This work deals with traveling salesman problem (TSP) and examines it’s possibilities to use in business. It is about the optimization of the travel cost, saving time and unnecessary mileage. Part of the work is a program with a GUI written in program MATLAB. Program uses neural networks to calculate the most effective path between places, where the trader has to reach. It’s possible to use the algorithm for many purposes, e.g. distribution of goods, store management, planning of PCBs or rescue services. Program communicates with the Google Maps API server, which provides the actual information of the path.

As complimentary metal-oxide semiconductor (CMOS) technologies scale and field-effect transistor (FET) architectures change, the factors in deciding to utilize analog or digital transistor behaviors evolve. This thesis examines three case studies where traditionally analog or digital circuitry has dominated published works but I show that the opposite regime has significant benefits in scaled CMOS technologies. I present a highly digital operational amplifier (traditionally analog) and two artificial neurons (traditionally digital). In Chapters 2 and 3 I present a highly-digital five-stage zero-crossing-based amplifier which breaks the trade-off between slew rate and settling accuracy. I investigate the optimal charge pump design by analyzing the effects of the current scaling factor, number of current sources, maximum current value, and input amplitude on the settling performance including overshoot and settling time. I find that there exists an optimal number of stages that yields the fastest settling for a given total current and load capacitance. The proposed amplifier achieves a signal-to-noise ratio of 57 dB at a sampling rate of 40 MHz and consumes 1.45 mW under a 1V supply. In Chapters 4 and 5, I propose two novel analog artificial spiking neurons, operating in the voltage domain and phase domain respectively. The voltage domain neuron presented in Chapter 4 implements a novel fine-tuning method called neuromodulatory tuning which reduced the number of parameters to be tuned by four orders of magnitude as compared with traditional fine-tuning methods. Chapter 5 presents the design of a novel phase-domain neuron. Voltage domain neurons mimic biological neurons by integrating charge on a capacitor. I instead integrate phase in a voltage-controlled ring oscillator (VCO). I also propose a novel bidirectional switched-capacitor synapse which saves significant area compared to bidirectional current based synapses. The proposed neuron, synapse and weight memory occupy only 21x27um, and consume 134fJ/spike under a 0.35V supply.

La représentation de l’information mnésique est toujours une question d’intérêt majeur en neurobiologie, mais également, du point de vue informatique, en apprentissage artificiel. Dans certains modèles de réseaux de neurones artificiels, nous sommes confrontés au dilemme de la récupération de l’information sachant, sur la base de la performance du modèle, que cette information est effectivement stockée mais sous une forme inconnue ou trop complexe pour être facilement accessible. C’est le dilemme qui se pose pour les grands réseaux de neurones et auquel tente de répondre le paradigme du « reservoir computing ».Le « reservoir computing » est un courant de modèles qui a émergé en même temps que le modèle que nous présentons ici. Il s’agit de décomposer un réseau de neurones en (1) une couche d’entrée qui permet d’injecter les exemples d’apprentissage, (2) un « réservoir » composé de neurones connectés avec ou sans organisation particulière définie, et dans lequel il peut y avoir des mécanismes d’adaptation, (3) une couche de sortie, les « readout », sur laquelle un apprentissage supervisé est opéré. Nous apportons toutefois une particularité, qui est celle d’utiliser les délais axonaux, temps de propagation d’une information d’un neurone à un autre. Leur mise en oeuvre est un apport computationnel en même temps qu’un argument biologique pour la représentation de l’information. Nous montrons que notre modèle est capable d’un apprentissage artificiel efficace et prometteur même si encore perfectible. Sur la base de ce constat et dans le but d’améliorer les performances nous cherchons à comprendre les dynamiques internes du modèle. Plus précisément nous étudions comment la topologie du réservoir peut influencer sa dynamique. Nous nous aidons pour cela de la théorie des groupes polychrones. Nous avons développé, pour l’occasion, des algorithmes permettant de détecter ces structures topologico-dynamiques dans un réseau, et dans l’activité d’un réseau de topologie donnée.Si nous comprenons les liens entre topologie et dynamique, nous pourrons en tirer parti pour créer des réservoirs adaptés aux besoins de l’apprentissage. Finalement, nous avons mené une étude exhaustive de l’expressivité d’un réseau en termes de groupes polychrones, en fonction de différents types de topologies (aléatoire, régulière, petit-monde) et de nombreux paramètres (nombre de neurones, connectivité, etc.). Nous pouvons enfin formuler un certain nombre de recommandations pour créer un réseau dont la topologie peut être un support riche en représentations possibles. Nous tentons également de faire le lien avec la théorie cognitive de la mémoire à traces multiples qui peut, en principe, être implémentée et étudiée par le prisme des groupes polychrones. / How memory information is represented is still an open question in neurobiology, but also, from the computer science point of view, in machine learning. Some artificial neuron networks models have to face the problem of retrieving information, knowing that, in regard to the model performance, this information is actually stored but in an unknown form or too complex to be easily accessible. This is one of the problems met in large neuron networks and which « reservoir computing » intends to answer.« Reservoir computing » is a category of models that has emerged at the same period as, and has propoerties similar to the model we present here. It is composed of three parts that are (1) an input layer that allows to inject learning examples, (2) a « reservoir » composed of neurons connected with or without a particular predefined, and where there can be adaptation mecanisms, (3) an output layer, called « readout », on which a supervised learning if performed. We bring a particularity that consists in using axonal delays, the propagation time of information from one neuron to another through an axonal connexion. Using delays is a computational improvement in the light of machin learning but also a biological argument for information representation.We show that our model is capable of a improvable but efficient and promising artificial learning. Based on this observation and in the aim of improving performance we seek to understand the internal dynamics of the model. More precisely we study how the topology of the reservoir can influence the dynamics. To do so, we make use of the theory of polychronous groups. We have developped complexe algorithms allowing us to detect those topologicodynamic structures in a network, and in a network activity having a given topology.If we succeed in understanding the links between topology and dynamics, we may take advantage of it to be able to create reservoir with specific properties, suited for learning. Finally, we have conducted an exhaustive study of network expressivness in terms of polychronous groups, based on various types of topologies (random, regular, small-world) and different parameters (number of neurones, conectivity, etc.). We are able to formulate some recommandations to create a network whose topology can be rich in terms of possible representations. We propose to link with the cognitive theory of multiple trace memory that can, in principle, be implemented and studied in the light of polychronous groups.

Submitted by Irene Nascimento (irene.kessia@ufpe.br) on 2016-09-27T18:39:30Z No. of bitstreams: 2 license_rdf: 1232 bytes, checksum: 66e71c371cc565284e70f40736c94386 (MD5) RicardoDeAndradeAraujo.pdf: 2136922 bytes, checksum: 3bf9d638152b4cc1870ed7c533772fae (MD5) / Made available in DSpace on 2016-09-27T18:39:30Z (GMT). No. of bitstreams: 2 license_rdf: 1232 bytes, checksum: 66e71c371cc565284e70f40736c94386 (MD5) RicardoDeAndradeAraujo.pdf: 2136922 bytes, checksum: 3bf9d638152b4cc1870ed7c533772fae (MD5) Previous issue date: 2016-01-01 / CNPQ / Este trabalho apresenta um estudo sobre séries temporais financeiras, em alta-frequência, na tentativa de identificar as características do seu fenômeno gerador e, baseado neste estudo, propor um modelo, composto por uma combinação balanceada entre operadores lineares e operadores não-lineares crescentes e decrescentes, capaz de prever este tipo particular de série temporal. Para o processo de aprendizagem, é proposto um método baseado em gradiente descendente, utilizando ideias do algoritmo de retropropagação do erro (back propagation, BP) e uma abordagem alternativa para superar o problema da não-diferenciabilidade dos operadores não-lineares. Uma análise experimental é conduzida com o modelo proposto, utilizando um conjunto de séries temporais financeiras, em alta-frequência, do mercado de ações Brasileiro: Banco do Brasil SA, Banco Bradesco SA, Brasil Foods SA, BR Malls Participações SA e Companhia Energética Minas Gerais. Nestes experimentos, um conjunto relevante de medidas é utilizado para avaliar o desempenho preditivo do modelo proposto, e os resultados alcançados superam aqueles obtidos utilizando técnicas estatísticas, neurais e híbridas apresentadas na literatura. Também, são realizadas simulações com um sistema de apoio à decisão, baseado em previsão, para compra e venda de ações, tendo em vista demonstrar o desempenho econômico expressivo do modelo proposto no mercado de ações, em alta-frequência. / This work presents a study about high-frequency financial time series to identify the characteristics of their generator phenomenon and, based on such study, to propose a model, composed of a balanced combination of linear operators and increasing and decreasing nonlinear operators, able to predict this kind of time series. For the learning process, it is proposed a descent gradient-based method, using ideas from the back propagation (BP) algorithm and a systematic approach to overcome the problem of nondifferentiability of nonlinear operators. An experimental analysis is conducted with the proposed model, using a set of highfrequency financial time series of the Brazilian stock market: Banco do Brasil SA, Banco Bradesco SA, Brasil Foods SA, BR Malls Participações SA and Companhia Energética Minas Gerais. In these experiments, a relevant set of measures are used to assess the prediction performance of the proposed model, and the achieved results overcome those obtained by statistical, neural and hybrid techniques presented in the literature. Also, it is performed simulations with a prediction-based decision support system, for buy and sale of stocks, to demonstrate the significant economic performance of the proposed model in real high-frequency stock market

The presented thesis discusses the basic analysis methods of arteriovelocitograms. The core of this work rests in classification of signals and contribution to possibilities of noninvasive diagnostic methods for evaluation patients with peripheral ischemic occlusive arterial disease. The classification employs multivariate statistical methods and principles of neural networks. The data processing works with an angiographic verified set of arteriovelocitogram dates. The digital subtraction angiography classified them into 3 separable classes in dependence on degree of vascular stenosis. Classification AVG signals are represented in the program by the 6 parameters that are measured on 3 different places on each patient’s leg. Evaluation of disease appeared to be a comprehensive approach at signals acquired from whole patient’s leg. The sensitivity of clustering method compared with angiography is between 82.75 % and 90.90 %, specificity between 80.66 % and 88.88 %. Using neural networks sensitivity is in range of 79.06 % and 96.87 %, specificity is in range of 73.07 % and 91.30 %.

The thesis addresses the topic of Deep Neural Networks, in particular the methods regar- ding the field of Deep Learning, which is used to initialize the weight and learning process s itself within Deep Neural Networks. The focus is also put to the basic theory of the classical Neural Networks, which is important to comprehensive understanding of the issue. The aim of this work is to determine the optimal set of optional parameters of the algori- thms on various complexity levels of image recognition tasks through experimenting with created application applying Deep Neural Networks. Furthermore, evaluation and analysis of the results and lessons learned from the experimentation with classical and Deep Neural Networks are integrated in the thesis.

The master’s thesis is aimed at the expedience evaluation of use of a reverse leasing of investment financed by a long credit. First of all the fundamentals of a leasing, a long credit and the present state of the Czech leasing market are reviewed. The target of the effort is the comparison of several reverse leasing’s options as a source of financing.

This thesis focuses on fuzzy neural networks. The combination of the fuzzy logic and artificial neural networks leads to the development of more robust systems. These systems are used in various field of the research, such as artificial intelligence, machine learning and control theory. First, we provide a quick overview of underlying neural networks and fuzzy systems to explain fundamental ideas that form the basis of the fields, and follow with the introduction of the fuzzy neural network theory, classification and application. Then we describe a design and a realization of the fuzzy associative memory, as an example of these systems. Finally, we benchmark the realization using the pattern recognition and control tasks. The results are evaluated and compared against existing systems.