Jiang Minghu

A fast hybrid algorithm of global optimization for feedforward neural networks

This paper presents the hybrid algorithm of global optimization of dynamic learning rate for multilayer feedforward neural networks (MLFNN). The effect of inexact line search on conjugacy was studied and a generalized conjugate gradient method based on this effect was proposed and shown to have global convergence for error backpropagation of MLFNN. The descent property and global convergence was given for the improved hybrid algorithm of the conjugate gradient algorithm, the results of the proposed algorithm show a considerable improvement over the Fletcher-Reeves algorithm and the conventional backpropagation (BP) algorithm, it overcomes the drawback of conventional BP and Polak-Ribieve conjugate gradient algorithm that maybe plunge into local minima.

Segmentation of Mandarin Braille word and Braille translation based on multi-knowledge

This paper is about the segmentation of Braille words and the transformation from Mandarin Braille to Chinese characters. Braille word segmentation consists of the rules base, the signs base of segmentation and knowledge base for disambiguation and mistakes. By using adjacency constraints and bidirectional maximum matching with a dictionary, our systems segmentation precision is better than 99% for the common text. By incorporating a pinyin knowledge dictionary into the system, we perfectly solved the problem of ambiguity in the translation from Braille to pinyin and developed a statistical language model based on the transformation of pinyin into characters. By using a multi-knowledge base to carry out the disambiguation process for each pinyin sentence, we built a multi-level graph and used a Viterbi search to find the sequence of Chinese characters with maximum likelihood, and used an N-best algorithm to get the N most likely character sequences. The experimental results show that the systems overall precision for translation from Braille codes to Chinese characters is 94.38%.

The consonant/vowel (C/V) speech classification using high-rank function neural network (HRFNN)

The article provide an improvement of the method of Bendiksen et. al. (1990) adopting the backpropagation (BP) network for voiced/unvoiced speech classification, by using the HRFNN, adapting it to the non-linear pronunciation model. The comparison test has shown that the HRFNN has a 100 times higher training rate than the BP network and the recognition accuracy is better than the BP network. As for the dynamic time-changing characterization of the speech signals and non-right-cross distribution of the C/V features, it was very difficult to search for the accurate CV transforming point in the past. A time-delay HRFNN is put forward, it is very effective for recognition of the CV transforming point, and for the automatic segmentation of continuous speech, it has a fast training rate, high recognition accuracy, and good dynamic characterization. The theory and experiments have shown that the network model is of high robustness.

RESEARCH ON THE MODEL OF SPEECH RECOGNITION AND UNDERSTANDING BY USING HIERARCHICAL INFORMATION FEEDBACK

In this paper according to the process of cognitive of human being to speech is put forward a model of speech recognition and understanding in a noisy environment. For speech recognition, two level modular Extended Associative Memory Neural Networks (EAMNN) are adopted. The learning speed is 9 times faster than that of the conventional BP net. It has high self-adaptability, robustness, fault toleration and associative memory ability to the noisy signals. To speech understanding, the structure of hierarchical analysis and examining faults which is a combination of statistic inference and syntactic rules is adopted, to pick up the candidates of the speech recognition and to predict the next word by the statistic inference base; and the syntactic rule base reduces effectively the recognition errors and candidates of acoustic level; then by comparing and rectifying errors through information feedback and guiding the succeeding speech process, the recognition of the sentence is realized.

Fast learning algorithms for time-delay neural networks phoneme recognition

To counter the disadvantage that TDNN take up long training time, the paper puts forward several improved methods of TDNN in phoneme recognition. The comparison of proposed methods with early method shows that they are effective in increasing the convergence speed of TDNN: (1) the error backpropagation algorithm trains initially the weights of the network; (2) the single-extreme output is replaced by the double-extreme output; (3) changing the energy function updates weights according to output errors; (4) the weight update criterion of error backpropagation is changed from the average weights of all corresponding time-delay frames to the layers. All of these make the training time decrease from 23 hours and 25 minutes to 45 minutes. The convergence speed increases by tens of times when the complexity of the network increases just a little more.

The dynamic-semantic-framework of a literature QAS

In the paper we propose the dynamic-semantic-framework (DSF), which is a strategy, used in the analysis of question answer system (QAS) of Chinese question sentences. With the DSF, the system omits the traditional analysis on the syntax level, which is always quite complex in Chinese oral, and directly extracts keywords from the sentences dynamically. The DSF is translated to query mode in order to get information from the knowledge base, which contains the information of the Chinese authors and their works. The answering process outputs the answer abstracted from the knowledge base. The experimental results show that the DSF can be readjusted if the query in the knowledge base fails, and with this strategy, the system can avoid the query failure caused by the incorrectly analysis of the keywords.

Reliability estimations of control systems effected by several interference sources

In order to establish the sufficient and necessary condition that arbitrarily reliable systems can not be constructed with function elements under interference sources, it is very important to expand set of interference sources with the above property. In this paper, the models of two types of interference sources are raised respectively: interference source possessing real input vectors and constant reliable interference source. We study the reliability of the systems effected by the interference sources, and the lower bound of the reliability is presented. The results show that it is impossible that arbitrarily reliable systems can not be constructed with the elements effected by above interference sources.

Analysis of the effects of quantization in high-order function neural network

A statistical quantization model is used to analyze of the effects of quantization in digital implementation of high-order function neural network. From the theory we analyse the performance degradation and fault tolerance of the neural network caused by the number of quantization bits and by changing the order. We try to predict the error in the high-order function neural network (HOFNN) given the properties of the network and the number of quantization bits. Experimental results show the error rate is inversely proportional to quantized bits M for HRFNN. The recognition performance of the backpropagation (BP) network and the HRFNN are almost the same for different quantization bits. The networks performance degradation gets worse when the number of bits is lower than 4-bit quantization. The networks performance degradation gets worse when the number of bits is lower than 4-bit quantization.

A novel fast learning algorithms for time-delay neural networks

To counter the drawbacks of long training time required by Waibels time-delay neural networks (TDNN) in phoneme recognition, the paper puts forward several improved fast learning methods for TDNN. Merging the unsupervised Oja rule and the similar error backpropagation algorithm for initial training of TDNN weights can effectively increase the convergence speed. Improving the error energy function and updating the changing of weights according to size of output error, can increase the training speed. From backpropagation along layer, to average overlap part of backpropagation error of the first hidden layer along a frame, the training samples gradually increase the convergence speed increases. For multi-class phonemic modular TDNNs, we improve the architecture of Waibels modular networks, and obtain an optimum modular TDNNs of tree structure to accelerate its learning. Its training time is less than Waibels modular TDNNs.

A robust learning algorithm for the extended AM neural network

The extended associative memory (AM) neural network (EAMNN) has the advantage of performing the classification in noisy environments. We propose a faster robust learning algorithm of EAMNN and a new error cost function based on weighted sum of standard output error and Hamming distance of output error, and the additional derivatives term of first hidden layer neural activation functions. The fast backpropagation training is based on a modified steepest descent method derived by changing the error function to update weights according to output error, thus it speeds up significantly training speed of the MLP and BAM. The algorithm can force the hidden-layer activation to be saturated to reduce sensitivity of the output values to input variables effectively. It improves robustness on classification performance, increases associative memory ability and accelerates training speed of EAMNN. The experiments verify that it is more powerful than other networks. Then we proposed a two level tree structure modular EAMNN for large-set pattern classification.