Search

In the present work, we have proposed a novel Bangla word segmentation technique that is based on stroke-level busy zone formation procedure. In an unconstrained domain, people often write text where strokes may be poorly aligned (due to multi-directional skewness) and varied combination of strokes with various types of joining between them are possible while forming the words. Hence, a segmentation approach for stroke extraction is pertinent for any stroke-based word recognition system. The presence of a large volume of symbols set (58 basic symbols with more than 280 compound characters) in Bangla script makes the task more challenging. In the current experiment, our stroke-level segmentation approach effectively handles such type of Bangla words. A sub-zoning scheme within busy zone followed by a modified Down->Up->Down (DUD) concept within these sub-zones has been used to find valid segmentation points. This scheme avoids over and under-segmentation issues caused by either inherent writing pattern or due to writing style variations up to certain extent. The proposed segmentation approach has been tested on 6500 online handwritten Bangla word samples with 98.45% correct segmentation accuracy (compared with manually generated ground truth of the same database).

Ensemble models can achieve more accurate and robust predictions than single learners. A selective ensemble may further improve the predictions by selecting a subset of the models from the entire ensemble, based on a quality criterion. We consider reinforcement learning ensembles, where the members are artificial neural networks. In this context, we extensively evaluate a recently introduced algorithm for ensemble subset selection in reinforcement learning scenarios. The aim of the learning strategy is to select members whose weak decisions are compensated by strong decisions for collected states. The correctness of a decision is determined by the Bellman error. In our empirical evaluations, we compare the benchmark performances of the full ensembles and the selective ensembles in generalized maze and in SZ-Tetris. Both are large state environments. We found that while the selective ensembles have a small number of agents, they significantly outperform the large ensembles. We therefore conclude that selecting an informative subset of many agents may be more efficient than training single agents or full ensembles.

The paper presents a probabilistic echo-state network (π-ESN) for density estimation over variable-length sequences of multivariate random vectors. The π-ESN stems from the combination of the reservoir of an ESN and a parametric density model based on radial basis functions. A constrained maximum likelihood training algorithm is introduced, suitable for sequence classification. Extensions of the algorithm to unsupervised clustering and semi-supervised learning (SSL) of sequences are proposed. Experiments in emotion recognition from speech signals are conducted on the WaSeP? dataset. Compared with established techniques, the π-ESN yields the highest recognition accuracies, and shows interesting clustering and SSL capabilities.

Labelling real world data sets is a difficult problem. Often, the human expert is unsure about a class label of a specific sample point or, in case of very large data sets, it is impractical to label them manually. In semi-supervised clustering, the sample labels, which are external informations, are used to find better matching cluster partitions. Further, kernel-based clustering methods are able to partition the data with nonlinear boundaries in feature space. While these methods improve the clustering results, they have a quadratic computation time. In this paper, we propose a meta-algorithm that processes small-sized subsets of a large data set, clusters them with the sample labels and merges the points close to the resulting prototypes with the next points, until the whole data set has been processed. It has a linear computation time. The error function that this meta-algorithm minimizes is presented. Although we applied this meta-algorithm to Kernel Fuzzy C-Means, Relational Neural Gas and Kernel K-Means, it can be applied to a broad range of kernel-based clustering methods. The proposed method has been empirically evaluated on two real world benchmark data sets.

In this paper, we discuss radial basis function (RBF) classifiers and classifier fusion architectures to categorize bioacoustic time series. The proposed methods are evaluated on a data base of cricket songs of 35 different species recorded in Thailand and Ecuador. Local characteristic features are extracted from the recordings and then categorized by RBF classifiers. These local decisions are then combined through temporal fusion. Three different fusion techniques (Averaging, Percentiles and Voting) are considered. We present and discuss the classification results for the proposed classifier architectures on this data set of 35 species.

In this paper, learning algorithms for radial basis function (RBF) networks are discussed. Whereas multilayer perceptrons (MLP) are typically trained with backpropagation algorithms, starting the training procedure with a random initialization of the MLP's parameters, an RBF network may be trained in many different ways. We categorize these RBF training methods into one-, two-, and three-phase learning schemes. Two-phase RBF learning is a very common learning scheme. The two layers of an RBF network are learnt separately; first the RBF layer is trained, including the adaptation of centers and scaling parameters, and then the weights of the output layer are adapted. RBF centers may be trained by clustering, vector quantization and classification tree algorithms, and the output layer by supervised learning (through gradient descent or pseudo inverse solution). Results from numerical experiments of RBF classifiers trained by two-phase learning are presented in three completely different pattern recognition applications: (a) the classification of 3D visual objects; (b) the recognition hand-written digits (2D objects); and (c) the categorization of high-resolution electrocardiograms given as a time series (1D objects) and as a set of features extracted from these time series. In these applications, it can be observed that the performance of RBF classifiers trained with two-phase learning can be improved through a third backpropagation-like training phase of the RBF network, adapting the whole set of parameters (RBF centers, scaling parameters, and output layer weights) simultaneously. This, we call three-phase learning in RBF networks. A practical advantage of two- and three-phase learning in RBF networks is the possibility to use unlabeled training data for the first training phase. Support vector (SV) learning in RBF networks is a different learning approach. SV learning can be considered, in this context of learning, as a special type of one-phase learning, where only the output layer weights of the RBF network are calculated, and the RBF centers are restricted to be a subset of the training data. Numerical experiments with several classifier schemes including k-nearest-neighbor, learning vector quantization and RBF classifiers trained through two-phase, three-phase and support vector learning are given. The performance of the RBF classifiers trained through SV learning and three-phase learning are superior to the results of two-phase learning, but SV learning often leads to complex network structures, since the number of support vectors is not a small fraction of the total number of data points.

We investigate the pattern completion performance of neural auto-associative memories composed of binary threshold neurons for sparsely coded binary memory patterns. By focussing on iterative retrieval, we are able to introduce effective threshold control strategies. These are investigated by means of computer simulation experiments and analytical treatment. To evaluate the systems performance we consider the completion capacity C and the mean retrieval errors. The asymptotic completion capacity values for the recall of sparsely coded binary patterns in one-step retrieval is known to be ln 2 4 ≈ 17.32% for binary Hebbian learning, and 1 (8 ln 2) ≈ 18% for additive Hebbian learning. These values are accomplished with vanishing error probability and yet are higher than those obtained in other known neural memory models. Recent investigations on binary Hebbian learning have proved that iterative retrieval as a more refined retrieval method does not improve the asymptotic completion capacity of one step retrieval. In a finite size auto-associative memory we show that iterative retrieval achieves higher capacity and better error correction than one-step retrieval. One-step retrieval produces high retrieval errors at optimal memory load. Iterative retrieval reduces the retrieval errors within a few iteration steps (t ⩽ 5). Experiments with additive Hebbian learning show that in the finite model, binary Hebbian learning exhibits much better performance. Thus the main concern of this paper is binary Hebbian learning. We examine iterative retrieval in experiments with up to n = 20,000 threshold neurons. With this system size one-step retrieval yields a completion capacity of about 16%, the second retrieval step increases this value to 17.9% and with iterative retrieval we obtain 19%. The first two retrieval steps in the finite system have also been treated analytically. For one-step retrieval the asymptotic capacity value is approximated from below with growing system size. In the second retrieval step (and as the experiments suggest also for iterative retrieval) the finite size behaviour is different. The capacity exceeds the asymptotic value, reaches an optimum for finite system size, and decreases to the asymptotic limit.