Artificial neural networks have popularized problem-solving with algorithms, which rely on input-output examples instead of causal models to operate. The deep learning (DL) variety has notably been shown to efficiently solve complex classification and prediction problems when large training sets and big computational resources are available. However, these constraints are not always met in science and engineering, where adapted DL solutions are necessary. A case in point is hand prostheses where the training data is scarce, and the computational resources limited. The tutorial review explores these challenges when applying AI to such devices. Using the example of a surface electromyography-driven hand prosthesis control system, the sensing, signal processing, and machine learning pipelines of a wearable system are detailed.

A machine learning neural model is investigated for obstacle detection by autonomous trains drives. Transfer learning is used in conjunction with a pretrained Inception-ResNet-v2 convolutional neural network (CNN) that is fine-tuned with the RailSem19 set of images. Given the small size of the set and its data imbalance, various data augmentation techniques are explored to improve the model’s detection accuracy. The obtained results show that class size balancing and synthetic augmentation of the training data improve the average detection accuracy from 78% with the original RailSem19 training set to up to 94.06% with balanced data augmentation, with 91.66% precision and 96.85% recall, corresponding to an F1 score of 95.43%.

A deep neural network architecture for the automatic sizing of analog circuit components is proposed, with a focus on radio frequency (RF) applications in the 2 to 5 GHz region. It addresses the challenges of the typically small number of examples for network training and the existence of multiple solutions, of which impractical values for integrated circuit implementation. We address these issues by restricting the learning to one component size at a time, thanks to a cascade of dedicated shallow neural networks (SNN), where each network constrains the prediction of the next ones. Moreover, the SNNs are individually tuned by a genetic algorithm for the prediction order and accuracy. This reduction of the solution space at each step allows the use of small training sets, and the introduced constraints between SNNs handle component interdependencies. The method is successfully validated on three different types of RF microcircuits: a low-noise amplifier (LNA), a voltage-controlled oscillator (VCO), and a mixer, using 180 nm and 130 nm CMOS implementations. All the predictions were within 5 % of the true values, both at the component and performance levels, and all the responses were obtained in less than 5 s, after 4 to 47 min. training on a regular PC station. The obtained results show that the proposed method is fast and applicable to arbitrary analog circuit topologies, with no need to retrain the developed neural network for each new set of desired circuit performances.

Human dialogues often show underlying dependencies between turns, with each interlocutor influencing the queries/responses of the other. This article follows this by proposing a neural architecture for conversation modeling that looks at the dialogue history of both sides. It consists of a generative model where one encoder feeds three decoders to process three successive turns of dialogue for predicting the next utterance, with a multidimension attention mechanism aggregating the past and current contexts for a cascade effect on each decoder. As a result, a more comprehensive account of the dialogue evolution is obtained than by focusing on a single turn or the last encoder context, or on the user side alone. The response generation performance of the model is evaluated on three corpora of different sizes and topics, and a comparison is made with six recent generative neural architectures, using both automatic metrics and human judgments. Our results show that the proposed architecture equals or improves the state-of-the-art for adequacy and fluency, particularly when large open-domain corpora are used in the training. Moreover, it allows better tracking of the dialogue state evolution for response explainability.

We present a generative neural model for open and multi-turn dialog response generation that relies on a multi-dimension attention process to account for the semantic interdependence between the generated words and the conversational history, so as to identify all the words and utterances that influence each generated response. The performance of the model is evaluated on the wide scope DailyDialog corpus and a comparison is made with two other generative neural architectures, using several machine metrics. The results show that the proposed model improves the state of the art for generation accuracy, and its multi-dimension attention allows for a more detailed tracking of the influential words and utterances in the dialog history for response explainability by the dialog history.

The presence of unwanted signals in the radio frequency (RF) spectrum, called RF interference (RFI), is a major drawback in wireless communication systems. The detection of REI has been dealt mainly with signal processing and supervising machine learning approaches. In this paper, we investigate two unsupervised machine learning alternatives for REI detection, the one-class support vector machine (SVM) and the support vector data description (SVDD) algorithm, which delimit the class boundaries of normal signals in high-dimensional space, and view REI contaminated signal as novelty, i.e., outsiders from unknown classes. Similar to the popular binary SVM classifier, these two algorithms can learn from a relatively small training set, and they use unsupervised training to learn from typically unbalanced RFI data sets without need for data augmentation techniques as in supervised training. The experimental results for detecting three types of RFI, using scaling features to a range as a standardization method, show that SVDD has a low computational complexity and an accuracy of 90.74 % versus 91.67 % for the One-class SVM.

Studying drought in a greenhouse setting allows to analyze plant growth under controlled environmental conditions. However, simulating different drought intensities by varying the soil moisture is challenging. This study describes a sensory and control system to simulate drought conditions for wheat, within a framework to study this crop's genetic responses under drought stress. The system uses drip irrigation and allows to maintain the soil moisture within a specified range on potted wheat plants. It allows identifying the amount of water required for irrigation in wheat growth stages, and conducting biological experiments to understand the effects of drought stress on wheat growth.

A domain ontology (DO) is a machine-readable knowledge repository which, whenever properly exploited, can help to discover meaningful and intelligible patterns from compatible datasets. Yet since such data is naturally graph-shaped, the corresponding task amounts to mining what we call ontologically-generalized graph patterns. We study the underlying problem within a dairy production context where a dedicated DO has been designed beforehand. Two alternative mining approaches have been designed, both representing adaptations of methods from the literature. We evaluated them on an excerpt from our dairy production dataset and report here their respective limitations. We also sketch a way to approach the design of ontology-powered graph miner.

Several obstacles face the development of noninvasive blood glucose sensing using printed electronics. They relate both to the sensing technique and the appropriate substrates to use. This article reviews three potential approaches based on tears, sweat and interstitial fluid (ISF) analysis, whence the mechanism, materials and substrates, performance and limitations are reported. A comparison of the three techniques leads to conclude that using iontophoresis followed by electrochemical-enzymatic glucose sensing may be the most promising approach given the current state-of-the-art. However, the definite relationship between ISF glucose concentration, and blood glucose concentration, is yet to be established to ensure the method's reliability.

Myopathy is a disease of the skeletal muscle tissue that is generally diagnosed by electromyography (EMG). Computer-aided-diagnosis (CAD) systems that use the auto-regressive coefficients of the EMG signal for detecting myopathy are popular due to their simplicity and fast response. However, those systems do not consider the full linear dynamics of the EMG signal. We propose a CAD system that uses the estimated parameters of an auto-regressive moving average (ARMA) model and the variance of its disturbances to fully capture the linear dynamics and variability in the EMG signal. Then, those parameters are fed to a linear discriminant analysis (LDA) classifier to distinguish between healthy and myopathic signals. The experimental results of processing 160 EMG records, using 10-fold cross validation show strong improvement in accuracy, sensitivity, and specificity performance measures along with very low variability. Moreover, the system outperformed other works reported in the literature.

In this paper, we report experimental results of hybrid system using Hidden Markov Models/Multi-Layer Perceptron (HMM/MLP) model as acoustic model and based on the Fuzzy C-Means (FCM) clustering with optimization with Genetic Algorithm (GA). In this context, we use the result of FCM clustering as initial population of GA, this allows training the GA with a population of empirically generated chromosomes and not randomly initialized. Our results on speech recognition tasks show an increase in the estimates of the posterior probabilities of the correct words after training. We demonstrate the effectiveness of the proposed clustering approach in large-vocabulary speaker-independent continuous speech recognition with regard to the three baseline systems : Discrete HMM, hybrid HMM/MLP with K-Means and FCM clustering.

Electromyography (EMG) is commonly used for the diagnosis of neuromuscular and motor control disorders. This work presents an accurate system for the classification of healthy and neuropathic EMG signals. The proposed system employs an auto-regressive moving average (ARMA) model to capture the evolution in time of the EMG signal. Then, the obtained ARMA coefficients serve as inputs to a linear discriminant analysis (LDA) classifier to distinguish between healthy and neuropathic signals. The experimental results of processing 160 EMG records, using tenfold cross validation, indicate 99.74% classification accuracy, with 100% sensitivity and 99.66% specificity. The system outperformed LDA trained with only auto-regressive (AR) coefficients and other works reported in the literature. Moreover, it is fast for clinical applications.

The effect of partial volume related to anatomical MRI and functional images limit the diagnostic potential of brain imaging. To remedy for this problem, we propose a fuzzy-genetic brain segmentation scheme for the assessment of white matter, gray matter and cerebrospinal fluid volumes, from brain images of Alzheimer patients from a real database. This clustering process based on Possibilistic C-Means (PCM) algorithm, which allows modeling the degree of relationship between each voxels and a given tissue; and based on fuzzy genetic initialization for the centers of clusters by a Fuzzy C-Means (FCM) algorithm, and for which the result is optimized by genetic process. The visual results show a concordance between the ground truth segmentation and the hybrid algorithm results, which allows efficient tissue classification. The superiority was also proved with the quantitative results of the proposed method in comparison with the both conventional FCM and PCM algorithms.

A robot is presented whose behavior is based on two fundamental types of learning in the animal world: Classical Conditioning (CC) and Operant Conditioning (OC). It is shown how both share Spike-Timing-Dependent-Plasticity (STDP) as learning process for a Spiking Neural Network (SNN). STDP was implemented on a Field-Programmable Gate Array (FPGA) with very low-demanding resources, using an adaptation of the Synapto-dendritic Kernel Adapting Neuron (SKAN) model. Moreover, it is shown how a 3-way version of STDP is needed to allow for OC. The robot was designed to use the CC and OC neuronal architectures proposed in this paper and was tested in a dynamic environment, which consisted of a maze with changing features. It was successful in presenting both types of learning. This paper thus validates an architecture with an important potential for very large scale time-dependent parallel data analysis, with high capacity of adaptation in a dynamic environment.

Proper acquisition of the photoplethysmography signals is essential in a pulse oximetry system and sensor placement plays an important role in this respect. Due to the complex structure of the finger tissue, inadequate sensor placement will have an adverse effect on the light path and high signal quality may become impossible to achieve [1]. In this paper, we present a ring shaped oximeter that uses six sets of light emitting diodes and photodetectors, uniformly distributed around the finger to identify the best signal path, thus making the signal acquisition immune to ring position. Moreover it uses a radio transceiver to eliminate the connection wires to a base station. In this proof of concept study, this novel ring oximeter was implemented with commercial low power consumption off-the-shelf components mounted on a rigid-flex board that connects to a remote host for signal processing and oxygen level calculation.

We show how a minimal components requirement and very low resource demanding field-programmable gate array (FPGA) implementation of an adapted version of the synapto-dendritic Kernel Adapting Neuron (SKAN) model can be used to underlie two of the most basic learning processes: classical conditioning (CC) and operant conditioning (OC). In the CC architecture, this adapted SKAN model is used in a spiking neural network (SNN) to implement spike-timing-dependant-plasticity (STDP). However, in order to achieve a functioning OC architecture, a new STDP-inspired learning process is introduced. The modified CC architecture, new OC architecture, adapted SKAN model and new STDP-inspired process represent the four contributions presented here, along with simulation results on a FPGA which shows their adequacy in supporting CC and OC learning behaviors.

SysML/UML activity diagrams are widely used for the modeling and analysis of complex systems and they have become a de-facto standard for software and embedded systems. Previously in our group, we formalized SysML activity diagrams by developing a calculus called New Activity Calculus (NuAC). In this work, we redefine NuAC terms to support code generation for ARM Cortex-M processors and we present an automated SysML activity diagram to RTX (Keil Real-Time Operating System) code generator that uses mapping rules expressed in NuAC. To demonstrate the capability of the developed tool, we use it for scheduling of a JPEG Encoder on an ARM Cortex-M4 device.

The use of renewable energy at roadside units (RSUs) in vehicular ad hoc networks is a great alternative to the electric grid, since it lowers the carbon footprint, and the cost of deployment and maintenance. This paper describes a scheduler for serving vehicles by RSUs that use energy harvesting, with the aim to maximize the number of served vehicles. We start by defining an integer linear programming model for finding the optimal offline schedule. The model is shown to be NP-hard and hence we propose a greedy heuristic to solve it. We compare the optimal solution and near- optimal offline heuristic with an energy-efficient scheduler for RSUs. Our simulation results show that the proposed scheduler for energy harvesting RSUs can reduce the service delay of vehicles. It also provides good performance with respect to the percentage of served vehicles, in comparison to energy-efficient scheduler in grid-powered RSUs.

We describe a method for circuit synthesis that determines the parameter values by using a set of artificial neural networks (ANNs) that learn in sequence. Each ANN is optimized to output only one design parameter, and the latter constrains the learning/recall of its successor(s). Two competing ANN architectures are considered, the multilayer perceptron (MLP) and the radial basis functions (RBF) network, and each one has its internal parameters tuned by a genetic algorithm. The method was tested on the design of a radio-frequency, low-noise amplifier (RF-LNA) with ten design parameters to set, and it yielded one-hundred percent success rate in specifying the parameter values at five percent tolerance.

We describe a method for physiological signal denoising based on the variational mode decomposition (VMD), the discrete wavelet transform (DWT), and constrained least squares (CLS) optimization. First, the noisy signal is decomposed into a sum of variational mode functions (VMFs) by VMD. Next, the DWT thresholding technique is applied to each VMF for denoising. Then, a weighted sum of the denoised VMFs is performed after weight estimation by CLS. The summation ignores the residue. This approach is compared to others based on empirical mode decomposition (EMD) and DWT thresholding of the obtained intrinsic mode functions (IMFs) and residue, followed by the unweighted summation of the results. The comparisons were performed with two EEG signals from the left and right cortex of a rat, and one ECG signal from a human subject. Using the signal-to-noise ratio and mean squared error as performance metrics, the results show strong evidence of the superiority of the VMD-DWT-CLS approach over the standard EMD-DWT. It is concluded that using CLS in the final reconstruction stage and ignoring the residue may bring significant improvement to the denoising process.

We present a fractal-based methodology to analyze brain magnetic resonance images (MRI) for the automated detection of cerebral arteriovenous malformations (AVM). First, the MRI is split into right and left hemispheres components whose fractal dimensions (FD) are estimated using detrended fluctuation analysis (DFA). Then, the obtained FD values are used to characterize healthy and AVM-affected brain MRIs. Using a database of twenty-eight images, and ten-fold cross validation, classification by a support vector machine (SVM) was 100% accurate when using either a linear or a radial basis Gaussian kernel, and the total image processing time was 32.75 s on a midrange PC station. It is concluded that the presented cerebral AVM detection system is both simple and accurate, and its processing time makes it compatible for use in a clinical environment, should it performance be confirmed with a larger image database.

The purpose of this study is the prediction of Standard & Poor's (S&P500) trends (ups and downs) with macroeconomic variables, technical indicators, and investor moods using k-NN algorithm and probabilistic neural networks. More precisely, eleven economic factors, twelve technical indicators and four measures of investor's mood were selected as potential predictive variables. Then, the Granger causality test was performed to identify among them the predictive variables that show a strong relationship with the stock market. Finally, the identified inputs are fed to k-NN and PNN separately and the correct detection of stock market ups (+0.5%)-aggressive investment strategy - is computed using the obtained hit ratios. The simulations results from 10-fold experiments show that the average detection rate of k-NN and PNN are respectively 93.45% (±0.0019, standard deviation) and 92.4% (±0.006, standard deviation). The results suggest that aggregating the three categories of information (economic, technical, and psychological information) along with k-NN as classifier leads to high detection accuracy of future stock market ups and downs.

The increasing possibilities of the multicore chip and system on a chip have brought task scheduling to the forefront of efficient system design. In previous work, we described a method to predict the effectiveness of a scheduling algorithm for a given application. It was based on association rules between the attribute values of a set of sample applications and one or more performance metrics for the scheduling algorithm. We then used the approach to devise a methodology to compare scheduling algorithms two at a time, using rule confidence differences. In this paper, we describe a visualization approach that allows comparing an arbitrary number of algorithms, by showing the relevant data in different perspectives thanks to a rules map. Three examples illustrate the effectiveness of the method.

An automated processing system of brain magnetic resonance (MR) images is presented with application to normal versus glioma diagnosis. It exploits lobe asymmetry to distinguish the normal and abnormal brain MR images. Each MR image is first processed to emphasize edges before splitting it into right lobe and left lobe components. These are transformed into one-dimensional signals and the corresponding power spectral density functions (PSDF) are estimated. Then, a four-dimensional feature vector is formed with the energy of each PSDF and their correlation coefficient calculated by two approaches. Using leave-one-out cross validation on a dataset of seven normal and seven glioma affected MR images, 100% classification accuracy was achieved by a support vector machine classifier, with near real-time processing time.

The preliminary work on a new methodology to design low noise amplifiers (LNAs) for use in radiofrequency (RF) wireless systems is presented. The methodology aims to find the relevant design parameters faster than current analytical models and optimization procedures. To reach this goal, an artificial neural network (ANN) is used to learn the design task by being exposed to successful design examples. Our preliminary results, using a training set of two hundred design examples, show that a radial basis functions ANN can learn the provided designs perfectly, but a larger training set is required for definite conclusions regarding the prediction of component values for new designs.

This work compares two CMOS front-end transimpedance amplifiers (TIA) for use in optical biosensors. They are the shunt-feedback and current-mode circuits, the most widely used for wideband operation. The former consists of a three-stage nested-Miller-compensated (NMC) amplifier in non-inverting mode with a photodiode (PD) bootstrapping and a controlled voltage gain; the latter comprises a wideband common-gate feedback (CGFB) current mirror coupled to a current-to-voltage conversion stage and two common-source gain stages. The simulation results show that the shunt-feedback TIA achieves a maximal gain of 112 dBΩ over a 2 MHz bandwidth, whereas the current-mode TIA has a flat gain of roughly 83 dBΩ over a 115 MHz bandwidth. The overall input rms noise of each circuit was 185pA/√Hz and 53nA/√Hz, respectively, with power consumptions of 0.5 mW and 28.6 mW. It is concluded that the shunt-feedback TIA is a better choice for low to mid-frequency applications.

A dual-mode, low-power and low-noise front-end transimpedance amplifier (TIA), intended for use in optical biosensors, is presented. It consists of a three-stage nested-Miller-compensated (NMC) amplifier in a non-inverting mode with a photodiode (PD) bootstrapping and a controlled voltage gain (CVG). The PD front-end can handle properly waveforms with soft and sharp edge transitions. In the former case, it reduces the voltage gain at DC while preserving a high gain in the frequency range of interest; in the latter, it provides a constant voltage gain over the whole TIA bandwidth. The simulation results show that a maximal gain of about 112 dBΩ can be achieved over a 2 MHz bandwidth. The empirical results obtained with a prototype mounted on PCB for blood glucose monitoring are closely correlated with those obtained using a commercial glucose/lactate analyzer.

We propose an evolutionary approach to automatically generate model transformation rules from a set of examples. To this end, genetic programming is adapted to the problem of model transformation in the presence of complex input/output relationships (i.e., models conforming to meta-models) by generating declarative programs (i.e., transformation rules in this case). Our approach does not rely on prior transformation traces for the model-example pairs, and directly generates executable, many-to-many rules with complex conditions. The applicability of the approach is illustrated with the well-known problem of transforming UML class diagrams into relational schemas, using examples collected from the literature.

Building powerful machine-learning assessment models is an important achievement of empirical software engineering research, but it is not the only one. Intelligibility of such models is also needed, especially, in a domain, software engineering, where exploration and knowledge capture is still a challenge. Several algorithms, belonging to various machine-learning approaches, are selected and run on software data collected from medium size applications. Some of these approaches produce models with very high quantitative performances, others give interpretable, intelligible, and "glass-box" models that are very complementary. We consider that the integration of both, in automated decision-making systems for assessing software product quality, is desirable to reach a compromise between performance and intelligibility.

This paper presents a framework for the automatic creation of netlists for arbitrary electronic circuits. The methodology relies on defining interfaces that allow a set of integrated circuits and other electronic components to be interconnected without user intervention. The framework, called "Intelligent Netlist Creator", has been successfully tested on several circuits. The results show that the proposed flow for netlist creation assists the user by automating some connections whenever possible.

The problem of automatic classification of brain images obtained by magnetic resonance imaging (MRI) is considered. In order to design the classification system, a three- stage approach is used. It consists of wavelet decomposition of the image under study, feature extraction from the LH and HL sub- bands using first order statistics, and final classification by support vector machines (SVM). The proposed approach shows higher performance than when using features extracted from the LL sub-band. It is concluded that the horizontal and vertical sub- bands of the wavelet transform can effectively encode the discriminating features of normal and pathological images.

Detecting and fixing defects make programs easier to understand by developers. We propose an automated approach for the detection and correction of various types of design defects in source code. Our approach allows to automatically find detection rules, thus relieving the designer from doing so manually. Rules are defined as combinations of metrics/thresholds that better conform to known instances of design defects (defect examples). The correction solutions, a combination of refactoring operations, should minimize, as much as possible, the number of defects detected using the detection rules. In our setting, we use genetic programming for rule extraction. For the correction step, we use genetic algorithm. We evaluate our approach by finding and fixing potential defects in four open-source systems. For all these systems, we found, in average, more than 80% of known defects, a better result when compared to a state-of-the-art approach, where the detection rules are manually or semi-automatically specified. The proposed corrections fix, in average, more than 78%of detected defects.

Building efficient machine-learning assessment models is an important achievement of empirical software engineering research. Their integration in automated decision-making systems is one of the objectives of this work. It aims at empirically verify the relationships between some software internal artifacts and two quality attributes: maintainability and reusability. Several algorithms, belonging to various machine-learning approaches, are selected and run on software data collected from medium size applications. Some of these approaches produce models with very high quantitative performances; others give interpretable and "glass-box" models that are very complementary.

This work studies the use of Particle Swarm Optimization (PSO) as a classification technique. Beyond assessing classification accuracy, it investigates the following questions: does PSO present limitations for high dimensional application domains? Is it less efficient for multi class problems? To answer the questions, an experimental set up was realized that uses three high dimensional data sets. Our results are that, depending on the mechanisms controlling confinement and dispersion in the PSO algorithm, the classification accuracy varied with the dimensionality of the data and the cardinality of the output space.

This work studies the use of Particle Swarm Optimization (PSO) as a classification technique for a fluorescence measurements substance database. Beyond assessing classification accuracy, it investigates the following questions: does PSO present limitations for high dimensional application domains? Is it less efficient for multi class problems? To answer the questions, an experimental set up was realized leading to interesting conclusions.

Most bidirectional associative memory (BAM) networks use a symmetrical output function for dual fixed-point behavior. In this paper, we show that by introducing an asymmetry parameter into a recently introduced chaotic BAM output function, prior knowledge can be used to momentarily disable desired attractors from memory, hence biasing the search space to improve recall performance. This property allows control of chaotic wandering, favoring given subspaces over others. In addition, reinforcement learning can then enable a dual BAM architecture to store and recall nonlinearly separable patterns. Our results allow the same BAM framework to model three different types of learning: supervised, reinforcement, and unsupervised. This ability is very promising from the cognitive modeling viewpoint. The new BAM model is also useful from an engineering perspective; our simulations results reveal a notable overall increase in BAM learning and recall performances when using a hybrid model with the general regression neural network (GRNN).

This paper addresses important performance issues in delay-line-based timing applications targeting FPGA devices. The circuit under test is a TDC circuit implemented on a low-cost FPGA from XILINX. Various performance limitations such as uncertainty and non-uniformity in cell delays are described and corresponding optimization and improvement suggestions are made. Experimental results were obtained using ring oscillator-based test structures to inspect intra-die delay mismatches along the target FPGA's surface.

An optoelectronic front end for a high-speed solid-state based time-domain fluorescence measurement system is investigated. It consists of a source light emitting diode (LED) and photodetector (PD) pair connected to an original transimpedance amplifier. The amount of light reaching the detector from the fluorescence and the effect of different noise sources on the proposed system are evaluated. The simulation results show that a bandwidth greater than 80MHz could be reached for PD currents with amplitudes as low as 20nA. An actual circuit was built to confirm the projected capabilities of the system.

Bidirectional associative memories (BAMs) have been widely used for auto and heteroassociative learning. However, few research efforts have addressed the issue of multistep vector pattern recognition. We propose a model that can perform multi step pattern recognition without the need for a special learning algorithm, and with the capacity to learn more than two pattern series in the training set. The model can also learn pattern series of different lengths and, contrarily to previous models, the stimuli can be composed of gray-level images. The paper also shows that by adding an extra autoassociative layer, the model can accomplish one-to-many association, a task that was exclusive to feedforward networks with context units and error backpropagation learning.

Typical bidirectional associative memories (BAM) use an offline, one-shot learning rule, have poor memory storage capacity, are sensitive to noise, and are subject to spurious steady states during recall. Recent work on BAM has improved network performance in relation to noisy recall and the number of spurious attractors, but at the cost of an increase in BAM complexity. In all cases, the networks can only recall bipolar stimuli and, thus, are of limited use for grey-level pattern recall. In this paper, we introduce a new bidirectional heteroassociative memory model that uses a simple self-convergent iterative learning rule and a new nonlinear output function. As a result, the model can learn online without being subject to overlearning. Our simulation results show that this new model causes fewer spurious attractors when compared to others popular BAM networks, for a comparable performance in terms of tolerance to noise and storage capacity. In addition, the novel output function enables it to learn and recall grey-level patterns in a bidirectional way.

Autoassociative memories are known for their capacity to learn correlated patterns, complete these patterns and, once the learning phase completed, filter noisy inputs. However, no autoassociative memory as of yet was able to learn noisy patterns without preprocessing or special procedure. In this paper, we show that a new unsupervised learning rule enables associative memory models to locally learn online noisy correlated patterns. The learning is carried out by a dual Hebbian rule and the convergence is asymptotic. The asymptotic convergence results in an unequal eigenvalues spectrum, which distinguish SCRAM from optimal linear associative memories (OLAMs). Therefore, SCRAM develops less spurious attractors and has better recall performance under noise degradation.

Current object-oriented (OO) software systems must satisfy new requirements that include quality aspects. These, contrary to functional requirements, are difficult to determine during the test phase of a project. Predictive and estimation models offer an interesting solution to this problem. This paper describes an original approach to build rule-based predictive models that are based on fuzzy logic and that enhance the performance of classical decision trees. The approach also attempts to bridge the cognitive gap that may exist between the antecedent and the consequent of a rule by turning the latter into a chain of sub rules that account for domain knowledge. The whole framework is evaluated on a set of OO applications.

Managing the evolution of an OO system constitutes a complex and resource-consuming task. This is particularly true for reusable class libraries since the user interface must be preserved for version compatibility. Thus, the symptomatic detection of potential instabilities during the design phase of such libraries may help avoid later problems. This paper introduces a fuzzy logic-based approach for evaluating the stability of a reusable class library interface, using structural metrics as stability indicators. To evaluate this new approach, we conducted a preliminary study on a set of commercial C++ class libraries. The obtained results are very promising when compared to those of two classical machine learning approaches, top down induction of decision trees and Bayesian classifiers.

Many sources agree that managing the evolution of an OO system constitutes a complex and resource-consuming task. This is particularly true for reusable class libraries, as the user interface must be preserved to allow for version compatibility. Thus, the symptomatic detection of potential instabilities during the design phase of such libraries may serve to avoid later problems. This paper presents a fuzzy logic-based approach for evaluating the interface stability of a reusable class library, by using structural metrics as stability indicators.