Informatica

Text categorization – the assignment of natural language documents to one or more predefined categories based on their semantic content – is an important component in many information organization and management tasks. Performance of neural networks learning is known to be sensitive to the initial weights and architecture. This paper discusses the use multilayer neural network initialization with decision tree classifier for improving text categorization accuracy. Decision tree from root node until a final leave is used for initialization of each single unit. Growing decision trees with increasingly larger amounts of training data will result in larger decision tree sizes. As a result, the neural networks constructed from these decision trees are often larger and more complex than necessary. Appropriate choice of certainty factor is able to produce trees that are essentially constant in size in the face of increasingly larger training sets. Experimental results support the conclusion that error based pruning can be used to produce appropriately sized trees, which are directly mapped to optimal neural network architecture with good accuracy. The experimental evaluation demonstrates this approach provides better classification accuracy with Reuters‐21578 corpus, one of the standard benchmarks for text categorization tasks. We present results comparing the accuracy of this approach with multilayer neural network initialized with traditional random method and decision tree classifiers.

The development of Lithuanian HMM/ANN speech recognition system, which combines artificial neural networks (ANNs) and hidden Markov models (HMMs), is described in this paper. A hybrid HMM/ANN architecture was applied in the system. In this architecture, a fully connected three‐layer neural network (a multi‐layer perceptron) is trained by conventional stochastic back‐propagation algorithm to estimate the probability of 115 context‐independent phonetic categories and during recognition it is used as a state output probability estimator. The hybrid HMM/ANN speech recognition system based on Mel Frequency Cepstral Coefficients (MFCC) was developed using CSLU Toolkit. The system was tested on the VDU isolated‐word Lithuanian speech corpus and evaluated on a speaker‐independent ∼750 distinct isolated‐word recognition task. The word recognition accuracy obtained was about 86.7%.

The article presents a limited‐vocabulary speaker independent continuous Estonian speech recognition system based on hidden Markov models. The system is trained using an annotated Estonian speech database of 60 speakers, approximately 4 hours in duration. Words are modelled using clustered triphones with multiple Gaussian mixture components. The system is evaluated using a number recognition task and a simple medium‐vocabulary recognition task. The system performance is explored by employing acoustic models of increasing complexity. The number recognizer achieves an accuracy of 97%. The medium‐vocabulary system recognizes 82.9% words correctly if operating in real time. The correctness increases to 90.6% if real‐time requirement is discarded.

Color constancy is the perceived stability of the color of objects under different illuminants. Four-layer neural network for color constancy has been developed. It has separate input channels for the test chip and for the background. Input of network was RGB receptors. Second layer consisted of color opponent cells and output have three neurons signaling x, y, Y coordinates (1931 CIE). Network was trained with the back-propagation algorithm. For training and testing we used nine illuminants with wide spectrum. Neural network was able to achieve color constancy. Input of background coordinates and nonlinearity of network have crucial influence for training.

Neural networks are often characterized as highly nonlinear systems of fairly large amount of parameters (in order of 10³ – 10⁴). This fact makes the optimization of parameters to be a nontrivial problem. But the astonishing moment is that the local optimization technique is widely used and yields reliable convergence in many cases. Obviously, the optimization of neural networks is high-dimensional, multi-extremal problem, so, as usual, the global optimization methods would be applied in this case. On the basis of Perceptron-like unit (which is the building block for the most architectures of neural networks) we analyze why the local optimization technique is so successful in the field of neural networks. The result is that a linear approximation of the neural network can be sufficient to evaluate the start point for the local optimization procedure in the nonlinear regime. This result can help in developing faster and more robust algorithms for the optimization of neural network parameters.