Stefano Rovetta

Circular backpropagation networks for classification

The class of mapping networks is a general family of tools to perform a wide variety of tasks. This paper presents a standardized, uniform representation for this class of networks, and introduces a simple modification of the multilayer perceptron with interesting practical properties, especially well suited to cope with pattern classification tasks. The proposed model unifies the two main representation paradigms found in the class of mapping networks for classification, namely, the surface-based and the prototype-based schemes, while retaining the advantage of being trainable by backpropagation. The enhancement in the representation properties and the generalization performance are assessed through results about the worst-case requirement in terms of hidden units and about the Vapnik-Chervonenkis dimension and cover capacity. The theoretical properties of the network also suggest that the proposed modification to the multilayer perceptron is in many senses optimal. A number of experimental verifications also confirm theoretical results about the models increased performances, as compared with the multilayer perceptron and the Gaussian radial basis functions network.

Vector quantization for license-plate location and image coding

License-plate location in sensor images plays an important role in vehicle identification for automated transport systems (ATS). This paper presents a novel method based on vector quantization (VQ) to process vehicle images. The proposed method makes it possible to perform superior picture compression for archival purposes and to support effective location at the same time. As compared with classical approaches, VQ encoding can give some hints about the contents of image regions; such additional information can be exploited to boost location performance. The VQ system can be trained by way of examples; this gives the advantages of adaptiveness and on-field tuning. The approach has been tested in a real industrial application and included satisfactorily in a complete ATS for vehicle identification.

Soft transition from probabilistic to possibilistic fuzzy clustering

In the fuzzy clustering literature, two main types of membership are usually considered: A relative type, termed probabilistic, and an absolute or possibilistic type, indicating the strength of the attribution to any cluster independent from the rest. There are works addressing the unification of the two schemes. Here, we focus on providing a model for the transition from one schema to the other, to exploit the dual information given by the two schemes, and to add flexibility for the interpretation of results. We apply an uncertainty model based on interval values to memberships in the clustering framework, obtaining a framework that we term graded possibility. We outline a basic example of graded possibilistic clustering algorithm and add some practical remarks about its implementation. The experimental demonstrations presented highlight the different properties attainable through appropriate implementation of a suitable graded possibilistic model. An interesting application is found in automated segmentation of diagnostic medical images, where the model provides an interactive visualization tool for this task

K-winner machines for pattern classification

The paper describes the K-winner machine (KWM) model for classification. KWM training uses unsupervised vector quantization and subsequent calibration to label data-space partitions. A K-winner classifier seeks the largest set of best-matching prototypes agreeing on a test pattern, and provides a local-level measure of confidence. A theoretical analysis characterizes the growth function of a K-winner classifier, and the result leads to tight bounds to generalization performance. The method proves suitable for high-dimensional multiclass problems with large amounts of data. Experimental results on both a synthetic and a real domain (NIST handwritten numerals) confirm the approach effectiveness and the consistency of the theoretical framework.

Objective quality assessment of MPEG-2 video streams by using CBP neural networks

The increasing use of compression standards in broadcasting digital TV has raised the need for established criteria to measure perceived quality. Novel methods must take into account the specific artifacts introduced by digital compression techniques. This paper presents a methodology using circular backpropagation (CBP) neural networks for the objective quality assessment of motion picture expert group (MPEG) video streams. Objective features are continuously extracted from compressed video streams on a frame-by-frame basis; they feed the CBP network estimating the corresponding perceived quality. The resulting adaptive modeling of subjective perception supports a real-time system for monitoring displayed video quality. The overall system mimics perception but does not require an analytical model of the underlying physical phenomenon. The ability to process compressed video streams represents a crucial advantage over existing approaches, as avoiding the decoding process greatly enhances the systems real-time performance. Experimental evidence confirmed the approach validity. The system was tested on real test videos; they included different contents ranging from fiction to sport. The neural model provided a satisfactory, continuous-time approximation for actual scoring curves, which was validated statistically in terms of confidence analysis. As expected, videos with slow-varying contents such as fiction featured the best performances.

Efficient training of neural gas vector quantizers with analog circuit implementation

This paper presents an algorithm for training vector quantizers with an improved version of the neural gas model, and its implementation in analog circuitry. Theoretical properties of the algorithm are proven that clarify the performance of the method in terms of quantization quality, and motivate design aspects of the hardware implementation. The architecture for vector quantization training includes two chips, one for Euclidean distance computation, the other for programmable sorting of codevectors. Experimental results obtained in a real application (image coding) support both the algorithms effectiveness and the hardware performance, which can speed up the training process by up to two orders of magnitude.

Automated diagnosis and disease characterization using neural network analysis

A neural network approach is used to analyze and diagnose a rather new and uncommon disease, Lyme borreliosis. To fully exploit the methods generalizing power, a significance analysis split the set of inputs of a trained network into two classes that were important and unimportant. The results of this analysis lead to a new structured network, whose topology and architecture reflect the estimated relevance of symptoms. The diagnostic performance thus obtained showed a dramatic improvement which reached an average error rate of around 6%.<<ETX>>

Circular backpropagation networks embed vector quantization

This letter proves the equivalence between vector quantization (VQ) classifiers and circular backpropagation (CBP) networks. The calibrated prototypes for a VQ schema can be plugged in a CBP feedforward structure having the same number of hidden neurons and featuring the same mapping. The letter describes how to exploit such equivalence by using VQ prototypes to perform a meaningful initialization for BP optimization. The approach effectiveness was tested considering a real classification problem (NIST handwritten digits).

Plastic algorithm for adaptive vector quantisation

A plastic algorithm for building vector quantisers adaptively attains a dynamic representation of observed data; an unsupervised version of classical crossvalidation rules the algorithms stopping condition. Combining plasticity with empirical generalisation-based control yields an adaptive methodology for VQ. The paper analyses the methods convergence properties and discusses the models generalisation performance. Experimental results on synthetic and real, complex testbeds support the models validity.

Applying the Possibilistic c-Means Algorithm in Kernel-Induced Spaces

In this paper, we study a kernel extension of the classic possibilistic c-means. In the proposed extension, we implicitly map input patterns into a possibly high-dimensional space by means of positive semidefinite kernels. In this new space, we model the mapped data by means of the possibilistic clustering algorithm. We study in more detail the special case where we model the mapped data using a single cluster only, since it turns out to have many interesting properties. The modeled memberships in kernel-induced spaces yield a modeling of generic shapes in the input space. We analyze in detail the connections to one-class support vector machines and kernel density estimation, thus, suggesting that the proposed algorithm can be used in many scenarios of unsupervised learning. In the experimental part, we analyze the stability and the accuracy of the proposed algorithm on some synthetic and real datasets. The results show high stability and good performances in terms of accuracy.