Tesis sobre el tema "DBN (Deep Belief Network)"

The objective of reinforcement learning is to find "good" actions in an environment where feedback is provided through a numerical reward, and the current state (i.e. sensory input) is assumed to be available at each time step. The notion of "good" is defined as maximizing the expected cumulative returns over time. Sometimes it is useful to construct models of the environment to aid in solving the problem. We investigate Dyna-style reinforcement learning, a powerful approach for problems where not much real data is available. The main idea is to supplement real trajectories with simulated ones sampled from a learned model of the environment. However, in large state spaces, the problem of learning a good generative model of the environment has been open so far. We propose to use deep belief networks to learn an environment model. Deep belief networks (Hinton, 2006) are generative models that have been effective in learning the time dependency relationships among complex data. It has been shown that such models can be learned in a reasonable amount of time when they are built using energy models. We present our algorithm for using deep belief networks as a generative model for simulating the environment within the Dyna architecture, along with very promising empirical results.
L'objectif de l'apprentissage par renforcement est de choisir de bonnes actions dansun environnement où les informations sont fournies par une récompense numérique, etl'état actuel (données sensorielles) est supposé être disponible à chaque pas de temps. Lanotion de "correct" est définie comme étant la maximisation des rendements attendus cumulatifsdans le temps. Il est parfois utile de construire des modèles de l'environnementpour aider à résoudre le problème. Nous étudions l'apprentissage par renforcement destyleDyna, une approche performante dans les situations où les données réelles disponiblesne sont pas nombreuses. L'idée principale est de compléter les trajectoires réelles aveccelles simulées échantillonnées partir d'un modèle appri de l'environnement. Toutefois,dans les domaines à plusieurs états, le problème de l'apprentissage d'un bon modèlegénératif de l'environnement est jusqu'à présent resté ouvert. Nous proposons d'utiliserles réseaux profonds de croyance pour apprendre un modèle de l'environnement. Lesréseaux de croyance profonds (Hinton, 2006) sont des modèles génératifs qui sont efficaces pourl'apprentissage des relations de dépendance temporelle parmi des données complexes. Ila été démontré que de tels modèles peuvent être appris dans un laps de temps raisonnablequand ils sont construits en utilisant des modèles de l'énergie. Nous présentons notre algorithmepour l'utilisation des réseaux de croyance profonds en tant que modèle génératifpour simuler l'environnement dans l'architecture Dyna, ainsi que des résultats empiriquesprometteurs.

Cette thèse aborde le problème de la détection des feux de forêt par des outils de traitement d’images et apprentissage machine. Un incendie de forêt est un feu qui se propage sur une étendue boisée. Il peut être d'origine naturelle (dû à la foudre ou à une éruption volcanique) ou humaine. Dans le monde entier, l’impact des feux de forêts sur de nombreux aspects de notre vie quotidienne se fait de plus en plus apparente sur l’écosystème entier. De nombreuses méthodes ont montré l’efficacité pour la détection des incendies de forêt. L’originalité du présent travail réside dans la détection précoce des incendies par la détection de la fumée de forêt et la classification des régions de fumée et de non fumée à l’aide d’apprentissage profond et des outils de traitement d’image. Un ensemble de techniques de prétraitement nous a aidé à avoir une base de donnée importante (ajout du bruit aux entrées, augmentation des données) qui nous a permis après de tester la robustesse du modèle basée sur le DBN qu’on a proposé et évaluer la performance en calculant les métriques suivantes (IoU, Précision, Rappel, F1 score). Finalement, l’algorithme proposé est testé sur plusieurs images afin de valider son efficacité. Les simulations de notre algorithme ont été comparées avec celles traités dans l’état de l’art (Deep CNN, SVM…) et ont fourni de très bons résultats
This thesis deals with the problem of forest fire detection using image processing and machine learning tools. A forest fire is a fire that spreads over a wooded area. It can be of natural origin (due to lightning or a volcanic eruption) or human. Around the world, the impact of forest fires on many aspects of our daily lives is becoming more and more apparent on the entire ecosystem.Many methods have been shown to be effective in detecting forest fires. The originality of the present work lies in the early detection of fires through the detection of forest smoke and the classification of smoky and non-smoky regions using deep learning and image processing tools. A set of pre-processing techniques helped us to have an important database which allowed us afterwards to test the robustness of the model based on deep belief network we proposed and to evaluate the performance by calculating the following metrics (IoU, Accuracy, Recall, F1 score). Finally, the proposed algorithm is tested on several images in order to validate its efficiency. The simulations of our algorithm have been compared with those processed in the state of the art (Deep CNN, SVM...) and have provided very good results. The results of the proposed methods gave an average classification accuracy of about 96.5% for the early detection of smoke

Recent breakthroughs in machine learning are paving the way to the vision of software 2.0 era, which foresees the replacement of traditional software development with such techniques for many applications. In the context of agent-oriented programming, we believe that mixing together cognitive architectures like the BDI one and learning techniques could trigger new interesting scenarios. In that view, our previous work presents Jason-RL, a framework that integrates BDI agents and Reinforcement Learning (RL) more deeply than what has been already proposed so far in the literature. The framework allows the development of BDI agents having both explicitly programmed plans and plans learned by the agent using RL. The two kinds of plans are seamlessly integrated and can be used without differences. Here, we take autonomous driving as a case study to verify the advantages of the proposed approach and framework. The BDI agent has hard-coded plans that define high-level directions while fine-grained navigation is learned by trial and error. This approach – compared to plain RL – is encouraging as RL struggles in temporally extended planning. We defined and trained an agent able to drive in a track with an intersection, at which it has to choose the correct path to reach the assigned target. A first step towards porting the system in the real-world has been done by building a 1/10 scale racecar prototype which learned how to drive in a simple track.

Restricted Boltzmann Machines (RBMs) and autoencoders have been used - in several variants - for similar tasks, such as reducing dimensionality or extracting features from signals. Even though their structures are quite similar, they rely on different training theories. Lately, they have been largely used as building blocks in deep learning architectures that are called deep belief networks (instead of stacked RBMs) and stacked autoencoders. In light of this, the student has worked on this thesis with the aim to understand the extent of the similarities and the overall pros and cons of using either RBMs, autoencoders or denoising autoencoders in deep networks. Important characteristics are tested, such as the robustness to noise, the influence on training of the availability of data and the tendency to overtrain. The author has then dedicated part of the thesis to study how the three deep networks in exam form their deep internal representations and how similar these can be to each other. In result of this, a novel approach for the evaluation of internal representations is presented with the name of F-Mapping. Results are reported and discussed.

Sometimes finding labeled data for machine learning tasks is difficult. This is a problem for purely supervised models like the Multilayered perceptron(MLP). A Discriminative Deep Belief Network(DDBN) is a semi-supervised model that is able to use both labeled and unlabeled data. This research aimed to move towards a rule of thumb of when it is beneficial to use a DDBN instead of an MLP, given the proportions of labeled and unlabeled data. Several trials with different amount of labels, from the MNIST and Rectangles-Images datasets, were conducted to compare the two models. It was found that for these datasets, the DDBNs had better accuracy when few labels were available. With 50% or more labels available, the DDBNs and MLPs had comparable accuracies. It is concluded that a rule of thumb of using a DDBN when less than 50% of labels are available for training, would be in line with the results. However, more research is needed to make any general conclusions.
Märkt data kan ibland vara svårt att hitta för maskininlärningsuppgifter. Detta är ett problem för modeller som bygger på övervakad inlärning, exem- pelvis Multilayerd Perceptron(MLP). Ett Discriminative Deep Belief Network (DDBN) är en semi-övervakad modell som kan använda både märkt och omärkt data. Denna forskning syftar till att närma sig en tumregel om när det är för- delaktigt att använda en DDBN i stället för en MLP, vid olika proportioner av märkt och omärkt data. Flera försök med olika mängd märkt data, från MNIST och Rectangle-Images datamängderna, genomfördes för att jämföra de två mo- dellerna. Det konstaterades att för dessa datamängder hade DDBNerna bättre precision när ett fåtal märkt data fanns tillgängligt. När 50% eller mer av datan var märkt, hade DDBNerna och MLPerna jämförbar noggrannhet. Slutsatsen är att en tumregel att använda en DDBN när mindre än 50% av av träningsdatan är märkt, skulle vara i linje med resultaten. Det behövs dock mer forskning för att göra några generella slutsatser.

Dans ce travail de thèse, nous présentons nos travaux qui portent sur l’identification des cibles en général par un radar Ultra-Large Bande (ULB) et en particulier l’identification des cibles dont la surface équivalente radar est faible telles que les piétons et les cyclistes. Ce travail se décompose en deux parties principales, la détection et la reconnaissance. Dans la première approche du processus de détection, nous avons proposé et étudié un détecteur de radar ULB robuste qui fonctionne avec des données radar 1-D (A-scan) à une dimension. Il exploite la combinaison des statistiques d’ordres supérieurs et du détecteur de seuil automatique connu sous le nom de CA-CFAR pour Cell-Averaging Constant False Alarm Rate. Cette combinaison est effectuée en appliquant d’abord le HOS sur le signal reçu afin de supprimer une grande partie du bruit. Puis, après avoir éliminé le bruit du signal radar reçu, nous implémentons le détecteur de seuil automatique CA-CFAR. Ainsi, cette combinaison permet de disposer d’un détecteur de radar ULB à seuil automatique robuste. Afin d’améliorer le taux de détection et aller plus loin dans le traitement, nous avons évalué l’approche des données radar 2-D (B-Scan) à deux dimensions. Dans un premier temps, nous avons proposé une nouvelle méthode de suppression du bruit, qui fonctionne sur des données B-Scan. Il s’agit d’une combinaison de WSD et de HOS. Pour évaluer les performances de cette méthode, nous avons fait une étude comparative avec d’autres techniques de suppression du bruit telles que l’analyse en composantes principales, la décomposition en valeurs singulières, la WSD, et la HOS. Les rapports signal à bruit -SNR- des résultats finaux montrent que les performances de la combinaison WSD et HOS sont meilleures que celles des autres méthodes rencontrées dans la littérature. A la phase de reconnaissance, nous avons exploité les données des deux approches à 1-D et à 2-D obtenues à partir du procédé de détection. Dans la première approche à 1-D, les techniques SVM et le DBN sont utilisées et évaluées pour identifier la cible en se basant sur la signature radar. Les résultats obtenus montrent que la technique SVM donne de bonnes performances pour le système proposé où le taux de reconnaissance global moyen atteint 96,24%, soit respectivement 96,23%, 95,25% et 97,23% pour le cycliste, le piéton et la voiture. Dans la seconde approche à 1-D, les performances de différents types d’architectures DBN composées de différentes couches ont été évaluées et comparées. Nous avons constaté que l’architecture du réseau DBN avec quatre couches cachées est meilleure et la précision totale moyenne peut atteindre 97,80%. Ce résultat montre que les performances obtenues avec le DBN sont meilleures que celles obtenues avec le SVM (96,24%) pour ce système de reconnaissance de cible utilisant un radar ULB. Dans l’approche bidimensionnelle, le réseau de neurones convolutifs a été utilisé et évalué. Nous avons proposé trois architectures de CNN. La première est le modèle modifié d’Alexnet, la seconde est une architecture avec les couches de convolution arborescentes et une couche entièrement connectée, et la troisième est une architecture avec les cinq couches de convolution et deux couches entièrement connectées. Après comparaison et évaluation des performances de ces trois architectures proposées nous avons constaté que la troisième architecture offre de bonnes performances par rapport aux autres propositions avec une précision totale moyenne qui peut atteindre 99,59%. Enfin, nous avons effectué une étude comparative des performances obtenues avec le CNN, DBN et SVM. Les résultats montrent que CNN a les meilleures performances en termes de précision par rapport à DBN et SVM. Cela signifie que l’utilisation de CNN dans les données radar bidimensionnels permet de classer correctement les cibles radar ULB notamment pour les cibles à faible SER et SNR telles que les cyclistes ou les piétons
In this thesis work, we focused on the study and development of a system identification using UWB-Ultra-Wide-Band short range radar to detect the objects and particularly the vulnerable road users (VRUs) that have low RCS-Radar Cross Section- such as cyclist and pedestrian. This work is composed of two stages i.e. detection and recognition. In the first approach of detection stage, we have proposed and studied a robust UWB radar detector that works on one dimension 1-D radar data ( A-scan). It relies on a combination of Higher Order Statistics (HOS) and the well-known CA-CFAR (Cell-Averaging Constant False Alarm Rate) detector. This combination is performed by firstly applying the HOS to the received radar signal in order to suppress the noise. After eliminating the noise of the received radar signal, we apply the CA-CFAR detector. By doing this combination, we finally have an UWB radar detector which is robust against the noise and works with the adaptive threshold. In order to enhance the detection performance, we have evaluated the approach of using two dimensions 2-D (B-Scan) radar data. In this 2-D radar approach, we proposed a new method of noise suppression, which works on this B-Scan data. The proposed method is a combination of WSD (Wavelet Shrinkage Denoising) and HOS. To evaluate the performance of this method, we performed a comparative study with the other noise removal methods in literature including Principal Component Analysis (PCA), Singular Value Decomposition (SVD), WSD and HOS. The Signal-to-Noise Ratio (SNR) of the final result has been computed to compare the effectiveness of individual noise removal techniques. It is observed that a combination of WSD and HOS has better capability to remove the noise compared to that of the other applied techniques in the literature; especially it is found that it allows to distinguish efficiency the pedestrian and cyclist over the noise and clutters whereas other techniques are not showing significant result. In the recognition phase, we have exploited the data from the two approaches 1-D and 2-D, obtained from the detection method. In the first 1-D approach, Support Vector Machines (SVM) and Deep Belief Networks (DBN) have been used and evaluated to identify the target based on the radar signature. The results show that the SVM gives good performances for the proposed system where the total recognition accuracy rate could achieve up to 96,24%. In the second approach of this 1-D radar data, the performance of several DBN architectures compose of different layers have been evaluated and compared. We realised that the DBN architecture with four hidden layers performs better than those of with two or three hidden layers. The results show also that this architecture achieves up to 97.80% of accuracy. This result also proves that the performance of DBN is better than that of SVM (96.24%) in the case of UWB radar target recognition system using 1-D radar signature. In the 2-D approach, the Convolutional Neural Network (CNN) has been exploited and evaluated. In this work, we have proposed and investigated three CNN architectures. The first architecture is the modified of Alexnet model, the second is an architecture with three convolutional layers and one fully connected layer, and the third is an architecture with five convolutional layers and two fully connected layers. The performance of these proposed architectures have been evaluated and compared. We found that the third architecture has a good performance where it achieves up to 99.59% of accuracy. Finally, we compared the performances obtained using CNN, DBN and SVM. The results show that CNN gives a better result in terms of accuracy compared to that of DBN and SVM. It allows to classify correctly the UWB radar targets like cyclist and pedestrian

Les réseaux de neurones profonds (DNN) ont obtenu un succès remarquable sur de nombreuses applications du monde réel (par exemple, la reconnaissance de formes et la segmentation sémantique), mais sont toujours confrontés au problème de la gestion de l'incertitude. La théorie de Dempster-Shafer (DST) fournit un cadre bien fondé et élégant pour représenter et raisonner avec des informations incertaines. Dans cette thèse, nous avons proposé un nouveau framework utilisant DST et DNNs pour résoudre les problèmes d'incertitude. Dans le cadre proposé, nous hybridons d'abord DST et DNN en branchant une couche de réseau neuronal basée sur DST suivie d'une couche utilitaire à la sortie d'un réseau neuronal convolutif pour la classification à valeur définie. Nous étendons également l'idée à la segmentation sémantique en combinant des réseaux entièrement convolutifs et DST. L'approche proposée améliore les performances des modèles DNN en attribuant des modèles ambigus avec une incertitude élevée, ainsi que des valeurs aberrantes, à des ensembles multi-classes. La stratégie d'apprentissage utilisant des étiquettes souples améliore encore les performances des DNN en convertissant des données d'étiquettes imprécises et non fiables en fonctions de croyance. Nous avons également proposé une stratégie de fusion modulaire utilisant ce cadre proposé, dans lequel un module de fusion agrège les sorties de la fonction de croyance des DNN évidents selon la règle de Dempster. Nous utilisons cette stratégie pour combiner des DNN formés à partir d'ensembles de données hétérogènes avec différents ensembles de classes tout en conservant des performances au moins aussi bonnes que celles des réseaux individuels sur leurs ensembles de données respectifs. De plus, nous appliquons la stratégie pour combiner plusieurs réseaux superficiels et obtenir une performance similaire d'un DNN avancé pour une tâche compliquée
Deep neural networks (DNNs) have achieved remarkable success on many realworld applications (e.g., pattern recognition and semantic segmentation) but still face the problem of managing uncertainty. Dempster-Shafer theory (DST) provides a wellfounded and elegant framework to represent and reason with uncertain information. In this thesis, we have proposed a new framework using DST and DNNs to solve the problems of uncertainty. In the proposed framework, we first hybridize DST and DNNs by plugging a DSTbased neural-network layer followed by a utility layer at the output of a convolutional neural network for set-valued classification. We also extend the idea to semantic segmentation by combining fully convolutional networks and DST. The proposed approach enhances the performance of DNN models by assigning ambiguous patterns with high uncertainty, as well as outliers, to multi-class sets. The learning strategy using soft labels further improves the performance of the DNNs by converting imprecise and unreliable label data into belief functions. We have also proposed a modular fusion strategy using this proposed framework, in which a fusion module aggregates the belief-function outputs of evidential DNNs by Dempster’s rule. We use this strategy to combine DNNs trained from heterogeneous datasets with different sets of classes while keeping at least as good performance as those of the individual networks on their respective datasets. Further, we apply the strategy to combine several shallow networks and achieve a similar performance of an advanced DNN for a complicated task

With the diffusion of cheap sensors, sensor-equipped devices (e.g., drones), and sensor networks (such as Internet of Things), as well as the development of inexpensive human-machine interaction interfaces, the ability to quickly and effectively process sequential data is becoming more and more important. There are many tasks that may benefit from advancement in this field, ranging from monitoring and classification of human behavior to prediction of future events. Most of the above tasks require pattern recognition and machine learning capabilities. There are many approaches that have been proposed in the past to learn in sequential domains, especially extensions in the field of Deep Learning. Deep Learning is based on highly nonlinear systems, which very often reach quite good classification/prediction performances, but at the expenses of a substantial computational burden. Actually, when facing learning in a sequential, or more in general structured domain, it is common practice to readily resort to nonlinear systems. Not always, however, the task really requires a nonlinear system. So the risk is to run into difficult and computational expensive training procedures to eventually get a solution that improves of an epsilon (if not at all) the performances that can be reached by a simple linear dynamical system involving simpler training procedures and a much lower computational effort. The aim of this thesis is to discuss about the role that linear dynamical systems may have in learning in sequential domains. On one hand, we like to point out that a linear dynamical system (LDS) is able, in many cases, to already provide good performances at a relatively low computational cost. On the other hand, when a linear dynamical system is not enough to provide a reasonable solution, we show that it can be used as a building block to construct more complex and powerful models, or how to resort to it to design quite effective pre-training techniques for nonlinear dynamical systems, such as Echo State Networks (ESNs) and simple Recurrent Neural Networks (RNNs). Specifically, in this thesis we consider the task of predicting the next event into a sequence of events. The datasets used to test various discussed models involve polyphonic music and contain quite long sequences. We start by introducing a simple state space LDS. Three different approaches to train the LDS are then considered. Then we introduce some brand new models that are inspired by the LDS and that have the aim to increase the prediction/classification capabilities of the simple linear models. We then move to study the most common nonlinear models. From this point of view, we considered the RNN models, which are significantly more computationally demanding. We experimentally show that, at least for the addressed prediction task and the considered datasets, the introduction of pre-training approaches involving linear systems leads to quite large improvements in prediction performances. Specifically, we introduce pre-training via linear Autoencoder, and an alternative based on Hidden Markov Models (HMMs). Experimental results suggest that linear models may play an important role for learning in sequential domains, both when used directly or indirectly (as basis for pre-training approaches): in fact, when used directly, linear models may by themselves return state-of-the-art performance, while requiring a much lower computational effort with respect to their nonlinear counterpart. Moreover, even when linear models do not perform well, it is always possible to successfully exploit them within pre-training approaches for nonlinear systems.
Con la diffusione di dispositivi a basso costo, e reti di sensori (come ad esempio l'Internet of Things), nonché lo sviluppo di interfacce di interazione uomo-macchina a basso costo, la capacità di processare dati sequenziali in maniera veloce, e assicurando un basso consumo di risorse, è diventato sempre più importante. Molti sono i compiti che trarrebbero beneficio da un avanzamento in questo ambito, dal monitoraggio e classificazione di comportamenti umani fino alla predizioni di eventi futuri. Molti dei task citati richiedono l'uso di tecniche di pattern recognition e di abilità correlate con metodi tipici dell’apprendimento automatico. Molti sono gli approcci per eseguire apprendimento su domini sequenziali proposti nel recente passato, e molti sono basati su tecniche tipiche dell'ambito del Deep Learning. I metodi di Deep Learning sono tipicamente basati su sistemi fortemente non lineari, capaci di ottenere ottimi risultati in problemi di predizione/classificazione, ma che risultano anche essere molto costosi dal punto di vista computazionale. Quando si cerca di eseguire un compito di apprendimento su domini sequenziali, e più in generale su dati strutturati, tipicamente si ricorre all'utilizzo di sistemi non lineari. Non è però sempre vero che i task considerati richiedono modelli non lineari. Quindi il rischio è di andare ad utilizzare metodi troppo complessi, e computazionalmente costosi, per poi ottenere alla fine soluzioni che migliorano di un’epsilon (o anche no migliorano) i risultati ottenibili tramite l'utilizzo di sistemi lineari dinamici, che risultano essere molto meno costosi dal punto di vista dell'apprendimento, e del costo computazionale. L'obiettivo di questa tesi è di discutere del ruolo che i sistemi lineari dinamici possono avere nelle esecuzioni di compiti di apprendimento su dati strutturati. In questa tesi vogliamo mettere in luce le capacità dei sistemi lineari dinamici (LDS) di ottenere soluzioni molto buone ad un costo computazionale relativamente basso. Inoltre risulta interessante vedere come, nel caso in cui un sistema lineare non sia sufficiente per ottenere il risultato sperato, esso possa essere usato come base per costruire modelli più complessi, oppure possa essere utilizzato per eseguire la fase di pre-training per un modello non lineare, come ad esempio Echo State Networks (ESNs) e Recurrent Neural Networks (RNNs). Nello specifico in questa tesi è stato considerato un task di predizione dell'evento successivo, data una sequenza di eventi. I dataset usati per testare i vari modelli proposti nella tesi, contengono sequenze di musica polifonica, che risultano essere particolarmente lunghe e complesse. Nella prima parte della tesi viene proposto l'utilizzo del semplice modello LDS per affrontare il compito considerato. In particolare vengono considerati tre approcci diversi per eseguire l'apprendimento con questo modello. Viene poi introdotti nuovi modelli, ispirati al modello LDS, che hanno l'obiettivo di migliorare le prestazioni di quest'ultimo nei compiti di predizione/classificazione. Vengono poi considerati i più comuni modelli non lineari, in particolare il modello RNN il quale risulta essere significativamente più complesso e computazionalmente costoso da utilizzare. Viene quindi empiricamente dimostrato che, almeno per quanto riguarda il compito di predizione e i dataset considerati, l'introduzione di una fase di pre-training basati su sistemi lineari porta ad un significativo miglioramento delle prestazioni e della accuratezza nell'eseguire la predizione. In particolare 2 metodi di pre-training vengono proposti, il primo chiamato pre-training via Linear Autoencoder, ed il secondo basato su Hidden Markov Models (HMMs). I risultati sperimentali suggeriscono che i sistemi lineari possono giocare un ruolo importante per quanto riguarda il compito di apprendimento in domini sequenziali, sia che siano direttamente usati oppure siano usati indirettamente (come base per eseguire la fase di pre-training): infatti, usandoli direttamente, essi hanno permesso di raggiungere risultati che rappresentano lo stato dell'arte, andando però a richiedere uno sforzo computazionale molto limitato se confrontato con i più comuni modelli non lineari. Inoltre, anche quando le performance ottenute sono risultate non soddisfacenti, si è dimostrato che è possibile utilizzarli con successo per eseguire la fase di pre-training di sistemi non lineari.

Multimodal biometric systems have been widely applied in many real-world applications due to its ability to deal with a number of significant limitations of unimodal biometric systems, including sensitivity to noise, population coverage, intra-class variability, non-universality, and vulnerability to spoofing. This PhD thesis is focused on the combination of both the face and the left and right irises, in a unified hybrid multimodal biometric identification system using different fusion approaches at the score and rank level. Firstly, the facial features are extracted using a novel multimodal local feature extraction approach, termed as the Curvelet-Fractal approach, which based on merging the advantages of the Curvelet transform with Fractal dimension. Secondly, a novel framework based on merging the advantages of the local handcrafted feature descriptors with the deep learning approaches is proposed, Multimodal Deep Face Recognition (MDFR) framework, to address the face recognition problem in unconstrained conditions. Thirdly, an efficient deep learning system is employed, termed as IrisConvNet, whose architecture is based on a combination of Convolutional Neural Network (CNN) and Softmax classifier to extract discriminative features from an iris image. Finally, The performance of the unimodal and multimodal systems has been evaluated by conducting a number of extensive experiments on large-scale unimodal databases: FERET, CAS-PEAL-R1, LFW, CASIA-Iris-V1, CASIA-Iris-V3 Interval, MMU1 and IITD and MMU1, and SDUMLA-HMT multimodal dataset. The results obtained have demonstrated the superiority of the proposed systems compared to the previous works by achieving new state-of-the-art recognition rates on all the employed datasets with less time required to recognize the person’s identity.Multimodal biometric systems have been widely applied in many real-world applications due to its ability to deal with a number of significant limitations of unimodal biometric systems, including sensitivity to noise, population coverage, intra-class variability, non-universality, and vulnerability to spoofing. This PhD thesis is focused on the combination of both the face and the left and right irises, in a unified hybrid multimodal biometric identification system using different fusion approaches at the score and rank level. Firstly, the facial features are extracted using a novel multimodal local feature extraction approach, termed as the Curvelet-Fractal approach, which based on merging the advantages of the Curvelet transform with Fractal dimension. Secondly, a novel framework based on merging the advantages of the local handcrafted feature descriptors with the deep learning approaches is proposed, Multimodal Deep Face Recognition (MDFR) framework, to address the face recognition problem in unconstrained conditions. Thirdly, an efficient deep learning system is employed, termed as IrisConvNet, whose architecture is based on a combination of Convolutional Neural Network (CNN) and Softmax classifier to extract discriminative features from an iris image. Finally, The performance of the unimodal and multimodal systems has been evaluated by conducting a number of extensive experiments on large-scale unimodal databases: FERET, CAS-PEAL-R1, LFW, CASIA-Iris-V1, CASIA-Iris-V3 Interval, MMU1 and IITD and MMU1, and SDUMLA-HMT multimodal dataset. The results obtained have demonstrated the superiority of the proposed systems compared to the previous works by achieving new state-of-the-art recognition rates on all the employed datasets with less time required to recognize the person’s identity.
Higher Committee for Education Development in Iraq

This thesis investigates the use of deep learning techniques to address the problem of machine understanding of human affective behaviour and improve the accuracy of both unimodal and multimodal human emotion recognition. The objective was to explore how best to configure deep learning networks to capture individually and jointly, the key features contributing to human emotions from three modalities (speech, face, and bodily movements) to accurately classify the expressed human emotion. The outcome of the research should be useful for several applications including the design of social robots.

Recent advancements in processing power, storage capabilities, and availability of data, has led to improvements in many applications through the use of machine learning. Using machine learning in control systems was first suggested in the 1990s, but is more recently being implemented. In this thesis, an embedded climate system, which is a type of control system, will be looked at. The ways in which machine learning can be used to replicate portions of the climate system is looked at. Deep Belief Networks are the machine learning models of choice. Firstly, the functionality of a PID controller is replicated using a Deep Belief Network. Then, the functionality of a more complex control path is replicated. The performance of the Deep Belief Networks are evaluated at how they compare to the original control portions, and the performance in hardware. It is found that the Deep Belief Network can quite accurately replicate the behaviour of a PID controller, whilst the performance is worse for the more complex control path. It was seen that the use of delays in input features gave better results than without. A climate system with a Deep Belief Network was also loaded onto hardware. The minimum requirements of memory usage and CPU usage were met. However, the CPU usage was greatly affected, and if this was to be used in practice, work should be done to decrease it.
Många applikationer har förbättras genom användningen av maskininlärning. Maskininlärning för reglersystem föreslogs redan på 1990-talet och har nu börjat tillämpas, eftersom processorkraft, lagringsmöjligheter och tillgänglighet till rådata ökat. I detta examensarbete användes ett inbäddat klimatsystem, som är en typ av reglersystem. Maskininlärningsmodellen Deep Belief Network användes för att undersöka hur delar av klimatsystemet skulle kunna återskapas. Först återskapades funktionaliteten hos en PID-regulator och sedan funktionaliteten av en mer komplex del av reglersystemet Prestandan hos nätverken utvärderades i jämförelse med prestandan i de ursprungliga kontrolldelarna och hårdvaran. Det visade sig att Deep Belief Network utmärkt kunde replikera PID-regulatorns beteende, medan prestandan var lägre för den komplexa delen av reglersystemet. Användningen av fördröjningar i indata till nätverken gav bättre resultat än utan. Ett klimatsystem med ett Deep Belief Network laddades också över på hårdvaran. Minimikrav för minnesanvändning och CPU- användning var uppfyllda, men CPU- användningen påverkades kraftigt. Detta gör, att om maskininlärning ska kunna användas i verkligheten, bör CPU-användningen minskas.

Advances in unsupervised learning and deep neural networks have led to increased performance in a number of domains, and to the ability to draw strong comparisons between the biological method of self-organization conducted by the brain and computational mechanisms. This thesis aims to use real-world data to tackle two areas in the domain of computer vision which have biological equivalents: feature detection and motion tracking. The aforementioned advances have allowed efficient learning of feature representations directly from large sets of unlabeled data instead of using traditional handcrafted features. The first part of this thesis evaluates such representations by comparing regularization and preprocessing methods which incorporate local neighbouring information during training on a single-layer neural network. The networks are trained and tested on the Hollywood2 video dataset, as well as the static CIFAR-10, STL-10, COIL-100, and MNIST image datasets. The induction of topography or simple image blurring via Gaussian filters during training produces better discriminative features as evidenced by the consistent and notable increase in classification results that they produce. In the visual domain, invariant features are desirable such that objects can be classified despite transformations. It is found that most of the compared methods produce more invariant features, however, classification accuracy does not correlate to invariance. The second, and paramount, contribution of this thesis is a biologically-inspired model to explain the emergence of motion tracking behaviour in early development using unsupervised learning. The model’s self-organization is biased by an original concept called retinal constancy, which measures how similar visual contents are between successive frames. In the proposed two-layer deep network, when exposed to real-world video, the first layer learns to encode visual motion, and the second layer learns to relate that motion to gaze movements, which it perceives and creates through bi-directional nodes. This is unique because it uses general machine learning algorithms, and their inherent generative properties, to learn from real-world data. It also implements a biological theory and learns in a fully unsupervised manner. An analysis of its parameters and limitations is conducted, and its tracking performance is evaluated. Results show that this model is able to successfully follow targets in real-world video, despite being trained without supervision on real-world video.

One of the foremost characteristics of the mammalian visual system is the retinotopic mapping observed in the low-level visual processing centres; the spatial pattern of activation in the lateral geniculate nucleus and primary visual cortex corresponds topologically to the pattern of light falling on the retina. Various vision systems have been developed that take advantage of structured input such as retinotopy, however these systems are often not biologically plausible. Using a parsimonious approach for implementing retinotopy, one that is based on the biology of our visual pathway, we run simulations of visual learning using a deep belief network (DBN). Experiments show that we can successfully produce receptive fields and activation maps typical of the LGN and visual cortex respectively. These results may indicate a possible avenue of exploration into discovering the workings of the early visual system (and possibly more) on a neuronal level.

Recent developments in social media and web technologies offer new opportunities to access, analyze and process ever-increasing amounts of fashion-related data. In the appealing context of design and fashion, our main goal is to automatically suggest fashionable outfits based on the preferences extracted from real-world data provided either by individual users or gathered from the internet. In our case, the clothing items have the form of 2D-images. Especially for visual data processing tasks, recent models of deep neural networks are known to surpass human performance. This fact inspired us to apply the idea of transfer learning to understand the actual variability in clothing items. The principle of transfer learning consists in extracting the internal representa- tions formed in large convolutional networks pre-trained on general datasets, e.g., ImageNet, and visualizing its (similarity) structure. Together with transfer learn- ing, clustering algorithms and the image color schemes can be, namely, utilized when searching for related outfit items. Viable means applicable to generating new out- fits include deep belief networks and genetic algorithms enhanced by a convolutional network that models the outfit fitness. Although fashion-related recommendations remain highly subjective, the results we have achieved...

In the area of image recognition, the so-called deep neural networks belong to the most promising models these days. They often achieve considerably better results than traditional techniques even without the necessity of any excessive task-oriented preprocessing. This thesis is devoted to the study and analysis of three basic variants of deep neural networks-namely the neocognitron, convolutional neural networks, and deep belief networks. Based on extensive testing of the described models on the standard task of handwritten digit recognition, the convolutional neural networks seem to be most suitable for the recognition of general image data. Therefore, we have used them also to classify images from two very large data sets-CIFAR-10 and ImageNet. In order to optimize the architecture of the applied networks, we have proposed a new pruning algorithm based on the Principal Component Analysis. Powered by TCPDF (www.tcpdf.org)

碩士
國立高雄科技大學
電子工程系
107
The deep belief network (DBN) is a classic and representative neural network designed to solve classification problems. Stochastic computing (SC) is a highly efficient and attractive paradigm with low-cost hardware, the computation operation can be implemented by simple logic gates. The range of the conventional SC in the bipolar format is limited in the interval of [-1, 1], while the integral stochastic computing (ISC) expands the range to [-m, m], where m is the number of input streams. The new integral stochastic computing (NISC) has recently been introduced to improve hardware cost of ISC by reducing the number of states in the finite state machine (FSM). In this thesis, we propose a novel NISC-DBN architecture to improve hardware cost of the conventional ISC-DBN framework. The four-layer DBN structure 784-100-200-10 is considered. Simulation results reveal NISC-DBN outperform ISC-DBN in terms of the mean-square error (MSE). The classification accuracy of the NISC-DBN is also superior to that of ISC-DBN by applying the modified national institute of standards and technology (MNIST) dataset. The proposed NISC-DBN only increases the hardware cost by 1.6% over ISC-DBN in implementation the stochastic neuron of the first layer.

碩士
龍華科技大學
資訊管理系碩士班
106
The issue of air pollution is more and more important because influence of air pollution is also increasing in the world. The environment is affected by air pollution which makes the plant is slowly grow, genetic mutations, or diseases in humans. This study proposed a suitable prediction models for air pollutants in various regions, and the prediction model can obtain better performance. When anomalies are predicted, early warning can be provided to increase the time for prevention. This study collected data on air pollutants from four monitoring stations which are respectively Linkou Station in the New Taipei City, Erlin Station in the Changlin County, Hualien Station in Hualien County, and Nanxun Station in the Kaohsiung City. The air quality monitoring stations of the Environmental Protection Agency collected SO2, CO, O3, PM10, NOx, NO, and NO2. This study used the DBN, SDBN, ARIMA, and SARIMA models to predict air pollution concentration. This study can provide suitable prediction model for air pollutants in various areas.

Though adaptational effects are found throughout the visual system, the underlying mechanisms and benefits of this phenomenon are not yet known. In this work, the visual system is modeled as a Deep Belief Network, with a novel “post-training” paradigm (i.e. training the network further on certain stimuli) used to simulate adaptation in vivo. An optional sparse variant of the DBN is used to help bring about meaningful and biologically relevant receptive fields, and to examine the effects of sparsification on adaptation in their own right. While results are inconclusive, there is some evidence of an attractive bias effect in the adapting network, whereby the network’s representations are drawn closer to the adapting stimulus. As a similar attractive bias is documented in human perception as a result of adaptation, there is thus evidence that the statistical properties underlying the adapting DBN also have a role in the adapting visual system, including efficient coding and optimal information transfer given limited resources. These results are irrespective of sparsification. As adaptation has never been tested directly in a neural network, to the author’s knowledge, this work sets a precedent for future experiments.
text

碩士
國立交通大學
電信工程研究所
103
Binaural speech separation and de-reverberation are popular research topics and we have developed an unsupervised clustering method for these purposes. In this thesis, we adopt a supervised classification method for binaural speech separation and de-reverberation using the ideal binary mask (IBM) as the training target and a deep belief network (DBN) as the classifier. We extract the interaural time difference (ITD) and the interaural level difference (ILD) of each T-F unit as the binaural features. To boost the performance of de-reverberation, the interaural coherence (IC) is considered when building the target IBM. We propose three different DBN architectures, the side-by-side training (monaural training), the joint training (binaural training) and the multitask learning, and compare their binaural de-reverberation performance with the performance of our previously developed unsupervised clustering method in terms of many objective criteria.

碩士
元智大學
資訊工程學系
106
Compared to the past, people pay more and more attention to investment, how to make better use of limited funds to become the important of thinking. Certificate of exchange in the Taiwan 104 years the latest statistics of the cumulative number of shares of more than 17 million households, we can see that stock trading in Taiwan has become an indispensable investment pipeline. Stock price data provide a successful example in the stock forecast market where artificial intelligence (AI) techniques such as the Neuron Network have been widely used to predict stock prices and assist in investment strategies. However, the traditional neural network has been rapidly developed by the deep belief network (DBN) in image processing and semantic identification and other areas beyond, because the deep belief network has developed a number of skilled technology can be applied, will be more complex price information to do subtle and high-level abstract features of the expression, you can use the recent popular deep belief network to achieve. The purpose of this paper is to use deep belief network and TensorFlow system for rapid modeling and training to help investors can quickly digest and learn these stock information into a useful strategy for investors. We mainly use the machine learning model is deep belief network can achieve high-dimensional data feature expression, and the use of restricted Boltzmann machine from the non-marked data to learn the non-linear representation, which is the future deep learning Trends. the ease of use of data, and the amount of data can make us predict that the data will be more complex and full of noise, and expensive manual tagging data will become increasingly scarce, how to train from the unmarked data with high accuracy The system is still in the study, but the system of this paper in the use of a smaller amount of marker data to predict the stock when the ups and downs have a certain accuracy of the ability to predict, and the efficient establishment and training deep belief network model.

碩士
國立成功大學
高階管理碩士在職專班(EMBA)
105
The use of new artificial intelligent (AI) techniques, such as machine learning (ML) in the accounting domains, have unleashed great potential for researchers to improve accounting information systems (AIS). An automatic AIS reports financial statements from the supporting documents. The basic four financial statements provide information about the results of operations and financial position of an enterprise. As a result, those financial statements could be used to establish a diagnosis model for financial distress prediction (FDP). This study proposes an FDP model based on two ML approaches. Sixteen selected financial variables are calculated from financial statements to establish an FDP model using the deep belief network (DBN) algorithm coupled with the support vector machine (SVM) algorithm. Thrity-two distressed and thirty-two non-distressed companies (as matching samples) companies are selected from the Taiwan Economic Journal (TEJ) database, spanning the 2010-to-2016 sample period, to construct the FDP model. Three latent features of the financial data are extracted from 16 selected ratios by DBN and then divided into validation and training sets for SVM classification model construction. The constructed model is further used for prediction and evaluated by cross-validation. Our empirical results demonstrate that the proposed model could accurately predict the financial distress of a company. When using the previous two consecutive quarters of financial data before any event of distress, the prediction accuracy of the model could reach around 89% with the type I error of 4.7% and the type II error of 6.2%.

碩士
國立交通大學
電信工程研究所
103
Speech emotion recognition is a popular research topic from the last decade. Meanwhile, since the revival of deep learning in 2007, it has been adopted in various research fields. In this thesis, we use a deep belief network (DBN) as the classifier and examine its performance in detecting emotion states of noisy speech signals using rate-scale features (RS features) extracted from an auditory model. The noisy speech is derived by adding white and babble noises to clean utterances from the Berlin Emotional Speech database under various SNR levels. Afterward, the official feature set (Inter384) used in INTERSPEECH 2009 Emotion Challenge and a conventional support vector machine (SVM) classifier are considered for comparisons with the RS feature set and the DBN classifier, respectively. Furthermore, we propose an extended architecture of DBN based on the concept of multi-task learning (MTL) by adding a task of recognizing a different language (eNTERFACE 2005 Emotional Database) into the system. We postulate that one task could help speech emotion recognition performance of the other task. Simulation results demonstrate that (1) RS features yield higher recognition rates than Inter384 features; (2) DBN outperforms SVM using the RS features; (3) MTL-based DBN produces higher recognition rates than the original DBN.

Static authentication, where user identity is checked once at login time, can be circumvented no matter how strong the authentication mechanism is. Through attacks such as man-in-the-middle and its variants, an authenticated session can be hijacked later after the initial login process has been completed. In the last decade, continuous authentication (CA) using biometrics has emerged as a possible remedy against session hijacking. CA consists of testing the authenticity of the user repeatedly throughout the authenticated session as data becomes available. CA is expected to be carried out unobtrusively, due to its repetitive nature, which means that the authentication information must be collectible without any active involvement of the user and without using any special purpose hardware devices (e.g. biometric readers). Stylometry analysis, which consists of checking whether a target document was written or not by a specific individual, could potentially be used for CA. Although stylometric techniques can achieve high accuracy rates for long documents, it is still challenging to identify an author for short documents, in particular when dealing with large author populations. In this dissertation, we propose a new framework for continuous authentication using authorship verification based on the writing style. Authorship verification can be checked using stylometric techniques through the analysis of linguistic styles and writing characteristics of the authors. Different from traditional authorship verification that focuses on long texts, we tackle the use of short messages. Shorter authentication delay (i.e. smaller data sample) is essential to reduce the window size of the re-authentication period in CA. We validate our method using different block sizes, including 140, 280, and 500 characters, and investigate shallow and deep learning architectures for machine learning classification. Experimental evaluation of the proposed authorship verification approach based on the Enron emails dataset with 76 authors yields an Equal Error Rate (EER) of 8.21% and Twitter dataset with 100 authors yields an EER of 10.08%. The evaluation of the approach using relatively smaller forgery samples with 10 authors yields an EER of 5.48%.
Graduate

碩士
國立中興大學
統計學研究所
106
This thesis mainly uses deep belief network (DBN) pre-trained by restricted Boltzmann machine (RBM) to recognize high confused mandarin vowels such as ㄢ, ㄤ>, ㄛ , ㄨㄛ>, ㄥ, ㄣ>, etc. First, we would record the phonetic data of 20 speakers, and then perform a series of pre-processing such as digital sampling, endpoint detection, frame cutting, and windowing. Then take Mel-frequency cepstral coefficients (MFCC) as the features of the phonetic data, and use these features as the input to train the model. Different from multilayer perceptron (MLP) which uses random initial weights and biases, DBN uses RBM to pre-train the initial parameters in order to get a set of better initial parameters. After pre-training, take these initial parameters as the initial weights and biases of MLP, and then fine-tune these parameters by method of gradient descent. Since DBN obtains better initial parameters by pre-training, in the stage of using MLP to fine-tune parameters, the model converges faster than general MLP, and the recognition result is better, too. This research uses vowel data, each vowel has 25 frames, each frame has 39 features, and the model is DBN pre-trained by RBM which has one or two hidden layers. The identification rate of this method is at least 0.67% higher than that of MLP, and can increase by 9.61% at most. On average, DBN pre-trained by RBM has 4.59% higher identification rate than MLP.

Becoming a face expert takes years of learning and development. Many research programs are devoted to studying face perception, particularly given its prerequisite role in social interaction, yet its fundamental neural operations are poorly understood. One reason is that there are many possible explanations for a change in facial appearance, such as lighting, expression, or identity. Despite general agreement that the brain extracts multiple layers of feature detectors arranged into hierarchies to interpret causes of sensory information, very little work has been done to develop computational models of these processes, especially for complex stimuli like faces. The studies presented in this thesis used nonlinear generative models developed within machine learning to solve several face perception problems. Applying a deep hierarchical neural network, we showed that it is possible to learn representations capable of perceiving facial actions, expressions, and identities, better than similar non-hierarchical architectures. We then demonstrated that a generative architecture can be used to interpret high-level neural activity by synthesizing images in a top-down pass. Using this approach we showed that deep layers of a network can be activated to generate faces corresponding to particular categories. To facilitate training models to learn rich and varied facial features, we introduced a new expression database with the largest number of labeled faces collected to date. We found that a model trained on these images learned to recognize expressions comparably to human observers. Next we considered models trained on pairs of images, making it possible to learn how faces change appearance to take on different expressions. Modeling higher-order associations between images allowed us to efficiently match images of the same type according to a learned pairwise similarity measure. These models performed well on several tasks, including matching expressions and identities, and demonstrated performance superior to competing models. In sum, these studies showed that neural networks that extract highly nonlinear features from images using architectures inspired by the brain can solve difficult face perception tasks with minimal guidance by human experts.

Yes
The most recent face recognition systems are mainly dependent on feature representations obtained using either local handcrafted-descriptors, such as local binary patterns (LBP), or use a deep learning approach, such as deep belief network (DBN). However, the former usually suffers from the wide variations in face images, while the latter usually discards the local facial features, which are proven to be important for face recognition. In this paper, a novel framework based on merging the advantages of the local handcrafted feature descriptors with the DBN is proposed to address the face recognition problem in unconstrained conditions. Firstly, a novel multimodal local feature extraction approach based on merging the advantages of the Curvelet transform with Fractal dimension is proposed and termed the Curvelet–Fractal approach. The main motivation of this approach is that theCurvelet transform, a newanisotropic and multidirectional transform, can efficiently represent themain structure of the face (e.g., edges and curves), while the Fractal dimension is one of the most powerful texture descriptors for face images. Secondly, a novel framework is proposed, termed the multimodal deep face recognition (MDFR)framework, to add feature representations by training aDBNon top of the local feature representations instead of the pixel intensity representations. We demonstrate that representations acquired by the proposed MDFR framework are complementary to those acquired by the Curvelet–Fractal approach. Finally, the performance of the proposed approaches has been evaluated by conducting a number of extensive experiments on four large-scale face datasets: the SDUMLA-HMT, FERET, CAS-PEAL-R1, and LFW databases. The results obtained from the proposed approaches outperform other state-of-the-art of approaches (e.g., LBP, DBN, WPCA) by achieving new state-of-the-art results on all the employed datasets.

Commodity brain-computer interfaces (BCI) are beginning to accompany everything from toys and games to sophisticated health care devices. These contemporary interfaces allow for varying levels of interaction with a computer. Not surprisingly, the more intimately BCIs are integrated into the nervous system, the better the control a user can exert on a system. At one end of the spectrum, implanted systems can enable an individual with full body paralysis to utilize a robot arm and hold hands with their loved ones [28, 62]. On the other end of the spectrum, the untapped potential of commodity devices supporting electroencephalography (EEG) and electromyography (EMG) technologies require innovative approaches and further research. This thesis proposes a modularized software architecture designed to build flexible systems based on input from commodity BCI devices. An exploratory study using a commodity EEG provides concrete assessment of the potential for the modularity of the system to foster innovation and exploration, allowing for a combination of a variety of algorithms for manipulating data and classifying results. Specifically, this study analyzes a pipelined architecture for researchers, starting with the collection of spatio temporal brain data (STBD) from a commodity EEG device and correlating it with intentional behaviour involving keyboard and mouse input. Though classification proves troublesome in the preliminary dataset considered, the architecture demonstrates a unique and flexible combination of a liquid state machine (LSM) and a deep belief network (DBN). Research in methodologies and techniques such as these are required for innovation in BCIs, as commodity devices, processing power, and algorithms continue to improve. Limitations in terms of types of classifiers, their range of expected inputs, discrete versus continuous data, spatial and temporal considerations and alignment with neural networks are also identified.
Graduate
0317
0984
jasoncummer@gmail.com