In a large number of real world dilemmas and related applications the modeling of complex behavior is the central point. Over the past decades, new approaches based on Artificial Neural Networks (ANN) have been proposed to solve problems related to optimization, modeling, decision making, classification, data mining or nonlinear functions (behavior) approximation. Inspired from biological nervous systems and brain structure, Artificial Neural Networks could be seen as information processing systems, which allow elaboration of many original techniques covering a large field of applications. Among their most appealing properties, one can quote their learning and generalization capabilities. The main goal of this paper is to present, through some of main ANN models and based techniques, their real application capability in real world industrial dilemmas. Several examples through industrial and real world applications have been presented and discussed.

Mahinda Pathegama ¹⁾, Özdemir Göl ²⁾

    Computer-aided analysis for cell images acquired by an electron microscope involves a range of image processing steps including edge detection and thresholding. The major problem encountered in automatic cell analysis is the possible presence of  incomplete boundaries of cell features, which prevent the generation of cell feature details including all measurements as the boundaries include very tiny gaps. This paper presents a novel edge-linking technique based on an artificial neural process, which uses directional sensitivity derivatives from an edged image. The input signals applied to the neural layer are integrated with direction-sensitive information produced by an auxiliary  algorithm, which interrogates all the pixels in the 2-D image in order to designate the specified direction in which each edge-end pixel should propagate. The proposed edge-linking technique, implemented as an image-processing algorithm for direction-sensitive selectiveness, provides an effective solution to the problem of porous boundaries  encountered in biological cell image analysis.

HOW A PEAK CAN BE SEARCHED FOR IN AN ALMOST EVERYWHER FLATLAND OF ALTITUDE ZERO? – TINY FLAT ISLAND IN HUGE LAKE

    We are exploring a weight configuration space searching for solutions to make our neural network with spiking neurons do some tasks. For the task of simulating an associative memory model, we have already known one such solution – a weight configuration learned a set of patterns using Hebb’s rule, and we guess we have many others which we have not known so far. In searching for such solutions, we observed that the so-called fitness landscape was almost everywhere completely flatland of altitude zero in which the Hebbian weight configuration is the only unique peak, and in addition, the sidewall of the peak is not gradient at all. In such circumstances how could we search for the other peaks? This paper is a call for challenges to the problem.

    After a brief survey of work dealing with dynamic neurocontrollers changing their internal structure during the “lifetime” of a mobile autonomous robot, we present experiments employing a standard sensor–motor neurocontroller with self–adapting weights.  The change of behavior of the robot is linked to inputs from the environment that cause the emission of artificial neuromodulators (ANMs) in the robot’s neurocontroller.  In its simplest form an outside teacher (human or machine) constantly evaluates the robot’s actions by transmitting positive or negative feedback signals to the robot initiating the internal changes.  The focus of investigations is put on the mechanisms of the interaction of teaching input and structural changes.  A well–known concept for this interaction is Hebbian learning, which is regulated by ANMs in the presented approach.  In extension to related work in evolutionary robotics (ER), we analyze important details of robotic (ontogenetic) learning by experiments measuring the ability of robots to learn simple tasks in a simulated environment without employing evolution. Specifically, we are interested in the comparison of Hebb learning variants, and the crucial question of the correct interpretation of reward or punishment signals by the robot.

    In paper the outcomes of mathematical modeling of statistical recognition of binary images are proposed. The offered hypothesis that the pixels constituting boundary of recognition object and an image background are secondary attributes at recognition is experimentally confirmed. As consequence, recognition reliability can be raised due to exception of these pixels at recognition. Basing on the offered example of training of models on the fixed training sample and taking into account that the prototype model is a special case of the rejecting model we have noted that recognition reliability of offered model cannot be lower than reliability of the prototype. The obtained results will be used in hardware realization of binary comparators.

    A new method for object recognition and classification is presented in this paper. It merges two well-known and tested methods: neural networks and method of minimal eigenvalues. Each of these methods answers for a different part of recognition process. Method of minimal eigenvalues makes preparatory stage of analysis – of coordinates of characteristic points we get the vector describing given image. Next, it is recognized and classified with neural network. Gathering of characteristic points we perform with our view-based algorithm, but other methods should also do. In this work, method was applied for words in Latin alphabet – handwritten and machine-printed. The obtained results are promising.

    In this ongoing research, we present a Self-Observing Heuristic (SOH). SOH is a hybrid computing method. It roots in natural selection and optimization techniques to provide an environmentally driven evolutionary computation scheme, capable of autonomic cumulative learning. Our aim is to realize an adaptive learning system based on neo-Darwinian evolution of neural units. We proceed in two complementary directions. On one hand, we try to automatically compute the costly tuning phase of the configuration and learning parameters of neural networks (NNs). On the other hand, we use meiosis cellular growth as a natural computation technique to bypass palimpsest effects observed when adding new knowledge to previous one. The main idea is to build an event guided growing competitive NN that develops while it learns to tune other NNs’ parameters. Other NNs can be models more or less similar – or even identical – to it. The system adapts itself, learning to teach other models how to learn well.

A COMPARATIVE STUDY BETWEEN A MULTI-MODELS BASED APPROACH AND AN ARTIFICIAL NEURAL NETWORK BASED TECHNIQUE FOR NONLINEAR SYSTEMS IDENTIFICATION

Lamine Thiaw, Mariusz Rybnik, Rachid Malti, Abdennasser Chebira, Kurosh Madani

    Recently, a number of works propose multi-model based approaches to model non-linear systems. Such approach could also been seen as some “specific” approach, inspired from ANN operation mode, where each neurone, represented by one of the local models, realizes some higher level transfer function. In this paper we present two different neural based approaches to such multiple models concept: one issued from conventional structure and the other based on self-organizing dynamic architecture. A comparative study between a multi-model based architecture and an ANN based one, in the frame of nonlinear system identification is reported.

AN ANTHROPOMORPHIC SPEECH PROCESSING BASED ON THE COCHLEAR MODEL AND ITS APPLICATION FOR CODING TASK

    According to antisymmetry of basilar membrane (BM) movements, a new mathematical model of cochlea is derived using viscous cochlear fluid theory, and then transformed into a digital cochlear model with bilinear transformation. The frequency responses are found to be quite consistent with the experimental data, especially the high frequency slope is much more improved. A new cochlear map and 3 dB bandwidth characteristics for cochlear filter banks are obtained and presented, which will make applications of cochlear model more quantitative and accurate. Due to simplicity of its structure and reality of its characteristics, it will be proved the model can be used easily in speech processing system.

Qiangfu Zhao

    Models for machine learning can be categorized roughly into two groups: symbolic and non-symbolic Generally speaking, symbolic model based learning can provide understandable results, but cannot adapt to changing environments efficiently. On the other hand, non-symbolic model based learning can adapt to changing environments, but the results are usually "black-boxes”. In our study, we introduced a hybrid model called neural network tree (NNTree). An NNTree is a decision tree (DT) with each non-terminal node containing an expert neural network (ENN). Results obtained so far show that an NNTree can be re-trained incrementally using new data. In addition, an NNTree can be interpreted easily if we restrict the number of inputs for each ENN. Thus, it is possible to perform recognition, learning and understand using the NNTree model alone.

COMPUTING OF LYAPUNOV EXPONENTS TECHNIQUES USING NEURAL NETWORKS

    The authors examine neural network techniques for computing of Lyapunov spectrum using observations from unknown dynamical system. Such an approach is based on applying of multilayer perceptron (MLP) for forecasting the next state of dynamical system from the previous one. It allows for evaluating the Lyapunov spectrum of unknown dynamical system accurately and efficiently only by using scalar time series. The results of experiments are discussed.

I. A. Reznik ¹⁾, R.Kh. Sadykhov ²⁾

    In the given work the system of the hidden transfer of the graphic information is considered. For hiding the graphic information the digital correlations based on the complex BIFORE transform are used. A crypto stability of a technique is provided by a secret key, with that the hidden image is embedded in the container-image. The problems of efficiency, robustness, accuracy and performance of the suggested method are considered.