Descubre nuestras contribuciones a la comunidad científica internacional a través de publicaciones en revistas de alto impacto y conferencias prestigiosas.
Thermal comfort is the condition in which a person feels satisfaction with the thermal environment through a subjective evaluation. In this work, a compact and efficient estimation of thermal comfort perception by human subjects is performed using a constructive neurocomputational model trained with data generated in controlled conditions with 49 volunteers giving 705 different scenarios, allowing, thanks to the versatility of the model, an interpretable and simple resulting function facilitating an easy handling of the results by people from different fields. The results have been compared with two of the most used standard methods for modelling thermal comfort: Fanger and COMFA models, and they show an improvement in terms of accuracy and mean square error both in a binary decision scenario (comfort or not) as well as for a discrete decision-making case in which different thermal comfort regions are considered. The flexibility of the neural model permits the incorporation of extra subject-related variables that increases further the thermal comfort estimation and, also, permits the implementation of the model in distributed and low cost/low consumption systems.
Background and objectives: There is an increasing and renewed interest in Electronic Health Records (EHRs) as a substantial information source for clinical decision making. Consequently, automatic de-identification of EHRs is an indispensable task, since their dissociation from personal data is a necessary prerequisite for their dissemination. Nevertheless, the bulk of prior research in this domain has been conducted using English EHRs, given the limited availability of annotated corpora in other languages, including Spanish. Methods: In this study, the automatic de-identification of medical documents in Spanish was explored. A private corpus comprising 599 genuine clinical cases was annotated with eight different categories of protected health information. The prediction problem was approached as a named entity recognition task and two deep learning-based methodologies were developed. The first strategy was based on recurrent neural networks (RNN) and the second, an end-to-end approach, was based on Transformers. In addition, we have implemented a procedure to expand the amount of texts employed for model training. Results: Our findings demonstrate that Transformers surpass RNNs in the de-identification of clinical data in Spanish. Particularly noteworthy is the excellent performance of the XLM-RoBERTa large Transformer, achieving a rigorous strict-match micro-average of 0.946 for precision, 0.954 for recall, and an F1 score of 0.95 when applied to the amplified version of the corpus. Furthermore, a web-based application has been created to assist specialized clinicians in de-identifying EHRs through the aid of the implemented models. Conclusion: The study's conclusions showcase the practical applicability of the state-of-the-art Transformers models for precise de-identification of clinical notes in real-world medical settings in Spanish, with the potential to improve performance if continual pre-training strategies are implemented.
Bib numbers are used in mass competitions to identify participants, especially in long-distance races where runners commonly wear tags to verify that they pass mandatory checkpoints. In this paper, we delve deeper into the use of existing computer vision techniques for recognizing the digits present in bib numbers. Our analysis of bib recognition involves evaluating OCRs (Optical Character Recognition) techniques and a YOLOv7 digit detector on two public datasets: RBNR and TGCRBNW. The results reveal that the former scenario is solvable, while the latter presents extremely in-the-wild challenges. However, the findings suggest that more than relying solely on RBN for runner identification, other appearance-based cues, e.g., clothing and accessories, may be required due to various circumstances, such as occlusion or incomplete bib recognition. In any case, all those cues do not necessarily imply that the same person is wearing the RBN across the competition track, as they are not biome tric traits.
Most of the clinical information stored in Spanish healthcare systems is found as unstructured text in electronic medical records. The automatic extraction of valuable information contained in these documents is a critical task. Valuable information for clinical analysis units in oncology includes the location of a patient's neoplasm. This location, included in the ICD-10-ES coding category, can be extracted from the texts using natural language processing. To this end, in this study we have developed methodologies based on the state of the art in natural language processing, the Transformer models. The results obtained show that the application of these models is of great help in this task. In particular, the RoBERTa-Base-Biomed model performed best, with a value of 0.946 in percentage of correct answers, 0.920 in precision, 0.898 in sensitivity and 0.908 in F1-score, showing great performance for most classes.
Breast cancer is a major public health problem, with 2.3M new cases diagnosed each year. Immunotherapy is an effective treatment for breast cancer depending on several factors like subtype of tumours or associated prognosis. However, the immune system’s efficiency depends on the local microenvironment and requires region-specific trials with a reduced number of samples. To minimise this drawback and improve the accuracy of patient prognosis predictions, we explore several data augmentation methods, i.e. noise injection, oversampling techniques and generative adversarial networks. The experiment was conducted through a set of immune system gene expression samples donated by 165 breast cancer patients from the Málaga region. Results showed a 5% increase in AUC and a 23- 36% increase in F1 score for subtype prediction.
Color quantization (CQ) is one of the most common and important procedures to be performed on digital images. In this paper, a new approach to hierarchical color quantization is described, presenting a novel neural network architecture integrated by a convolutional autoencoder and a Growing Hierarchical Bregman Neural Gas (GHBNG). GHBNG is a CQ algorithm that allows the compression of an image by choosing a reduced set of the most representative colors to generate a high-quality reproduction of the original image. In the technique proposed here, an autoencoder is used to translate the image into a latent representation with higher per-pixel dimensionality but reduced resolution, and GHBNG is then used to quantize it. Experimental results confirm the performance of this technique and its suitability for tasks related to color quantization.
A growing and renewed interest has emerged in Electronic Health Records (EHRs) as a source of information for decision-making in clinical practice. In this context, the automatic de-identification of EHRs constitutes an essential task, since their dissociation from personal data is a mandatory first step before their distribution. However, the majority of previous studies on this subject have been conducted on English EHRs, due to the limited availability of annotated corpora in other languages, such as Spanish. In this study, we addressed the automatic de-identification of medical documents in Spanish. A private corpus of 599 real-world clinical cases have been annotated with 8 different protected health information categories. We have tackled the predictive problem as a named entity recognition task, developing two different deep learning-based methodologies, namely a first strategy based on recurrent neural networks (RNN) and an end-to-end approach based on transformers. Additionally, we have developed a data augmentation procedure to increase the number of texts used to train the models. The results obtained show that transformers outperform RNN on the de-identification of Spanish clinical data. In particular, the best performance was obtained by the XLM-RoBERTa large transformer, with a strict-match micro-averaged value of 0.946 for precision, 0.954 for recall and 0.95 for F1-score, when trained on the augmented version of the corpus. The performance achieved by transformers in this study proves the viability of applying these state-of-the-art models in real-world clinical scenarios.
Healthcare systems currently store a large amount of clinical data, mostly unstructured textual information, such as electronic health records (EHRs). Manually extracting valuable information from these documents is costly for healthcare professionals. For example, when a patient first arrives at an oncology clinical analysis unit, clinical staff must extract information about the type of neoplasm in order to assign the appropriate clinical specialist. Automating this task is equivalent to text classification in natural language processing (NLP). In this study, we have attempted to extract the neoplasm type by processing Spanish clinical documents. A private corpus of 23, 704 real clinical cases has been processed to extract the three most common types of neoplasms in the Spanish territory: breast, lung and colorectal neoplasms. We have developed methodologies based on state-of-the-art text classification task, strategies based on machine learning and bag-of-words, based on embedding models in a supervised task, and based on bidirectional recurrent neural networks with convolutional layers (C-BiRNN). The results obtained show that the application of NLP methods is extremely helpful in performing the task of neoplasm type extraction. In particular, the 2-BiGRU model with convolutional layer and pre-trained fastText embedding obtained the best performance, with a macro-average, more representative than the micro-average due to the unbalanced data, of 0.981 for precision, 0.984 for recall and 0.982 for F1-score.
En la última década, el aprendizaje profundo (DL) se ha impuesto como el enfoque de inteligencia artificial con mayor progresión y éxito. El DL conforma el estado del arte en visión por computador y procesamiento del lenguaje natural, mostrando además un potencial prometedor en bioinformática, un campo de gran impacto económico y social. Sin embargo, estos modelos presentan una importante desventaja, requieren de miles de instancias de datos para lograr un buen nivel de éxito. Actualmente en bioinformática, la adquisición de datos sigue siendo un proceso difícil y costoso, especialmente trabajando con conjuntos genómicos, expresión molecular o metabolómica. Estos son significativamente difíciles de obtener, y su escasez es acuciante en estudios de enfermedades raras o regiones geográficas concretas. Para resolver esta limitación se puede emplear el aumento de datos (DA), el incremento del número de muestras disponibles mediante transformaciones o generación. En los últimos años, modelos de DA pertenecientes al DL han obtenido un rendimiento asombroso en generación de imágenes sintéticas. Sin embargo, aplicar estos modelos a conjuntos bioinformáticos sin información espacial o temporal es desafiante. A este respecto, el objetivo de la tesis doctoral es el desarrollo de métodos de DA y su aplicación en problemas bioinformáticos no estructurados. Se desarrollaron métodos de DA basados en adición de ruido gaussiano, ajuste específico de ruido, adaptación de modelos generativos profundos, y ajuste de un meta-clasificador. Los métodos desarrollados se aplicaron en tres problemas: predicción de eventos en cáncer mediante RNA-Seq; predicción de la enfermedad rara Niemann-Pick Tipo-C a partir de datos metabolómicos; y clasificación del subtipo molecular de cáncer de mama utilizando la expresión génica del sistema inmune. De las investigaciones se concluye el potencial del DA para generar muestras que replican información biomédica y conducen a un aumento en el rendimiento de predicción.
Categorical data obtained from real-world domains are commonly imbalanced, as they often present more number of samples belonging to one of the classes. Imbalanced data tends to be a problem for classifiers, as the majority class biased them and affects overall performance. Among the techniques used for dealing with imbalanced data sets, data augmentation (DA) constitutes an alternative, as it can improve the accuracy of prediction for the minority class (usually the relevant one), but usually at the cost of a loss regarding predictions of the majority one. To benefit from both behaviours, we introduce in this study a meta-classifier scheme that works as a mixture of two classifiers, one trained with the original data and the second one trained using augmented data. The experiments carried out with 12 imbalanced data sets, 5 of them obtained from the TCGA database related to cancer survival prediction, show an improvement in accuracy, area under the ROC curve and Matthews correlation coefficient values compared to the results obtained using the original data sets.
Niemann–Pick Class 1 (NPC1) disease is a rare and debilitating neurodegenerative lysosomal storage disease (LSD). Metabolomics datasets of NPC1 patients available to perform this type of analysis are often limited in the number of samples and severely unbalanced. In order to improve the predictive capability and identify new biomarkers in an NPC1 disease urinary dataset, data augmentation (DA) techniques based on computational intelligence have been employed to create synthetic samples, i.e. the addition of noise, oversampling techniques and conditional generative adversarial networks. These techniques have been used to evaluate their predictive capacities on a set of urine samples donated by 13 untreated NPC1 disease and 47 heterozygous (parental) carrier control participants. Results on the prediction have also been obtained using different machine learning classification models and the partial least squares techniques. These results provide strong evidence for the ability of DA techniques to generate good quality synthetic data. Results acquired show increases in sensitivity of 20%–50%, an F1 score of 6%–30%, and a predictive capacity of 0.3 (out of 1). Additionally, more conventional forms of multivariate data analysis have been employed. These have allowed the detection of unusual urinary metabolite profiles, and the identification of biomarkers through the use of synthetically augmented datasets. Results indicate that urinary branched-chain amino acids such as valine, 3-aminoisobutyrate and quinolinate, may be employable as valuable biomarkers for the diagnosis and prognostic monitoring of NPC1 disease.
Within the area of bioinformatics, Deep Learning (DL) models have shown exceptional results in applications in which histological images, scans and tomographies are used. However, when gene expression data are used, the performance often does not reach the expected results. The reason is that these datasets commonly have a high dimensionality and a low number of samples. To improve results in this type of data, Data Augmentation (DA) techniques can be used. DA techniques are methods that can generate synthetic samples from original data to increase the size of the dataset. In this work, three different DA techniques have been developed and tested on six different cancer datasets. Results show that DA techniques can improve classification results with significant improvements in sensitivity, specificity and F1-score when applied to cancer gene expression datasets.
Niemann-Pick Class 1 (NPC1) disease is a rare and neurodegenerative disease, and often metabolomics datasets of NPC1 patients are limited in the number of samples and severely imbalanced. In order to improve the predictive capability and identify new biomarkers in an NPC1 disease urinary dataset, data augmentation (DA) techniques based on computational intelligence are employed to create additional synthetic samples. This paper presents DA techniques, based on the addition of noise, on oversampling techniques and using conditional generative adversarial networks, to evaluate their predictive capacities on a set of Nuclear Magnetic Resonance (NMR) profiles of urine samples. Prediction results obtained show increases in sensitivity (30%) and in F score (20%). In addition, multivariate data analysis and variable importance in projection scores have been applied. These analyses show the ability of the DA methods to replicate the information of the metabolites and determined that selected metabolites (such as 3-aminoisobutyrate, 3-hidroxivaleric, quinolinate and trimethylamine) may be valuable biomarkers for the diagnosis of NPC1 disease.
In this paper, a new color quantization method based on a self-organized artificial neural network called the Growing Hierarchical Bregman Neural Gas (GHBNG) is proposed. This neural network is based on Bregman divergences, from which the squared Euclidean distance is a particular case. Thus, the best suitable Bregman divergence for color quantization can be selected according to the input data. Moreover, the GHBNG yields a tree-structured model that represents the input data so that a hierarchical color quantization can be obtained, where each layer of the hierarchy contains a different color quantization of the original image. Experimental results confirm the color quantization capabilities of this approach.
By the lack of personalization in education, students obtain low performance in different subjects in school, particularly in mathematics. Therefore, learning style identification is a crucial tool to improve academic performance. Although traditional methods such questionnaires have been extensively used to the learning styles detection in youths and adults by its high precision, it produces boredom in children and does not allow to adjust learning automatically to student characteristics and preferences over time. In this paper, two methods for learning style recognition: CHAEA-Junior questionnaire (static method) and Artificial Neural Networks (automatic method) are explored. The data for the second technique used answers from the survey and the percentage scores from mathematical mini-games (Competitor, Dreamer, Logician, Strategist) based on Kolb’s learning theory. To the validity between both methods, it was conducted a pilot study with primary level students in Ecuador. The experimental tests show that Artificial Neural Networks are a suitable alternative to accurate models for automatic learning recognition to provide personalized learning to Ecuadorian students, which achieved close detection results concerning CHAEA-Junior questionnaire results.
An Autoencoder is an artificial neural network used for unsupervised learning and for dimensionality reduction. In this work, an Autoencoder has been used to encrypt and decrypt digital information. So, it is implemented to code and decode characters represented in an 8-bit format, which corresponds to the size of ASCII representation. The Back-propagation algorithm has been used in order to perform the learning process with two different variant depends on when the discretization procedure is carried out, during (model I) or after (model II) the learning phase. Several tests were conducted to determine the best Autoencoder architectures to encrypt and decrypt, taking into account that a good encrypt method corresponds to a process that generate a new code with uniqueness and a good decrypt method successfully recovers the input data. A network that obtains a 100% in the two process is considered a good digital cryptography implementation. Some of the proposed architecture obtain a 1 00% in the processes to encrypt 52 ASCII characters (Letter characters) and 95 ASCII characters (printable characters), recovering all the data.
C-Mantec neural network constructive algorithm Ortega (C-Mantec neural network algorithm implementation on MATLAB. https://github.com/IvanGGomez/CmantecPaco, 2015) creates very compact architectures with generalization capabilities similar to feed-forward networks trained by the well-known back-propagation algorithm. Nevertheless, constructive algorithms suffer much from the problem of overfitting, and thus, in this work the learning procedure is first analyzed for networks created by this algorithm with the aim of trying to understand the training dynamics that will permit optimization possibilities. Secondly, several optimization strategies are analyzed for the position of class separating hyperplanes, and the results analyzed on a set of public domain benchmark data sets. The results indicate that with these modifications a small increase in prediction accuracy of C-Mantec can be obtained but in general this was not better when compared to a standard support vector machine, except in some cases when a mixed strategy is used.
In this paper, a new self-organizing artificial neural network called growing hierarchical neural forest (GHNF) is proposed. The GHNF is a hierarchical model based on the growing neural forest, which is a tree-based model that learns a set of trees (forest) instead of a general graph so that the forest can grow in size. This way, the GHNF faces three important limitations regarding the self-organizing map: fixed size, fixed topology, and lack of hierarchical representation for input data. Hence, the GHNF is especially amenable to datasets containing clusters where each cluster has a hierarchical structure since each tree of the GHNF forest can adapt to one of the clusters. Experimental results show the goodness of our proposal in terms of self-organization and clustering capabilities. In particular, it has been applied to text mining of tweets as a typical exploratory data analysis application, where a hierarchical representation of concepts present in tweets has been obtained. Moreover, it has been applied to foreground detection in video sequences, outperforming several methods specialized in foreground detection.
Data augmentation (DA) is a key element in the success of Deep Learning (DL) models, as its use can lead to better prediction accuracy values when large size data sets are used. DA was not very much used with earlier neural network models before 2012, and the reason might be related to the type of models and the size of the data sets used. We investigate in this work, applying several state-of-the-art models based on Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs), the effect of DA when using small size data sets, analyzing the results in terms of the prediction accuracy obtained according to the different characteristics of the training samples (number of instances and features, and class unbalance degree). We further introduce modifications to the standard methods used to generate the synthetic samples to alter the class balance representation, and the overall results indicate that with some computational effort a significant increase in prediction accuracy can be obtained when small data sets are considered.
Learning math is important for the academic life of students: the development of mathematical skills is influenced by different characteristics of students such as geographical position, economic level, parents’ education, achievement level, teacher objectives, social level, use of information and communication technologies by teachers, learner motivation, gender, age, preferences for playing video games, and the school year of the students. In this work, these previously mentioned characteristics were considered as the attributes (inputs) of a multilayer neural network that uses a backpropagation algorithm to predict the percentage of improvement in mathematics through a 2D mathematical video game that was developed to allow the children to practice addition and subtraction operations. After applying the neural model, using the twelve attributes mentioned before and the backpropagation algorithm, there was a network of one layer with ten neurons and another network of two layers wit h 5 neurons in the first layer and 20 neurons in the second layer. Both architectures produced a mean squared error smaller than 0.0069 in the prediction of the student’s percentage of improvement in mathematics, being the best configurations found in this study for the neural model. These results lead to the conclusion that we are able to predict the percentage of improvement in math that the students could achieve after playing the game, and therefore, claiming if the video game is recommendable or not for certain students.
In Evolutionary Artificial Neural Networks (EANN), evolutionary algorithms are used to give an additional alternative to adapt besides learning, specially for connection weights training and architecture design, among others. A type of EANNs known as Topology and Weight Evolving Artificial Neural Networks (TWEANN) are used to evolve topology and weights. In this work, we introduce a new encoding on an implementation of NeuroEvolution of Augmenting Topologies (NEAT), a type of TWEANN, by adopting the Red-Black Tree (RBT) as the main data structure to store the connection genes instead of using a list. This new version of NEAT efficacy was tested using as case of study some data sets from the UCI database. The accuracy of networks obtained through the new version of NEAT were compared with the accuracy obtained from feed-forward artificial neural networks trained using back-propagation. These comparisons yielded that the accuracy were similar, and in some cases the accuracy obtained by the new version were better. Also, as the number of patterns increases, the average number of generations increases exponentially. Finall
Since the origins of artificial neural network research, many models of feedforward networks have been proposed. This paper presents an algorithm which adapts the shape of the activation function to the training data, so that it is learned along with the connection weights. The activation function is interpreted as a piecewise polynomial approximation to the distribution function of the argument of the activation function. An online learning procedure is given, and it is formally proved that it makes the training error decrease or stay the same except for extreme cases. Moreover, the model is computationally simpler than standard feedforward networks, so that it is suitable for implementation on FPGAs and microcontrollers. However, our present proposal is limited to two-layer, one-output-neuron architectures due to the lack of differentiability of the learned activation functions with respect to the node locations. Experimental results are provided, which show the performance of the proposal algorithm for classification and regression applications.
The main objective of this work has been to implement a model to find equidistant real roots using a Successive Adaptive Linear Neural Modeling which uses two approaches: a Self Organized Map (SOM) and an Adaptative Linear Neuron (Adaline). A SOM model has been used with a new neighborhood function Λ, and a physical distance β with which the task is divided in sub-processes reducing the complexity of the task because the SOM model can delimited the areas where a single root exist. Then, through a successive approach, it is applied an Feed-forward neural model with a learning process base on Adaline neuron with pocket in each pair of regions for finding the real root values with a reduced precision. Finally, several experiments were done consider CPU time, relative error, distance between the roots and polynomial degrees. The results show that the time complexity grows in a linear or logarithmic way. Also, the error does not increase in a higher rate than the degree of polynomial or the root distance.
Defining a relationship among companies that belong to the BMEX group in order to provide investors with information, is a help to minimize the risk at the moment of investing. Using data taken from Yahoo finances and BMEX website, a SOM neural network was used to study the daily data of all companies, which belong to the IBEX35 and the Latibex indexes. Companies which are part of the IBEX35, and appear closer between them in the SOM mesh, were compared in profitable terms showing that eminently there exist economic and business line relationship between them. The opposite happened companies selected randomly from the IBEX35 group in some specific cases. Likewise, the companies of Latibex group, were joined to IBEX35 companies to compare a entire year evolution between them. In fact, demonstrating that the model proposes definitely find and shows associations between companies that are near enough, and which belongs to a determinant index in a stock market environment.
A portable an distributed expert system has been implemented in order to carry out voice and speech recognition of different people. The device consists of an open source microcontroller connected to a microphone that would be able to instantly recognize words and people, and could be distributed as a real time device in the security field. Two different methods of artificial neural network have been performed to implement the system: the first one has been the Continuous Hopfield Network to store and recover patterns previously stored, in order to know the word used to communicate with the system, the second one has been the Backpropagation Algorithm used to recognize between different speakers. However prior to performing these methods, it has been necessary to pre-process the information so that it can be computed in this type of device, transforming sound files into a set of useful data using a Fast Fourier Transform. The results obtained have been satisfactory with an accuracy above 80% for words recovery with a white noise of 40% of the signal and above 80% for recognizing a person with a new words introduced.
Most clustering algorithms are designed to minimize a distortion measure which quantifies how far the elements of the clusters are from their respective centroids. The assessment of the results is often carried out with the help of cluster quality measures which take into account the compactness and separation of the clusters. However, these measures are not amenable to optimization because they are not differentiable with respect to the centroids even for a given set of clusters. Here we propose a differentiable cluster quality measure, and an associated clustering algorithm to optimize it. It turns out that the standard k-means algorithm is a special case of our method. Experimental results are reported with both synthetic and real datasets, which demonstrate the performance of our approach with respect to several standard quantitative measures.
Data augmentation has been proven particularly effective for image classification tasks where a significant boost of prediction accuracy can be obtained when the technique is combined with the use of Deep Learning architectures. Unfortunately, for non-image data the situation is quite different and the positive effect of augmenting the training set size is much smaller. In this work, we propose a method that creates new samples by adjusting the level of noise for individual input variables previously ranked by their relevance level. Results from several tests are analyzed using nine benchmark data sets when the augmented and original data are used for supervised training on Deep Learning architectures.
Scheduling problems are very hard computational tasks with several applications in multitude of domains. In this work we solve a practical problem motivated by a real industry situation, in which we apply a genetic algorithm for finding an acceptable solution in a very short time interval. The main novelty introduced in this work is the use of a priority based chromosome codification that determines the precedence of a task with respect to other ones, permitting to introduce in a very simple way all problem constraints, including setup costs and workforce availability. Results show the suitability of the approach, obtaining real time solutions for tasks with up to 50 products.
Training of large scale neural networks, like those used nowadays in Deep Learning schemes, requires long computational times or the use of high performance computation solutions like those based on cluster computation, GPU boards, etc. As a possible alternative, in this work the Back-Propagation learning algorithm is implemented in an FPGA board using a multiplexing layer scheme, in which a single layer of neurons is physically implemented in parallel but can be reused any number of times in order to simulate multi-layer architectures. An on-chip implementation of the algorithm is carried out using a training/validation scheme in order to avoid overfitting effects. The hardware implementation is tested on several configurations, permitting to simulate architectures comprising up to 127 hidden layers with a maximum number of neurons in each layer of 60 neurons. We confirmed the correct implementation of the algorithm and compared the computational times against C and Matlab code executed in a multicore supercomputer, observing a clear advantage of the proposed FPGA scheme. The layer multiplexing scheme used provides a simple and flexible approach in comparison to standard implementations of the Back-Propagation algorithm representing an important step towards the FPGA implementation of deep neural networks, one of the most novel and successful existing models for prediction problems.
Recent advances in FPGA technology have permitted the implementation of neurocomputational models, making them an interesting alternative to standard PCs in order to speed up the computations involved taking advantage of the intrinsic FPGA parallelism. In this work, we analyse and compare the FPGA implementation of two neural network learning algorithms: the standard and well known Back-Propagation algorithm and C-Mantec, a constructive neural network algorithm that generates compact one hidden layer architectures with good predictive capabilities. One of the main differences between both algorithms is the fact that while Back-Propagation needs a predefined architecture, C-Mantec constructs its network while learning the input patterns. Several aspects of the FPGA implementation of both algorithms are analyzed, focusing in features like logic and memory resources needed, transfer function implementation, computation time, etc. The advantages and disadvantages of both methods in relationship to their hardware implementations are discussed.
In recent years predictive models based on Deep Learning strategies have achieved enormous success in several domains including pattern recognition tasks, language translation, software design, etc. Deep learning uses a combination of techniques to achieve its prediction accuracy, but essentially all existing approaches are based on multi-layer neural networks with deep architectures, i.e., several layers of processing units containing a large number of neurons. As the simulation of large networks requires heavy computational power, GPUs and cluster based computation strategies have been successfully used. In this work, a layer multiplexing scheme is presented in order to permit the simulation of deep neural networks in FPGA boards. As a demonstration of the usefulness of the scheme deep architectures trained by the classical Back-Propagation algorithm are simulated on FPGA boards and compared to standard implementations, showing the advantages in computation speed of the proposed scheme.
Thermal comfort conditions are important for the normal development of human tasks, and as such they have been analyzed in the context of several areas including human physiology, ergonomics, heating and cooling systems, architectural design, etc. In this work, we analyze the estimation of the thermal comfort perception by human subjects using a neurocomputational model based on the C-Mantec constructive neural network architecture, comparing it with two standard methods for modeling thermal comfort: Fanger and COMFA models. The results indicate a significative advantage of C-Mantec in terms of the predictive accuracy obtained, consider also that the flexibility of the neural model would permit the introduction of extra variables that can increase further the thermal comfort estimation.
The well-known backpropagation learning algorithm is implemented in a field-programmable gate array (FPGA) board and a microcontroller, focusing in obtaining efficient implementations in terms of a resource usage and computational speed. The algorithm was implemented in both cases using a training/validation/testing scheme in order to avoid overfitting problems. For the case of the FPGA implementation, a new neuron representation that reduces drastically the resource usage was introduced by combining the input and first hidden layer units in a single module. Further, a time-division multiplexing scheme was implemented for carrying out product computations taking advantage of the built-in digital signal processor cores. In both implementations, the floating-point data type representation normally used in a personal computer (PC) has been changed to a more efficient one based on a fixed-point scheme, reducing system memory variable usage and leading to an increase in computation speed. The results show that the modifications proposed produced a clear increase in computation speed in comparison with the standard PC-based implementation, demonstrating the usefulness of the intrinsic parallelism of FPGAs in neurocomputational tasks and the suitability of both implementations of the algorithm for its application to the real world problems.
A two-dimensional Ising model with nearest-neighbors ferromagnetic interactions is implemented in a Field Programmable Gate Array (FPGA) board. Extensive Monte Carlo simulations were carried out using an efficient hardware representation of individual spins and a combined global-local LFSR random number generator. Consistent results regarding the descriptive properties of magnetic systems, like energy, magnetization and susceptibility are obtained while a speed-up factor of approximately six times is achieved in comparison to previous FPGA-based published works and almost 104 times in comparison to a standard CPU simulation. A detailed description of the logic design used is given together with a careful analysis of the quality of the random number generator used. The obtained results confirm the potential of FPGAs for analyzing the statistical mechanics of magnetic systems.
Most current approaches to computer vision are based on expensive, high performance hardware to meet the heavy computational requirements of the employed algorithms. These system architectures are severely limited in their practical application due to financial and technical limitations. In this work a different strategy is used, namely the development of an inexpensive and easy to deploy computer vision system for motion detection. This is achieved by three means. First of all, an affordable and flexible hardware platform is employed. Secondly, the motion detection algorithm is specifically tailored to involve a very small computational load. Thirdly, a fixed point programming paradigm is followed in implementing the system so as to further reduce the computational requirements. The proposed system is experimentally compared to the standard motion detector for a wide range of benchmark videos. The reported results indicate that our proposal attains substantially better performance, while it remains affordable and easy to install in practice.
Este trabajo se presenta como un estudio sobre la relación que existe entre el sistema inmunitario y el cáncer de mama. Con este propósito se van a utilizar diversos algoritmos y modelos de aprendizaje computacional. Aumentar la capacidad del sistema inmunitario frente a las células cancerígenas es uno de los temas actuales de más interés, ya sea por la competencia que otorgaría para prevenir la formación de distintos tipos de cáncer, como por la posibilidad de experimentar con técnicas más eficaces de inmunoterapia, terapias basadas en la modificación del sistema inmunitario y menos agresivas que la radioterapia o la quimioterapia. Los datos que se van a usar para este propósito, presentan la expresión genética del sistema inmunitario de pacientes de la región de Málaga afectadas por cáncer de mama y el subtipo molecular del cáncer que presentan. Se pretende pues observar en qué medida se podría intentar predecir el subtipo del cáncer, a partir de la expresión genética mediante la aplicación de distintos modelos de aprendizaje supervisado, con modelos clásicos como los árboles de decisión o las redes neuronales artificiales, y con modelos más actuales como bagging, boosting o el modelo LASSO. Mediante este análisis también se intenta encontrar algún tipo de correlación entre las variables observadas dentro del conjunto, ayudado además de un análisis con métodos de aprendizaje no supervisado. Los resultados obtenidos nos indican que existe una correlación patente entre ellos, además de que hay posibles firmas genéticas en el conjunto de datos. Tras el análisis, los receptores CTLA y NKG2D, y los genes LMP y TAP, parecen ser las variables moleculares más influyentes en el subtipo molecular del cáncer.
Recent advances in FPGA technology have permitted the implementation of neurocomputational models, making them an interesting alternative to standard PCs in order to speed up the computations involved taking advantage of the intrinsic FPGA parallelism. In this work, we analyse and compare the FPGA implementation of two neural network learning algorithms: the standard Back-Propagation algorithm and C-Mantec, a constructive neural network algorithm that generates compact one hidden layer architectures. One of the main differences between both algorithms is the fact that while Back-Propagation needs a predefined architecture, C-Mantec constructs its network while learning the input patterns. Several aspects of the FPGA implementation of both algorithms are analysed, focusing in features like logic and memory resources needed, transfer function implementation, computation time, etc. Advantages and disadvantages of both methods are discussed in the context of their application to benchmark problems.
En la actualidad, existen un gran número de investigaciones que usan técnicas de aprendizaje automático basadas en árboles de decisión. Como evolución de dichos trabajos, se han desarrollado métodos que usan Multiclasificadores (Random forest, Boosting, Bagging) que resuelven los mismos problemas abordados con árboles de decisión simples, aumentando el porcentaje de acierto. El ámbito de los problemas resueltos tradicionalmente por dichas técnicas es muy variado aunque destaca la bio-informática. En cualquier caso, la clasificación siempre puede ser consultada a un experto considerándose su respuesta como correcta. Existen problemas donde un experto en la materia no siempre acierta. Un ejemplo, pueden ser las quinielas (1X2). Donde podemos observar que un conocimiento del dominio del problema aumenta el porcentaje de aciertos, sin embargo, predecir un resultado erróneo es muy posible. El motivo es que el número de factores que influyen en un resultado es tan grande que, en muchas ocasiones, convierten la predicción en un acto de azar. En este trabajo pretendemos encontrar un multiclasificador basado en los clasificadores simples más estudiados como pueden ser el Perceptrón Multicapa o Árboles de Decisión con el porcentaje de aciertos más alto posible. Con tal fin, se van a estudiar e implementar una serie de configuraciones de clasificadores propios junto a multiclasificadores desarrollados por terceros. Otra línea de estudio son los propios datos, es decir, el conjunto de entrenamiento. Mediante un estudio del dominio del problema añadiremos nuevos atributos que enriquecen la información que disponemos de cada resultado intentando imitar el conocimiento en el que se basa un experto. Los desarrollos descritos se han realizado en R. Además, se ha realizado una aplicación que permite entrenar un multiclasificador (bien de los propios o bien de los desarrollados por terceros) y como resultado obtenemos la matriz de confusión junto al porcentaje de aciertos. En cuanto a resultados, obtenemos porcentajes de aciertos entre el 50% y el 55%. Por encima del azar y próximos a los resultados de los expertos.
Este trabajo presenta la implementación en FPGA de dos algoritmos de aprendizaje de redes neuronales. Se analizan las consideraciones de diseño y se comparan las implementaciones en términos de rendimiento y utilización de recursos hardware.
This work presents a high precision FPGA implementation of neural network activation functions. The approach achieves superior numerical accuracy while maintaining computational efficiency, addressing the challenge of implementing non-linear functions in fixed-point hardware.
Competitive majority network trained by error correction (C-Mantec), a recently proposed constructive neural network algorithm that generates very compact architectures with good generalization capabilities, is implemented in a field programmable gate array (FPGA). A clear difference with most of the existing neural network implementations (most of them based on the use of the backpropagation algorithm) is that the C-Mantec automatically generates an adequate neural architecture while the training of the data is performed. All the steps involved in the implementation, including the on-chip learning phase, are fully described and a deep analysis of the results is carried on using the two sets of benchmark problems. The results show a clear increase in the computation speed in comparison to the standard personal computer (PC)-based implementation, demonstrating the usefulness of the intrinsic parallelism of FPGAs in the neurocomputational tasks and the suitability of the hardware version of the C-Mantec algorithm for its application to real-world problems.
The techniques currently developed for updating software in sensor nodes located in changing environments require usually the use of reprogramming procedures, which clearly increments the costs in terms of time and energy consumption. This work presents an alternative to the traditional reprogramming approach based on an on-chip learning scheme in order to adapt the node behaviour to the environment conditions. The proposed learning scheme is based on C-Mantec, a novel constructive neural network algorithm especially suitable for microcontroller implementations as it generates very compact size architectures. The Arduino UNO board was selected to implement this learning algorithm as it is a popular, economic and efficient open source single-board microcontroller. C-Mantec has been successfully implemented in a microcontroller board by adapting it in order to overcome the limitations imposed by the limited resources of memory and computing speed of the hardware device. Also, this work brings an in-depth analysis of the solutions adopted to overcome hardware resource limitations in the learning algorithm implementation (e.g., data type), together with an efficiency assessment of this approach when the algorithm is tested on a set of circuit design benchmark functions. Finally, the utility, efficiency and versatility of the system is tested in three different-nature case studies in which the environmental conditions change its behaviour over time.
A recently proposed constructive neural network algorithm, named C-Mantec, is fully implemented in a Arduino board. The C-Mantec algorithm generates very compact size neural architectures with good prediction abilities, and thus the board can be potentially used to learn on-site sensed data without needing to transmit information to a central control unit. An analysis of the more difficult steps of the implementation is detailed, and a test is carried out on a set of benchmark functions normally used in circuit design to show the correct functioning of the implementation.
This paper introduces Committee C-Mantec, a probabilistic constructive neural network approach that combines ensemble methods with constructive learning algorithms. The method demonstrates improved generalization and robustness compared to single network approaches.