratio' cannot unify in image coding system.

Lossless data compression scheme based on neural network and SVM

First, the compression-awaited data are regarded Lis character strings which are produced by virtual information source mapping M. then the model of the virtual information source M is established by neural network and SVM. Last we construct a lossless data compression (coding) scheme based oil neural network and SVM with the model, an integer function and a SVM discriminant. The scheme differs from the old entropy coding (compressions) inwardly, and it can compress some data compressed by the old entropy coding.

SOC dynamic power management using artificial neural network

Dynamic Power Management (DPM) is a technique to reduce power consumption of electronic system by selectively shutting down idle components. In this article we try to introduce back propagation network and radial basis network into the research of the system-level power management policies. We proposed two PM policies-Back propagation Power Management (BPPM) and Radial Basis Function Power Management (RBFPM) which are based on Artificial Neural Networks (ANN). Our experiments show that the two power management policies greatly lowered the system-level power consumption and have higher performance than traditional Power Management(PM) techniques-BPPM is 1.09-competitive and RBFPM is 1.08-competitive vs. 1.79, 1.45, 1.18-competitive separately for traditional timeout PM, adaptive predictive PM and stochastic PM.

Structure approximate based on double weighted neural network

Double weighted neural network; is a kind of new general used neural network, which, compared with BP and RBF network, may approximate the training samples with a move complicated geometric figure and possesses a even greater approximation. capability. we study structure approximate based on double weighted neural network and prove its rationality.