This paper describes a Deep Belief Neural Network (DBNN) and Bidirectional Long-Short Term Memory (LSTM) hybrid used as an acoustic model for Speech Recognition. It was demonstrated by many independent researchers that DBNNs exhibit superior performance to other known machine learning frameworks in terms of speech recognition accuracy. Their superiority comes from the fact that these are deep learning networks. However, a trained DBNN is simply a feed-forward network with no internal memory, unlike Recurrent Neural Networks (RNNs) which are Turing complete and do posses internal memory, thus allowing them to make use of longer context. In this paper, an experiment is performed to make a hybrid of a DBNN with an advanced bidirectional RNN used to process its output. Results show that the use of the new DBNN-BLSTM hybrid as the acoustic model for the Large Vocabulary Continuous Speech Recognition (LVCSR) increases word recognition accuracy. However, the new model has many parameters and in some cases it may suffer performance issues in real-time applications.
Skin cancer is the most common form of cancer affecting humans. Melanoma is the most dangerous type of skin cancer; and early diagnosis is extremely vital in curing the disease. So far, the human knowledge in this field is very limited, thus, developing a mechanism capable of identifying the disease early on can save lives, reduce intervention and cut unnecessary costs. In this paper, the researchers developed a new learning technique to classify skin lesions, with the purpose of observing and identifying the presence of melanoma. This new technique is based on a convolutional neural network solution with multiple configurations; where the researchers employed an International Skin Imaging Collaboration (ISIC) dataset. Optimal results are achieved through a convolutional neural network composed of 14 layers. This proposed system can successfully and reliably predict the correct classification of dermoscopic lesions with 97.78% accuracy.
The use of fly ash as a material for earth structures involves its proper compaction. Fly ash compaction tests have to be conducted on separately prepared virgin samples because spherical ash grains are crushed during compaction, so the laboratory compaction procedure is time-consuming and laborious. The aim of the study was to determine the neural models for prediction of fly ash compaction curve shapes. The attempt of applying the artificial neural networks type MLP was made. ANN inputs were new-created variables – principal components dependent on grain-size distribution (as D₁₀–D₉₀ and uniformity and curvature coefficients), compaction method, and fly ash specific density. The output vectors were presented by co-ordinates of generated compaction curve points. Each point (wᵢ, ρdi) was described by two independent ANNs. Using ANN-based modelling method, models which enable establishing the approximate compaction curve shape were obtained.
Various sectors of the economy such as transport and renewable energy have shown great interest in sea bed models. The required measurements are usually carried out by ship-based echo sounding, but this method is quite expensive. A relatively new alternative is data obtained by airborne lidar bathymetry. This study investigates the accuracy of these data, which was obtained in the context of the project ‘Investigation on the use of airborne laser bathymetry in hydrographic surveying’. A comparison to multi-beam echo sounding data shows only small differences in the depths values of the data sets. The IHO requirements of the total horizontal and vertical uncertainty for laser data are met. The second goal of this paper is to compare three spatial interpolation methods, namely Inverse Distance Weighting (IDW), Delaunay Triangulation (TIN), and supervised Artificial Neural Networks (ANN), for the generation of sea bed models. The focus of our investigation is on the amount of required sampling points. This is analyzed by manually reducing the data sets. We found that the three techniques have a similar performance almost independently of the amount of sampling data in our test area. However, ANN are more stable when using a very small subset of points.
In this paper the authors propose a decision support system for automatic blood smear analysis based on microscopic images. The images are pre-processed in order to remove irrelevant elements and to enhance the most important ones – the healthy blood cells (erythrocytes) and the pathologic ones (echinocytes). The separated blood cells are analysed in terms of their most important features by the eigenfaces method. The features are the basis for designing the neural network classifier, learned to distinguish between erythrocytes and echinocytes. As the result, the proposed system is able to analyse the smear blood images in a fully automatic way and to deliver information on the number and statistics of the red blood cells, both healthy and pathologic. The system was examined in two case studies, involving the canine and human blood, and then consulted with the experienced medicine specialists. The accuracy of classification of red blood cells into erythrocytes and echinocytes reaches 96%.
An array consisting of four commercial gas sensors with target specifications for hydrocarbons, ammonia, alcohol, explosive gases has been constructed and tested. The sensors in the array operate in the dynamic mode upon the temperature modulation from 350°C to 500°C. Changes in the sensor operating temperature lead to distinct resistance responses affected by the gas type, its concentration and the humidity level. The measurements are performed upon various hydrogen (17-3000 ppm), methane (167-3000 ppm) and propane (167-3000 ppm) concentrations at relative humidity levels of 0-75%RH. The measured dynamic response signals are further processed with the Discrete Fourier Transform. Absolute values of the dc component and the first five harmonics of each sensor are analysed by a feed-forward back-propagation neural network. The ultimate aim of this research is to achieve a reliable hydrogen detection despite an interference of the humidity and residual gases.
A simple analog circuit is presented which can play a neuron role in static-model-based neural networks implemented in the form of an integrated circuit. Operating in a transresistance mode it is suited to cooperate with transconductance synapses. As a result, its input signal is a current which is a sum of currents coming from the synapses. Summation of the currents is realized in a node at the neuron input. The circuit has two outputs and provides a step function signal at one output and a linear function one at the other. Activation threshold of the step output can be conveniently controlled by means of a voltage. Having two outputs, the neuron is attractive to be used in networks taking advantage of fuzzy logic. It is built of only five MOS transistors, can operate with very low supply voltages, consumes a very low power when processing the input signals, and no power in the absence of input signals. Simulation as well as experimental results are shown to be in a good agreement with theoretical predictions. The presented results concern a 0.35 1m CMOS process and a prototype fabricated in the framework of Europractice.
In order to make the analog fault classification more accurate, we present a method based on the Support Vector Machines Classifier (SVC) with wavelet packet decomposition (WPD) as a preprocessor. In this paper, the conventional one-against-rest SVC is resorted to perform a multi-class classification task because this classifier is simple in terms of training and testing. However, this SVC needs all decision functions to classify the query sample. In our study, this classifier is improved to make the fault classification task more fast and efficient. Also, in order to reduce the size of the feature samples, the wavelet packet analysis is employed. In our investigations, the wavelet analysis can be used as a tool of feature extractor or noise filter and this preprocessor can improve the fault classification resolution of the analog circuits. Moreover, our investigation illustrates that the SVC can be applicable to the domain of analog fault classification and this novel classifier can be viewed as an alternative for the back-propagation (BP) neural network classifier.
The paper focuses on the problem of robust fault detection using analytical methods and soft computing. Taking into account the model-based approach to Fault Detection and Isolation (FDI), possible applications of analytical models, and first of all observers with unknown inputs, are considered. The main objective is to show how to employ the bounded-error approach to determine the uncertainty of soft computing models (neural networks and neuro-fuzzy networks). It is shown that based on soft computing models uncertainty defined as a confidence range for the model output, adaptive thresholds can be described. The paper contains a numerical example that illustrates the effectiveness of the proposed approach for increasing the reliability of fault detection. A comprehensive simulation study regarding the DAMADICS benchmark problem is performed in the final part.
The paper deals with the application of the feed-forward and cascade-forward neural networks to mechanical state variable estimation of the drive system with elastic coupling. The learning procedure of neural estimators is described and the influence of the input vector size and neural network structure to the accuracy of state variable estimation is investigated. The quality of state estimation by neural estimators of different types is tested and compared. The simple optimisation procedure is proposed. Optimised neural estimators of the torsional torque and the load machine speed are tested in the open-loop and closed-loop control structure of the drive system with elastic joint, with additional feedbacks from the shaft torque and the difference between the motor and the load speeds. It is shown that torsional vibrations of the two-mass system are damped effectively using the closed-loop control structure with additional feedbacks obtained from the developed neural estimators. The simulation results are confirmed by laboratory experiments.
Laughter is one of the most important paralinguistic events, and it has specific roles in human conversation. The automatic detection of laughter occurrences in human speech can aid automatic speech recognition systems as well as some paralinguistic tasks such as emotion detection. In this study we apply Deep Neural Networks (DNN) for laughter detection, as this technology is nowadays considered state-of-the-art in similar tasks like phoneme identification. We carry out our experiments using two corpora containing spontaneous speech in two languages (Hungarian and English). Also, as we find it reasonable that not all frequency regions are required for efficient laughter detection, we will perform feature selection to find the sufficient feature subset.
The literature on exchange rate forecasting is vast. Many researchers have tested whether implications of theoretical economic models or the use of advanced econometric techniques can help explain future movements in exchange rates. The results of the empirical studies for major world currencies show that forecasts from a naive random walk tend to be comparable or even better than forecasts from more sophisticated models. In the case of the Polish zloty, the discussion in the literature on exchange rate forecasting is scarce. This article fills this gap by testing whether non-linear time series models are able to generate forecasts for the nominal exchange rate of the Polish zloty that are more accurate than forecasts from a random walk. Our results confirm the main findings from the literature, namely that it is difficult to outperform a naive random walk in exchange rate forecasting contest.
Presented are results of a research on the possibility of using artificial neural networks for forecasting mechanical and technological parameters of moulding sands containing water-glass, hardened in the innovative microwave heating process. Trial predictions were confronted with experimental results of examining sandmixes prepared on the base of high-silica sand, containing various grades of sodium water-glass and additions of a wetting agent. It was found on the grounds of obtained values of tensile strength and permeability that, with use of artificial neural networks, it is possible complex forecasting mechanical and technological properties of these materials after microwave heating and the obtained data will be used in further research works on application of modern analytic methods for designing production technology of high-quality casting cores and moulds.
The purpose of the work was to predict the selected product parameters of the dry separation process using a pneumatic sorter. From the perspective of application of coal for energy purposes, determination of process parameters of the output as: ash content, moisture content, sulfur content, calorific value is essential. Prediction was carried out using chosen machine learning algorithms that proved to be effective in forecasting output of various technological processes in which the relationships between process parameters are non-linear. The source of data used in the work were experiments of dry separation of coal samples. Multiple linear regression was used as the baseline predictive technique. The results showed that in the case of predicting moisture and sulfur content this technique was sufficient. The more complex machine learning algorithms like support vector machine (SVM) and multilayer perceptron neural network (MPL) were used and analyzed in the case of ash content and calorific value. In addition, k-means clustering technique was applied. The role of cluster analysis was to obtain additional information about coal samples used as feed material. The combination of techniques such as multilayer perceptron neural network (MPL) or support vector machine (SVM) with k-means allowed for the development of a hybrid algorithm. This approach has significantly increased the effectiveness of the predictive models and proved to be a useful tool in the modeling of the coal enrichment process.
The article is devoted to the problem of voice signals recognition means introduction in the system of distance learning. The results of the conducted research determine the prospects of neural network means of phoneme recognition. It is also shown that the main difficulties of creation of the neural network model, intended for recognition of phonemes in the system of distance learning, are connected with the uncertain duration of a phoneme-like element. Due to this reason for recognition of phonemes, it is impossible to use the most effective type of neural network model on the basis of a multilayered perceptron, at which the number of input parameters is a fixed value. To mitigate this shortcoming, the procedure, allowing to transform the non-stationary digitized voice signal to the fixed quantity of mel-cepstral coefficients, which are the basis for calculation of input parameters of the neural network model, is developed. In contrast to the known ones, the possibility of linear scaling of phoneme-like elements is available in the procedure. The number of computer experiments confirmed expediency of the fact that the use of the offered coding procedure of input parameters provides the acceptable accuracy of neural network recognition of phonemes under near-natural conditions of the distance learning system. Moreover, the prospects of further research in the field of development of neural network means of phoneme recognition of a voice signal in the system of distance learning is connected with an increase in admissible noise level. Besides, the adaptation of the offered procedure to various natural languages, as well as to other applied tasks, for instance, a problem of biometric authentication in the banking sector, is also of great interest.
Creep compliance of the hot-mix asphalt (HMA) is a primary input of the current pavement thermal cracking prediction model used in the US. This paper discusses a process of training an Artificial Neural Network (ANN) to correlate the creep compliance values obtained from the Indirect Tension (IDT) with similar values obtained on small HMA beams from the Bending Beam Rheometer (BBR). In addition, ANNs are also trained to predict HMA creep compliance from the creep compliance of asphalt binder and vice versa using the BBR setup. All trained ANNs exhibited a very high correlation of 97 to 99 percent between predicted and measured values. The binder creep compliance functions built on the ANN-predicted discrete values also exhibited a good correlation when compared with the laboratory experiments. However, the simulation of trained ANNs on the independent dataset produced a significant deviation from the measured values which was most likely caused by the differences in material composition, such as aggregate type and gradation, presence of recycled additives, and binder type.
Due to an increasing amount of music being made available in digital form in the Internet, an automatic organization of music is sought. The paper presents an approach to graphical representation of mood of songs based on Self-Organizing Maps. Parameters describing mood of music are proposed and calculated and then analyzed employing correlation with mood dimensions based on the Multidimensional Scaling. A map is created in which music excerpts with similar mood are organized next to each other on the two-dimensional display.
There were two aims of the research. One was to enable more or less automatic confirmation of the known associations – either quantitative or qualitative – between technological data and selected properties of concrete materials. Even more important is the second aim – demonstration of expected possibility of automatic identification of new such relationships, not yet recognized by civil engineers. The relationships are to be obtained by methods of Artificial Intelligence, (AI), and are to be based on actual results from experiments on concrete materials. The reason of applying the AI tools is that in Civil Engineering the real data are typically non perfect, complex, fuzzy, often with missing details, which means that their analysis in a traditional way, by building empirical models, is hardly possible or at least can not be done quickly. The main idea of the proposed approach was to combine application of different AI methods in a one system, aimed at estimation, prediction, design and/or optimization of composite materials. The paradigm of the approach is that the unknown rules concerning the properties of concrete are hidden in experimental results and can be obtained from the analysis of examples. Different AI techniques like artificial neural networks, machine learning and certain techniques related to statistics were applied. The data for the analysis originated from direct observations and from reports and publications on concrete technology. Among others it has been demonstrated that by combining different AI methods it is possible to improve the quality of the data, (e.g. when encountering outliers and missing values or in clustering problems), so that the whole data processing system will be giving better prediction, (when applying ANNs), or the newly discovered rules will be more effective, (e.g. with descriptions more complete and – at the same time – possibly more consistent, in case of ML algorithms).