With the increasing availability of large datasets machine learning techniques are becoming an increasingly attractive alternative to expert-designed approaches to solving complex problems in domains where data is abundant. In this thesis we introduce several models for large sparse discrete datasets. Our approach, which is based on probabilistic models that use distributed representations to alleviate the effects of data sparsity, is applied to statistical language modelling and collaborative filtering. We introduce three probabilistic language models that represent words using learned real-valued vectors. Two of the models are based on the Restricted Boltzmann Machine (RBM) architecture while the third one is a simple deterministic model. We show that the deterministic model outperforms the widely used n-gram models and learns sensible word representations. To reduce the time complexity of training and making predictions with the deterministic model, we introduce a hierarchical version of the model, that can be exponentially faster. The speedup is achieved by structuring the vocabulary as a tree over words and taking advantage of this structure. We propose a simple feature-based algorithm for automatic construction of trees over words from data and show that the resulting models can outperform non-hierarchical neural models as well as the best n-gram models. We then turn our attention to collaborative filtering and show how RBM models can be used to model the distribution of sparse high-dimensional user rating vectors efficiently, presenting inference and learning algorithms that scale linearly in the number of observed ratings. We also introduce the Probabilistic Matrix Factorization model which is based on the probabilistic formulation of the low-rank matrix approximation problem for partially observed matrices. The two models are then extended to allow conditioning on the identities of the rated items whether or not the actual rating values are known. Our results on the Netflix Prize dataset show that both RBM and PMF models outperform online SVD models.

A North American man has approximately a one in six chance of being diagnosed with prostate cancer in his lifetime. In most cases, there is no clearly optimal treatment, so he may be invited to participate in a treatment decision between several medically reasonable options, each with potential short- and long-term side effects. Information needs are high at diagnosis and can continue to be elevated for years or decades. Many men and their families seek information online, where, due partly to the array of websites available and high variation in information preferences, it can be difficult to find personally relevant and useful websites. This research sought to address this issue by developing methods to categorize prostate cancer websites and exploring quantitative and qualitative relationships between websites, information-seekers, and individuals’ assessments of websites. The research involved a series of three studies. In the first study, 29 men with prostate cancer participated in a needs assessment involving questionnaires, an interview, and interaction with a prototype website. In the second study, a detailed classification system was developed and applied to a set of forty websites selected to be representative of the variety of prostate cancer websites available. The third (online) study collected clinical, cognitive, and psychosocial details from 65 participants along with their ratings of websites from study two. A number of hypotheses were tested. One finding was that, compared to men with greater trust, men with lower trust in their physician tended to judge commercial websites as less relevant and useful, and found websites with descriptions of personal experiences more relevant and useful. Analyses also addressed a number of exploratory questions, including whether website and individual attributes might predict preferences for websites. Using discriminant analysis on 80% of the data, two functions were identified that predicted ratings significantly better than chance. These relationships were then validated with 20% of the data held back for testing. The results are discussed in terms of their implications for information tailoring and recommender systems for prostate cancer patients searching for information online. Limitations of the current research and recommendations for future research are also presented.

Communication through text messaging, SMS (Short Message Service), is nowadays a huge industry with billions of active users. Because of the huge userbase it has attracted many companies trying to market themselves through unsolicited messages in this medium in the same way as was previously done through email. This is such a common phenomenon that SMS spam has now become a plague in many countries. This report evaluates several established machine learning algorithms to see how well they can be applied to the problem of filtering unsolicited SMS messages. Each filter is mainly evaluated by analyzing the accuracy of the filters on stored message data. The report also discusses and compares requirements for hardware versus performance measured by how many messages that can be evaluated in a fixed amount of time. The results from the evaluation shows that a decision tree filter is the best choice of the filters evaluated. It has the highest accuracy as well as a high enough process rate of messages to be applicable. The decision tree filter which was found to be the most suitable for the task in this environment has been implemented. The accuracy in this new implementation is shown to be as high as the implementation used for the evaluation of this filter. Though the decision tree filter is shown to be the best choice of the filters evaluated it turned out the accuracy is not high enough to meet the specified requirements. It however shows promising results for further testing in this area by using improved methods on the best performing algorithms.

Denna rapport omfattar ett arbete kring att förbättra signalbehandling och målföljning av en förbränningsfrekvens i en androidapplikation för effektberäkning hos accelererande fordon. Den ursprungliga applikationen är utvecklad på i3tex AB och det var även där som arbetet utfördes. Effektberäkningen görs genom att först spela in ljudet i kupén under ett accelerationsförlopp, där inspelningen signalbehandlas och förbränningsfrekvensen målföljs, sedan sker transformering från frekvensdomänen till hastighetsdomänen, varpå effekten beräknas via multiplikation av fordonets vikt, hastighet och acceleration. Problemet med den ursprungliga implementationen av målföljningen var att algoritmen inte var tillräckligt robust mot lågt signal/brus-förhållande (snr). För att göra systemet mer robust utvecklades flermålsföljning med Kalmanfilter, där ett poängsystem bestämmer vilken av målföljarna som mest troligt har följt förbränningsfrekvensen. Den nya algoritmen presterar betydligt bättre än den ursprungliga i avseende på rmse, men är betydligt mer resurskrävande. Genom optimering av hur signalbehandlingen görs, såsom längd och typ av fönsterfunktioner och andra parametrar för korttidsfouriertransformen (stft), är exekveringstiden för hela analysen marginellt snabbare och betydligt snabbare på en androidenhet med respektive utan stöd för hårdvaruaccelererade flyttalsoperationer. Det visade sig även att, trots att inte hårdvaruspecifikationerna för Android, cdd, specificerar inspelningar av frekvenser under 100 Hz är det möjligt på alla testade androidtelefoner och med tillräckligt gott resultat för att genomföra frekvensföljning enligt ovan.

La utilización de filtros activos paralelo mejora la eficiencia del suministro eléctrico mediante la modificación de las características de la forma de onda de las corrientes de línea. Trabajos de investigación anteriores destacan la complicada estructura de los controladores empleados en este tipo de soluciones y su sensibilidad a la distorsión de la onda de tensión en el punto donde se realiza su conexión. El objetivo fundamental de esta tesis es el desarrollo de un controlador para filtros activos de potencia tolerante a desequilibrios de tensión, huecos de tensión y armónicos de tensión. Objetivo secundario de este trabajo es el diseño de los algoritmos necesarios para el control de un filtro activo de potencia paralelo polifásico mediante una tarjeta basada en un procesador digital de señal. / Shunt active power filters (APF) modify the phase current waveform characteristics, which allow to increase the efficiency of electrical power grids. Previous research works establish the complex structure of controllers applied to APFs and their sensibility to voltage waveform disturbances.The aim of this thesis is the developing of a controller for shunt active power filters with tolerance to voltage unbalances, voltage dips and voltage harmonics. The proposed controller, implemented on a DSP target board, will be tested on a three-phase active power filter.

No description available.

No description available.

No description available.

Telecommunications companies (telcos) and Internet Service Providers (ISPs) monitor the traffic passing through their networks for the purposes of network evaluation and planning for future growth. Most monitoring techniques currently use a form of packet sampling. However, exhaustive monitoring is a preferable solution because it ensures accurate traffic characterization and also allows encoding operations, such as compression and encryption, to be performed. To overcome the very high computational cost of exhaustive monitoring and encoding of data, this thesis suggests exploiting parallelism. By utilizing a parallel cluster in conjunction with load balancing techniques, a simulation is created to distribute the load across the parallel processors. It is shown that a very scalable system, capable of supporting a fairly high data rate can potentially be designed and implemented. A complete system is then implemented in the form of a transparent Ethernet bridge, ensuring that the system can be deployed into a network without any change to the network. The system focuses its encoding efforts on obtaining the maximum compression rate and, to that end, utilizes the concept of streams, which attempts to separate data packets into individual flows that are correlated and whose redundancy can be removed through compression. Experiments show that compression rates are favourable and confirms good throughput rates and high scalability.

In order to control any system one needs to know the system's current state. In many real-world scenarios the state of the system cannot be determined with certainty due to the sensors being noisy or simply missing. In cases like these one needs to use probabilistic inference techniques to compute the likely states of the system and because such cases are common, there are lots of techniques to choose from in the field of Artificial Intelligence. Formally, we must compute a probability distribution function over all possible states. Doing this exactly is difficult because the number of states is exponential in the number of variables in the system and because the joint PDF may not have a closed form. Many approximation techniques have been developed over the years, but none ideally suited the problem we faced. Particle filtering is a popular scheme that approximates the joint PDF over the variables in the system by a set of weighted samples. It works even when the joint PDF has no closed form and the size of the sample can be adjusted to trade off accuracy for computation time. However, with many variables the size of the sample required for a good approximation can still become prohibitively large. Factored particle filtering uses the structure of variable dependencies to split the problem into many smaller subproblems and scales better if such decomposition is possible. However, our problem was unusual because some normally independent variables would become strongly correlated for short periods of time. This dynamically-changing dependency structure was not handled effectively by existing techniques. Considering variables to be always correlated meant the problem did not scale, considering them to be always independent introduced errors too large to tolerate. It was necessary to develop an approach that would utilize variables' independence whenever possible, but not introduce large errors when variables become correlated. We have developed a new technique for monitoring the state of the system for a class of systems with context-specific correlations. It is based on the idea of caching the context in which correlations arise and otherwise keeping the variables independent. Our evaluation shows that our technique outperforms existing techniques and is the first viable solution for the class of problems we consider.