[ES] La sociedad moderna se ha visto afectada por los acelerados avances de la tecnología. La aplicación de la inteligencia artificial se puede encontrar en todas partes, desde la televisión inteligente hasta los coches autónomos. Una tarea esencial del aprendizaje automático es la clasificación. A pesar de la cantidad de técnicas y algoritmos de clasificación que existen, es un campo que sigue siendo relevante por todas sus aplicaciones. Así, frente a la clasificación tradicional multiclase en la que a cada instancia se le asigna una única etiqueta de clase, se han propuesto otros métodos como la clasificación jerárquica y la clasificación multietiqueta. Esta tesis tiene como objetivo resolver la clasificación multiclase mediante una descomposición jerárquica. Asimismo, se exploran diferentes métodos de extender la aproximación definida para su aplicación en contextos cambiantes. La clasificación jerárquica es una tarea de aprendizaje automático en la que el problema de clasificación original se divide en pequeños subproblemas. Esta división se realiza teniendo en cuenta una estructura jerárquica que representa las relaciones entre las clases objetivo. Como resultado el clasificador jerárquico es a su vez una estructura (un árbol o un grafo) compuesta por clasificadores de base. Hasta ahora, en la literatura, la clasificación jerárquica se ha aplicado a dominios jerárquicos, independientemente que la estructura jerárquica sea proporcionada explícitamente o se asume implícita (en cuyo caso se hace necesario inferir primero dicha estructura jerárquica). La clasificación jerárquica ha demostrado un mejor rendimiento en dominios jerárquicos en comparación con la clasificación plana (que no tiene en cuenta la estructura jerárquica del dominio). En esta tesis, proponemos resolver los problemas de clasificación multiclase descomponiéndolo jerárquicamente de acuerdo a una jerarquía de clases inferida por un clasificador plano. Planteamos dos escenarios dependiendo del tipo de clasificador usado en la jerarquía de clasificadores: clasificadores duros (crisp) y clasificadores suaves (soft). Por otra parte, un problema de clasificación puede sufrir cambios una vez los modelos han sido entrenados. Un cambio frecuente es la aparición de una nueva clase objetivo. Dado que los clasificadores no han sido entrenados con datos pertenecientes a la nueva clase, no podrán encontrar predicciones correctas para las nuevas instancias, lo que afectará negativamente en el rendimiento de los clasificadores. Este problema se puede resolver mediante dos alternativas: el reentrenamiento de todo el modelo o la adaptación del modelo para responder a esta nueva situación. Como parte del estudio de los algoritmos de clasificación jerárquica se presentan varios métodos para adaptar el modelo a los cambios en las clases objetivo. Los métodos y aproximaciones definidas en la tesis se han evaluado experimentalmente con una amplia colección de conjuntos de datos que presentan diferentes características, usando diferentes técnicas de aprendizaje para generar los clasificadores de base. En general, los resultados muestran que los métodos propuestos pueden ser una alternativa a métodos tradicionales y otras técnicas presentadas en la literatura para abordar las situaciones específicas planteadas. / [CA] La societat moderna s'ha vist afectada pels accelerats avenços de la tecnologia. L'aplicació de la intel·ligència artificial es pot trobar a tot arreu, des de la televisió intel·ligent fins als cotxes autònoms. Una tasca essencial de l'aprenentatge automàtic és la classificació. Tot i la quantitat de tècniques i algoritmes de classificació que existeixen, és un camp que segueix sent rellevant per totes les seves aplicacions. Així, enfront de la classificació tradicional multiclase en la qual a cada instància se li assigna una única etiqueta de classe, s'han proposat altres mètodes com la classificació jeràrquica i la classificació multietiqueta. Aquesta tesi té com a objectiu resoldre la classificació multiclase mitjançant una descomposició jeràrquica. Així mateix, s'exploren diferents mètodes d'estendre l'aproximació definida per a la seva aplicació en contextos canviants. La classificació jeràrquica és una tasca d'aprenentatge automàtic en la qual el problema de classificació original es divideix en petits subproblemes. Aquesta divisió es realitza tenint en compte una estructura jeràrquica que representa les relacions entre les classes objectiu. Com a resultat el classificador jeràrquic és al seu torn una estructura (un arbre o un graf) composta per classificadors de base. Fins ara, en la literatura, la classificació jeràrquica s'ha aplicat a dominis jeràrquics, independentment que l'estructura jeràrquica sigui proporcionada explícitament o s'assumeix implícita (en aquest cas es fa necessari inferir primer aquesta estructura jeràrquica). La classificació jeràrquica ha demostrat un millor rendiment en dominis jeràrquics en comparació amb la classificació plana (que no té en compte l'estructura jeràrquica de l'domini). En aquesta tesi, proposem resoldre els problemes de classificació multiclasse descomponent jeràrquicament d'acord a una jerarquia de classes inferida per un classificador pla. Plantegem dos escenaris depenent de el tipus de classificador usat en la jerarquia de classificadors: classificadors durs (crisp) i classificadors suaus (soft). D'altra banda, un problema de classificació pot patir canvis una vegada els models han estat entrenats. Un canvi freqüent és l'aparició d'una nova classe objectiu. Atès que els classificadors no han estat entrenats amb dades pertanyents a la nova classe, no podran trobar prediccions correctes per a les noves instàncies, el que afectarà negativament en el rendiment dels classificadors. Aquest problema es pot resoldre mitjançant dues alternatives: el reentrenament de tot el model o l'adaptació de el model per respondre a aquesta nova situació. Com a part de l'estudi dels algoritmes de classificació jeràrquica es presenten diversos mètodes per adaptar el model als canvis en les classes objectiu. Els mètodes i aproximacions definides en la tesi s'han avaluat experimentalment amb una àmplia col·lecció de conjunts de dades que presenten diferents característiques, usant diferents tècniques d'aprenentatge per generar els classificadors de base. En general, els resultats mostren que els mètodes proposats poden ser una alternativa a mètodes tradicionals i altres tècniques presentades en la literatura per abordar les situacions específiques plantejades. / [EN] The modern society has been affected by rapid advances in technology. The application of artificial intelligence can be found everywhere, from intelligent television to autonomous cars. An essential task of machine learning is classification. Despite the number of classification techniques and algorithms that exist, it is a field that remains relevant for all its applications. Thus, as opposed to the traditional multiclass classification in which each instance is assigned a single class label, other methods such as hierarchical classification and multi-label classification have been proposed. This thesis aims to solve multiclass classification by means of a hierarchical decomposition. Also, different methods of extending the defined approach are explored for application in changing contexts. Hierarchical classification is an automatic learning task in which the original classification problem is divided into small sub-problems. This division is made taking into account a hierarchical structure that represents the relationships between the target classes. As a result the hierarchical classifier is itself a structure (a tree or a graph) composed of base classifiers. Up to now, in the literature, hierarchical classification has been applied to hierarchical domains, regardless of whether the hierarchical structure is explicitly provided or assumed to be implicit (in which case it becomes necessary to first infer the hierarchical structure). Hierarchical classification has demonstrated better performance in hierarchical domains compared to flat classification (which does not take into account the hierarchical structure of the domain). In this thesis, we propose to solve the problems of multiclass classification by breaking it down hierarchically according to a class hierarchy inferred by a plane classifier. We propose two scenarios depending on the type of classifier used in the classifier hierarchy: hard classifiers (crisp) and soft classifiers (soft). On the other hand, a classification problem may change once the models have been trained. A frequent change is the appearance of a new target class. Since the existing classifiers have not been trained with data belonging to the new class, they will not be able to find correct predictions for the new instances, which will negatively affect the performance of the classifiers. This problem can be solved by two alternatives: retraining the entire model or adapting the model to respond to this new situation. As part of the study of hierarchical classification algorithms, several methods are presented to adapt the model to changes in target classes. The methods and approaches defined in the thesis have been evaluated experimentally with a large collection of data sets that have different characteristics, using different learning techniques to generate the base classifiers. In general, the results show that the proposed methods can be an alternative to traditional methods and other techniques presented in the literature to address the specific situations raised. / Silva Palacios, DA. (2021). Clasificación Jerárquica Multiclase [Tesis doctoral]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/167015 / TESIS

The purpose of this study was to investigate how metadata from Spotify could be used to identify the language of songs in a dataset containing nine languages. Features based on song name, album name, genre, regional popularity and vectors describing songs, playlists and users were analysed individually and in combination with each other in different classifiers. In addition to this, this report explored how different levels of prediction confidence affects performance and how it compared to a classifier based on audio input. A random forest classifier proved to have the best performance with an accuracy of 95.4% for the whole data set. Performance was also investigated when the confidence of the model was taken into account, and when only keeping more confident predictions from the model, accuracy was higher. When keeping the 70% most confident predictions an accuracy of 99.4% was achieved. The model also proved to be robust to input of other languages than it was trained on, and managed to filter out unwanted records not matching the languages of the model. A comparison was made to a classifier based on audio input, where the model using metadata performed better on the training and test set used. Finally, a number of possible improvements and future work were suggested.

<p>Connected and autonomous vehicle (CAV) technologies provide disruptive and transformational opportunities for innovations toward intelligent transportation systems. Compared with human driven vehicles (HDVs), the CAVs can reduce reaction time and human errors, increase traffic mobility and will be more knowledgeable due to vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communication. CAVs’ potential to reduce traffic accidents, improve vehicular mobility and promote eco-driving is immense. However, the new characteristics and capabilities of CAVs will significantly transform the future of transportation, including the dissemination of traffic information, traffic flow dynamics and network equilibrium flow. This dissertation seeks to realize and enhance the application of CAVs by specifically advancing the research in three connected topics: (1) modeling and controlling information flow propagation within a V2V communication environment, (2) designing a real-time deployable cooperative control mechanism for CAV platoons, and (3) modeling network equilibrium flow with a mix of CAVs and HDVs. </p> <p>Vehicular traffic congestion in a V2V communication environment can lead to congestion effects for information flow propagation due to full occupation of the communication channel. Such congestion effects can impact not only whether a specific information packet of interest is able to reach a desired location, but also the timeliness needed to influence traffic system performance. This dissertation begins with exploring spatiotemporal information flow propagation under information congestion effects, by introducing a two-layer macroscopic model and an information packet relay control strategy. The upper layer models the information dissemination in the information flow regime, and the lower layer model captures the impacts of traffic flow dynamics on information propagation. Analytical and numerical solutions of the information flow propagation wave (IFPW) speed are provided, and the density of informed vehicles is derived under different traffic conditions. Hence, the proposed model can be leveraged to develop a new generation of information dissemination strategies focused on enabling specific V2V information to reach specific locations at specific points in time.</p> <p>In a V2V-based system, multiclass information (e.g., safety information, routing information, work zone information) needs to be disseminated simultaneously. The application needs of different classes of information related to vehicular reception ratio, the time delay and spatial coverage (i.e., distance it can be propagated) are different. To meet the application needs of multiclass information under different traffic and communication environments, a queuing strategy is proposed for each equipped vehicle to disseminate the received information. It enables control of multiclass information flow propagation through two parameters: 1) the number of communication servers and 2) the communication service rate. A two-layer model is derived to characterize the IFPW under the designed queuing strategy. Analytical and numerical solutions are derived to investigate the effects of the two control parameters on information propagation performance in different information classes. </p> <p>Third, this dissertation also develops a real-time implementable cooperative control mechanism for CAV platoons. Recently, model predictive control (MPC)-based platooning strategies have been developed for CAVs to enhance traffic performance by enabling cooperation among vehicles in the platoon. However, they are not deployable in practice as they require anembedded optimal control problem to be solved instantaneously, with platoon size and prediction horizon duration compounding the intractability. Ignoring the computational requirements leads to control delays that can deteriorate platoon performance and cause collisions between vehicles. To address this critical gap, this dissertation first proposes an idealized MPC-based cooperative control strategy for CAV platooning based on the strong assumption that the problem can be solved instantaneously. It then develops a deployable model predictive control with first-order approximation (DMPC-FOA) that can accurately estimate the optimal control decisions of the idealized MPC strategy without entailing control delay. Application of the DMPC-FOA approach for a CAV platoon using real-world leading vehicle trajectory data shows that it can dampen the traffic oscillation effectively, and can lead to smooth deceleration and acceleration behavior of all following vehicles.</p> <p>Finally, this dissertation also develops a multiclass traffic assignment model for mixed traffic flow of CAVs and HDVs. Due to the advantages of CAVs over HDVs, such as reduced value of time, enhanced quality of travel experience, and seamless situational awareness and connectivity, CAV users can differ in their route choice behavior compared to HDV users, leading to mixed traffic flows that can significantly deviate from the single-class HDV traffic pattern. However, due to a lack of quantitative models, there is limited knowledge on the evolution of mixed traffic flows in a traffic network. To partly bridge this gap, this dissertation proposes a multiclass traffic assignment model. The multiclass model captures the effect of knowledge level of traffic conditions on route choice of both CAVs and HDVs. In addition, it captures the characteristics of mixed traffic flow such as the difference in value of time between HDVs and CAVs and the asymmetry in their driving interactions, thereby enhancing behavioral realism in the modeling. New solution algorithms will be developed to solve the multiclass traffic assignment model. The study results can assist transportation decision-makers to design effective planning and operational strategies to leverage the advantages of CAVs and manage traffic congestion under mixed traffic flows.</p> <p>This dissertation deepens our understanding of the characteristics and phenomena in domains of traffic information dissemination, traffic flow dynamics and network equilibrium flow in the age of connected and autonomous transportation. The findings of this dissertation can assist transportation managers in designing effective traffic operation and planning strategies to fully exploit the potential of CAVs to improve system performance related to traffic safety, mobility and energy consumption. </p>

The average cost of operating a queueing network depends on several factors such as the complexity of the network and the service policy used. Approximate linear programming (ALP) is a method that can be used to compute an accurate lower bound on the optimal average cost as well as generate policies to be used in operating the network. These average cost solutions and policies are dependent on the type of basis function used in the ALP. In this paper, the ALP average cost solutions and policies are analyzed for twelve networks with four different types of basis functions (quadratic, linear, pure exponential, and mixed exponential). An approximate bound on the optimality gap between the ALP average cost solution and the optimal average cost solution is computed for each system, and the size of this bound is determined relative to the ALP average cost solution. Using the same set of networks, the performance of ALP generated policies are compared to the performance of the heuristic policies first-buffer-first-served (FBFS), last-buffer-first-served (LBFS), highest-queue-first-served (HQFS), and random-queue-first-served (RQFS). In general, ALP generated average cost solutions are considerably smaller than the simulated average cost under the corresponding policy, and therefore the approximate bounds on the optimality gaps are quite large. This bound increases with the complexity of the queueing network. Some ALP generated policies are not stabilizing policies for their corresponding networks, especially those produced using pure exponential and mixed exponential basis functions. For almost all systems, at least one of the heuristic policies results in mean average cost less than or nearly equal to the smallest mean average cost of all ALP generated policies in simulation runs. This means that generally there exists a heuristic policy which can perform as well as or better than any ALP generated policy. In conclusion, a useful bound on the optimality gap between the ALP average cost solution and the optimal average cost solution cannot be computed with this method. Further, heuristic policies, which are more computationally tractable than ALP generated policies, can generally match or exceed the performance of ALP generated policies, and thus computing such policies is often unnecessary for realizing cost benefits in queueing networks. / text

abstract: Deep learning architectures have been widely explored in computer vision and have depicted commendable performance in a variety of applications. A fundamental challenge in training deep networks is the requirement of large amounts of labeled training data. While gathering large quantities of unlabeled data is cheap and easy, annotating the data is an expensive process in terms of time, labor and human expertise. Thus, developing algorithms that minimize the human effort in training deep models is of immense practical importance. Active learning algorithms automatically identify salient and exemplar samples from large amounts of unlabeled data and can augment maximal information to supervised learning models, thereby reducing the human annotation effort in training machine learning models. The goal of this dissertation is to fuse ideas from deep learning and active learning and design novel deep active learning algorithms. The proposed learning methodologies explore diverse label spaces to solve different computer vision applications. Three major contributions have emerged from this work; (i) a deep active framework for multi-class image classication, (ii) a deep active model with and without label correlation for multi-label image classi- cation and (iii) a deep active paradigm for regression. Extensive empirical studies on a variety of multi-class, multi-label and regression vision datasets corroborate the potential of the proposed methods for real-world applications. Additional contributions include: (i) a multimodal emotion database consisting of recordings of facial expressions, body gestures, vocal expressions and physiological signals of actors enacting various emotions, (ii) four multimodal deep belief network models and (iii) an in-depth analysis of the effect of transfer of multimodal emotion features between source and target networks on classification accuracy and training time. These related contributions help comprehend the challenges involved in training deep learning models and motivate the main goal of this dissertation. / Dissertation/Thesis / Doctoral Dissertation Electrical Engineering 2018

Submitted by Geandra Rodrigues (geandrar@gmail.com) on 2018-01-09T13:48:15Z No. of bitstreams: 1 renanpiazzarolifinottiamaral.pdf: 1172046 bytes, checksum: eb7bf10c813d64fbddcc572d39aecfc5 (MD5) / Approved for entry into archive by Adriana Oliveira (adriana.oliveira@ufjf.edu.br) on 2018-01-22T16:10:30Z (GMT) No. of bitstreams: 1 renanpiazzarolifinottiamaral.pdf: 1172046 bytes, checksum: eb7bf10c813d64fbddcc572d39aecfc5 (MD5) / Made available in DSpace on 2018-01-22T16:10:30Z (GMT). No. of bitstreams: 1 renanpiazzarolifinottiamaral.pdf: 1172046 bytes, checksum: eb7bf10c813d64fbddcc572d39aecfc5 (MD5) Previous issue date: 2017-09-01 / - / This thesis presents and discusses improvements in the type-1 and singleton fuzzy logic system for dealing with classification problems. Two training methods are addressed, the scaled conjugate gradient, which uses the second order information approximating the multiplication of the Hessian matrix H by the directional vector v (i.e. Hv), and the same method using the differential operator R {.} to compute the exact value of Hv. Also, in order to adapt the fuzzy model to handle multiclass classification problems, it is developed a novel fuzzy model with a vector as output. All proposals are tested through the performance metrics analysis based on data sets provided by UCI Machine Learning Repository. The reported results show the high convergence speed and better classification rates of the proposed training methods than others presented in the literature. Additionally, the novel fuzzy model has a significant reduction in computational and classifier complexity, especially when the number of classes in classification problem increases.

Today, we are living in a data-driven world. Due to a surge in data generation, there is a need for efficient and accurate techniques to analyze data. One such kind of data which is needed to be analyzed are text reviews given for movies. Rather than classifying the reviews as positive or negative, we will classify the sentiment of the reviews on the scale of one to ten. In doing so, we will compare two recurrent neural network algorithms Long short term memory(LSTM) and Gated recurrent unit(GRU). The main objective of this study is to compare the accuracies of LSTM and GRU models. For training models, we collected data from two different sources. For filtering data, we used porter stemming and stop words. We coupled LSTM and GRU with the convolutional neural networks to increase the performance. After conducting experiments, we have observed that LSTM performed better in predicting border values. Whereas, GRU predicted every class equally. Overall GRU was able to predict multiclass text data of movie reviews slightly better than LSTM. GRU was computationally expansive when compared to LSTM.

The selection of hospital once an ambulance has picked up its patient is today decided by the ambulance staff. This report describes a supervised machinelearning approach for predicting hospital selection. This is a multi-classclassification problem. The performance of random forest, logistic regression and neural network were compared to each other and to a baseline, namely the one rule-algorithm. The algorithms were applied to real world data from SOS-alarm, the company that operate Sweden’s emergency call services. Performance was measured with accuracy and f1-score. Random Forest got the best result followed by neural network. Logistic regression exhibited slightly inferior results but still performed far better than the baseline. The results point toward machine learning being a suitable method for learning the problem of hospital selection. / Beslutet om till vilket sjukhus en ambulans ska köra patienten till bestäms idag av ambulanspersonalen. Den här rapporten beskriver användandet av övervakad maskininlärning för att förutsåga detta beslut. Resultaten från algoritmerna slumpmässig skog, logistisk regression och neurala nätvärk jämförs med varanda och mot ett basvärde. Basvärdet erhölls med algorithmen en-regel. Algoritmerna applicerades på verklig data från SOS-alarm, Sveriges operatör för larmsamtal. Resultaten mättes med noggrannhet och f1-poäng. Slumpmässigskog visade bäst resultat följt av neurala nätverk. Logistisk regression uppvisade något sämre resultat men var fortfarande betydligt bättre än basvärdet. Resultaten pekar mot att det är lämpligt att använda maskininlärning för att lära sig att ta beslut om val av sjukhus.

Over the last decade, neural networks have reached almost every field of science and technology. They have become a crucial part of various real-world applications, such as medical imaging. Still, their deployment in safety-critical applications remains limited owing to their inability to provide reliable uncertainty estimates and frequently occurring overconfident predictions, which is normally the case in modern neural networks possessing a substantial number of layers. In this thesis, we leverage the capability of data mining algorithms like density clustering to explain the behavior of a medical image classifier responsible for classifying white blood cells. We know that any clustering algorithm acts on the feature vector of the input data and annotates the data into different clusters as per the features. In this work, we lay down and prove the hypothesis that the output discrete probability matrix of a multi-class classification problem can be used as a feature vector where the confidence value of every class can be considered as a degree of resemblance with that class. Before implementing clustering, one needs to make sure that these confidence values represent actual probabilities so that they can be used as features; hence certain calibration techniques were incorporated to improve the calibration of the network first. Having a better calibrated medical classifier, density clustering was implemented, which generated results that provided solid arguments to justify the behavior of the network. As far as the use case of this method is concerned, it was observed that we could identify pathologies like myelodysplastic syndromes, acute lymphocytic leukemia, and chronic myelomonocytic leukemia in a patient. This was possible due to the presence of the same class of White blood cells in multiple clusters indicating the presence of subpopulations separated into healthy and pathological cells of the same class depending upon the pathology that needs to be detected. This was proved visually by mapping cluster points to actual cell images and quantitatively as well by using entropy as a method of quantifying uncertainty. This method showed that there is a lot of information embedded in the output probability matrix. Hence one can employ various data mining techniques to extract more information and not just limit themselves to misclassifications and confusion matrices. / Under det senaste decenniet har neurala nätverk nått nästan alla områden inom vetenskap och teknik. De har blivit en avgörande del av olika verkliga tillämpningar, såsom medicinsk bildbehandling. Ändå förblir deras användning i säkerhetskritiska applikationer begränsad på grund av deras oförmåga att tillhandahålla tillförlitliga osäkerhetsuppskattningar och ofta förekommande övermodiga förutsägelser, vilket normalt är fallet i moderna neurala nätverk som har ett stort antal lager. I den här avhandlingen utnyttjar vi förmågan hos datautvinningsalgoritmer som densitetsklustring för att förklara beteendet hos en medicinsk bildklassificerare som är ansvarig för att klassificera vita blodkroppar. Vi vet att alla klustringsalgoritmer verkar på funktionsvektorn för indata och annoterar data i olika kluster enligt funktionerna. I detta arbete lägger vi ner och bevisar hypotesen att den utgående diskreta sannolikhetsmatrisen för ett klassificeringsproblem med flera klasser kan användas som en egenskapsvektor där konfidensvärdet för varje klass kan betraktas som en grad av likhet med den klassen. Innan man implementerar klustring måste man se till att dessa konfidensvärden representerar faktiska sannolikheter så att de kan användas som funktioner; därför införlivades vissa kalibreringstekniker för att först förbättra kalibreringen av nätverket. Med en bättre kalibrerad medicinsk klassificerare implementerades densitetsklustring, vilket genererade resultat som gav solida argument för att motivera nätverkets beteende. När det gäller användningsfallet för denna metod, observerades det att vi kunde identifiera patologier som myelodysplastiska syndrom, akut lymfatisk leukemi och kronisk myelomonocytisk leukemi hos en patient. Detta var möjligt på grund av närvaron av samma klass av vita blodkroppar i flera kluster, vilket indikerar närvaron av subpopulationer separerade i friska och patologiska celler av samma klass beroende på vilken patologi som behöver detekteras. Detta bevisades visuellt genom att kartlägga klusterpunkter till faktiska cellbilder och kvantitativt också genom att använda entropi som en metod för att kvantifiera osäkerhet. Denna metod visade att det finns mycket information inbäddad i utmatningssannolikhetsmatrisen. Därför kan man använda olika datautvinningstekniker för att extrahera mer information och inte bara begränsa sig till felklassificeringar och förvirringsmatriser.

No description available.