Narrow Results

In this paper, we propose a model of a neural network ensemble that can be trained with a supervisor having two kinds of input-output functions where the occurrence probability of each function is not even. This condition can be likened to a learning condition, in which the learning data are hampered by noise. In this case, the neural network has the impression that the learning supervisor (object) has a probabilistic behavior in which the supervisor generates correct learning data most of the time but occasionally generates erroneous ones. The objective is to train the neural network to approximate the greatest distributed input-output relation, which can be considered to be the principal nature of the supervisor, so that we can obtain a neural network that is able, to some extent, to suppress the ill effect of erroneous data encountered during the learning process.

We investigate the memory structure and retrieval of the brain and propose a hybrid neural network of addressable and content-addressable memory which is a special database model and can memorize and retrieve any piece of information (a binary pattern) both addressably and content-addressably. The architecture of this hybrid neural network is hierarchical and takes the form of a tree of slabs which consist of binary neurons with the same array. Simplex memory neural networks are considered as the slabs of basic memory units, being distributed on the terminal vertexes of the tree. It is shown by theoretical analysis that the hybrid neural network is able to be constructed with Hebbian and competitive learning rules, and some other important characteristics of its learning and memory behavior are also consistent with those of the brain. Moreover, we demonstrate the hybrid neural network on a set of ten binary numeral patters.

In this paper, we extend our greedy network-growing algorithm to multi-layered networks. With multi-layered networks, we can solve many complex problems that single-layered networks fail to solve. In addition, the network-growing algorithm is used in conjunction with teacher-directed learning that produces appropriate outputs without computing errors between targets and outputs. Thus, the present algorithm is a very efficient network-growing algorithm. The new algorithm was applied to three problems: the famous vertical-horizontal lines detection problem, a medical data problem and a road classification problem. In all these cases, experimental results confirmed that the method could solve problems that single-layered networks failed to. In addition, information maximization makes it possible to extract salient features in input patterns.

The research aim is to use three clustering technologies for establishing molecular data model of large size sets by comparison between low energy samples (LES) and local molecular samples (LMS). Hierarchical cluster of multi-level tree distance relation, competitive learning network of similar inputs falling into the same cluster and topological SOM are used to analyze 6242 LES and 5000 LMS. Our experiments show that in SOM, there are 24 to 25 Davies-Boulding clustering index and color map cluster units in the LES more than 10 to 12 in the LMS, which is consistent with the results of hierarchical cluster and competitive learning network in the rough. The hierarchical cluster reflects the biggest inter-cluster distance about 30 for the LES is far larger than that of LMS about 10. The intra-cluster distance of LES about 15 is also far bigger than that of LMS about 3. In SOM, there are more cluster borders of high values (black) reflecting large distance and more clusters in the D-matrix and U-matrix of LES than that of LMS, due to the biggest standard deviation range from -8 to 10 of samples feature of the LES is bigger than that of LMS from -2.5 to 2.5.

We present a new neural model which extends the classical competitive learning (CL) by performing a Probabilistic Principal Components Analysis (PPCA) at each neuron. The model also has the ability to learn the number of basis vectors required to represent the principal directions of each cluster, so it overcomes a drawback of most local PCA models, where the dimensionality of a cluster must be fixed a priori. Experimental results are presented to show the performance of the network with multispectral image data.

In this paper, a structurally enhanced incremental neural learning technique is proposed to learn a discriminative codebook representation of images for effective image classification applications. In order to accommodate the relationships such as structures and distributions among visual words into the codebook learning process, we develop an online codebook graph learning method based on a novel structurally enhanced incremental learning technique, called as "visualization-induced self-organized incremental neural network (ViSOINN)". The hidden structural information in the images is embedded into the graph representation evolving dynamically with the adaptive and competitive learning mechanism. Afterwards, image features can be coded using a sub-graph extraction process based on the learned codebook graph, and a classifier is subsequently used to complete the image classification task. Compared with other codebook learning algorithms originated from the classical Bag-of-Features (BoF) model, ViSOINN holds the following advantages: (1) it learns codebook efficiently and effectively from a small training set; (2) it models the relationships among visual words in metric scaling fashion, so preserving high discriminative power; (3) it automatically learns the codebook without a fixed pre-defined size; and (4) it enhances and preserves better the structure of the data. These characteristics help to improve image classification performance and make it more suitable for handling large-scale image classification tasks. Experimental results on the widely used Caltech-101 and Caltech-256 benchmark datasets demonstrate that ViSOINN achieves markedly improved performance and reduces the computational cost considerably.

One of the most important challenges in computer vision applications is the background modeling, especially when the background is dynamic and the input distribution might not be stationary, i.e. the distribution of the input data could change with time (e.g. changing illuminations, waving trees, water, etc.). In this work, an unsupervised learning neural network is proposed which is able to cope with progressive changes in the input distribution. It is based on a dual learning mechanism which manages the changes of the input distribution separately from the cluster detection. The proposal is adequate for scenes where the background varies slowly. The performance of the method is tested against several state-of-the-art foreground detectors both quantitatively and qualitatively, with favorable results.

Neural networks differ from traditional approaches to image processing in terms of their ability to adapt to regularities in image structure and to self-organize so as to implement directed transformations. Biomedical ultrasonic images are often degraded in quality by noise and other factors, making enhancement techniques particularly important. This paper describes use of back propagation and competitive learning for enhancement and segmentation of ultrasonic images of the eye. Of particular interest is the extension or these technique to segmentation of three-dimensional data sets, where simple thresholding and gradient operations are not entirely successful.

This paper proposes a novel method which enables a Chinese character recognition system to obtain reliable recognition. In this method, two thresholds, i.e. class region thresholdR_k and disambiguity thresholdA_k, are used by each Chinese character k when the classifier is designed based on the nearest neighbor rule, where R_k defines the pattern distribution region of character k, and A_k prevents the samples not belonging to character k from being ambiguously recognized as character k. A novel algorithm to derive the appropriate thresholds A_k and R_k is developed so that a better recognition reliability can be obtained through iterative learning. Experiments performed on the ITRI printed Chinese character database have achieved highly reliable recognition performance (such as 0.999 reliability with a 95.14% recognition rate), which shows the feasibility and effectiveness of the proposed method.

Recently many techniques, e.g., Google or AltaVista, are available for classifying well-organized, hierarchical crisp categories from human constructed web pages such as that in Yahoo. However, given the current rate of web-page production, there is an urgent need of classifiers that are able to autonomously classify web-page categories that have overlaps. In this paper, we present a competitive learning method for this problem, which based on a new objective function and gradient descent scheme. Experimental results on real-world data show that the approach proposed in this paper gives a better performance in classifying randomly-generated, knowledge-overlapped categories as well as hierarchical crisp categories.

Self-Organizing Mapping (SOM) neural network has been widely used in pattern classification, vector quantization, and image compression. We consider the problem of strengthening the reliability of a SOM neural network by the technique of fault immunization of the synaptic links of each neuron which is similar to the concept of biological immunization. Instead of assuming the stuck-at-0 and stuck-at-1 as in those studies, we consider a general case of stuck-at-a, where a is a real value. The only assumption that we consider is only one neuron can be faulty at any time. No restriction on the number of faulty links of the neuron. Let w_i,j be the weight of synaptic link j of neuron i obtained after the winner-take-all classification. Weight w_i,j is immunized by adding a constant ∊_i,j, either positive or negative, to w_i,j. A neuron reaches its maximum fault immunization if the value of w_i,j + ∊_i,j can be either increased or decreased as much as possible without creating any misclassification. Thus, the fault immunization problem is formulated as an optimization problem on finding the value of each ∊_i,j. A technique to find the value of w_i,j + ∊_i,j is proposed and its application to enhance the transmission reliability in image compression area is introduced.

Business simulators are powerful tools for both supporting the decision-making process of business managers as well as for business education. An example is SIMBA (SIMulator for Business Administration), a powerful simulator which is currently used as a web-based platform for business education in different institutions. In this paper, we propose the application of reinforcement learning (RL) for the creation of intelligent agents that can manage virtual companies in SIMBA. This application is not trivial, given the particular intrinsic characteristics of SIMBA: it is a generalized domain where hundreds of parameters modify the domain behavior; it is a multi-agent domain where both cooperation and competition among different agents can coexist; it is required to set dozens of continuous decision variables for a given business decision, which is made only after the study of hundreds of continuous variables. We will demonstrate empirically that all these challenges can be overcome through the use of RL, showing results for different learning scenarios.

Clustering methods, such as k-means, Fuzzy C-Means (FCM), and others have been developed. However, they are only partitioning a database, so it is difficult to discover the reason why each cluster is formed. This paper proposes a method to discover the knowledge of how the clusters are derived. To select the center vector of each cluster, we employ an unsupervised clustering method based on Self-Organizing Map (SOM) without giving the number of clusters. We define the degree of contribution calculated from the weights of the neural network which learned the center vectors. We then describe the knowledge discovery method from the degrees. We applied our method to the artificial data and the clustering problem. The results show that the degree of contribution is an efficient indicator to represent the knowledge of how the clusters are formed.

This paper proposes a novel method which enables a Chinese character recognition system to obtain reliable recognition. In this method, two thresholds, i.e. class region thresholdR_k and disambiguity thresholdA_k, are used by each Chinese character k when the classifier is designed based on the nearest neighbor rule, where R_k defines the pattern distribution region of character k and A_k prevents the samples not belonging to character k from being ambiguously recognized as character k. A novel algorithm to derive the appropriate thresholds A_k and R_k is developed so that a better recognition reliability can be obtained through iterative learning. Experiments performed on the ITRI printed Chinese character database have achieved highly reliable recognition performance (such as 0.999 reliability with a 95.14% recognition rate), which shows the feasibility and effectiveness of the proposed method.

Neural networks differ from traditional approaches to image processing in terms of their ability to adapt to regularities in image structure and to self-organize so as to implement directed transformations. Biomedical ultrasonic images are often degraded in quality by noise and other factors, making enhancement techniques particularly important. This paper describes use of back propagation and competitive learning for enhancement and segmentation of ultrasonic images of the eye. Of particular interest is the extension of these techniques to segmentation of three-dimensional data sets, where simple thresholding and gradient operations are not entirely successful.