Many real-world problems like hand-written digit recognition or semantic scene classification are treated as multiclass or multi-label classification prob-lems. Solutions to these problems using support vector machines (SVMs) are well studied in literature. In this work, we focus on building sparse max-margin classifiers for multiclass and multi-label classification. Sparse representation of the resulting classifier is important both from efficient training and fast inference viewpoints. This is true especially when the training and test set sizes are large.Very few of the existing multiclass and multi-label classification algorithms have given importance to controlling the sparsity of the designed classifiers directly. Further, these algorithms were not found to be scalable. Motivated by this, we propose new formulations for sparse multiclass and multi-label classifier design and also give efficient algorithms to solve them. The formulation for sparse multi-label classification also incorporates the prior knowledge of label correlations. In both the cases, the classification model is designed using a common set of basis vectors across all the classes. These basis vectors are greedily added to an initially empty model, to approximate the target function. The sparsity of the classifier can be controlled by a user defined parameter, dmax which indicates the max-imum number of common basis vectors. The computational complexity of these algorithms for multiclass and multi-label classifier designisO(lk2d2 max), Where l is the number of training set examples and k is the number of classes. The inference time for the proposed multiclass and multi-label classifiers is O(kdmax). Numerical experiments on various real-world benchmark datasets demonstrate that the proposed algorithms result in sparse classifiers that require lesser number of basis vectors than required by state-of-the-art algorithms, to attain the same generalization performance. Very small value of dmax results in significant reduction in inference time. Thus, the proposed algorithms provide useful alternatives to the existing algorithms for sparse multiclass and multi-label classifier design.

A epilepsia, cujas crises são resultantes de distúrbios elétricos temporários no cérebro, é a quarta enfermidade neurológica mais comum, atingindo aproximadamente 50 milhões de pessoas. Essa enfermidade pode ser diagnosticada por meio de eletroencefalogramas (EEG), que são de elevada importância para o diagnóstico de enfermidades cerebrais. As informações consideradas relevantes desses exames são descritas em laudos médicos, que são armazenados com o objetivo de manter o histórico clínico do paciente e auxiliar os especialistas da área médica na realização de procedimentos futuros, como a identificação de padrões de determinadas enfermidades. Entretanto, o crescente aumento no armazenamento de dados médicos inviabiliza a análise manual dos mesmos. Outra dificuldade para a análise de EEG é a variabilidade de opiniões de especialistas sobre um mesmo padrão observado, podendo aumentar a dificuldade para o diagnóstico de enfermidades cerebrais. Também, os exames de EEG podem conter padrões relevantes difíceis de serem observados, mesmo por profissionais experientes. Da mesma forma, nos laudos podem faltar informações e/ou conter erros de digitação devido aos mesmos serem preenchidos apressadamente por especialistas. Assim, neste trabalho foi desenvolvido o método computacional de geração de laudos médicos (automatic generation of medical report AutoGenMR), que tem o propósito de auxiliar especialistas da área médica no diagnóstico de epilepsia e em tomadas de decisão. Esse processo é aplicado em duas fases: (1) construção de classificadores por meio de métodos de aprendizado de máquina e (2) geração automática de laudos textuais. O AutoGenMR foi avaliado experimentalmente em dois estudos de caso, para os quais, em cada um foi utilizada uma base de EEG disponibilizada publicamente e gratuitamente. Nessas avaliações foram utilizadas as mesmas configurações experimentais para a extração de características e construção de classificadores (desconsiderando que um dos problemas de classificação é multiclasse e o outro, binário). No primeiro estudo de caso, os modelos preditivos geraram, em média, 89% das expressões de laudos. Na segunda avaliação experimental, em média, 76% das sentenças de laudos foram geradas corretamente. Desse modo, os resultados de ambos estudos são considerados promissores, constatando que o AutoGenMR pode auxiliar especialistas na identificação de padrões relacionados a eventos epiléticos, na geração de laudos textuais padronizados e em processos de tomadas de decisão. / Epilepsy, which seizures are due to temporary electrical disturbances in the brain, is the fourth most common neurological disorder, affecting 50 million people, approximately. This disease can be diagnosed by electroencephalograms (EEG), which have great importance for the diagnosis of brain diseases. The information considered relevant in these tests is described in textual reports, which are stored in order to maintain the patients medical history and assist medical experts in performing such other procedures as the standard identification of certain diseases. However, the increasing medical data storage makes it unfeasible for manual analysis. Another challenge for the EEG analysis is the diversity of expert opinions on particular patterns observed and may increase the difficulty in diagnosing diseases of the brain. Moreover, the EEG may contain patterns difficult to be noticed even by experienced professionals. Similarly, the reports may not have information and/or include typographical errors due to its rushed filling by experts. Thereby, in this work, the automatic generation of medical report (AutoGenMR) method was developed in order to assist medical experts in the diagnosis of epilepsy and decision making. This method is applied in two phases: (1) classifier building by machine learning techniques and (2) automatic report generation. The AutoGenMR was computed in two case studies, for which, a public and freely available EEG database was used in each one. In both studies, the same experimental settings for feature extraction and classifier building were used. In the first study case, the classifiers correctly generated, on average, 89% of the report expressions. In the second experiment, on average, 76% of the report sentences were successfully generated. In this sense, the results of both studies are considered promising, noting that the AutoGenMR can assist medical experts in the identification of patterns related to epileptic events, standardized textual report generation, and in decision-making processes.

No presente trabalho provamos resultados sobre as flutuações dos fluxos de partículas e das partículas marcadas no processo de Hammersley multiclasse. Os métodos das demonstrações são robustos, formulados de modo a serem aplicados em outros processos, em particular se aplicam ao processo de exclusão totalmente assimétrico multiclasse (TASEP multiclasse) e à seu respectivo modelo de percolação de última passagem. Os principais teoremas obtidos são um teorema central do limite para o choque, seu coeficiente de difusão e uma fórmula exata para a variância do fluxo de partículas de classe N >1 para o processo em equilíbrio multiclasse. / We prove fluctuations results concerning fluxes of particles and tagged particles on multiclass Hammersley process. The methods used are robust and apply to other processes, in particular all the proofs can be adapted to the Multiclass totally asymmetric simple exclusion process (Multiclass TASEP) and its respective last passage percolation model. The main theorems obtained are a central limit theorem for the shock, its diffusion coefficient and an exact formula for the variance of the $N$-th class particle flux in a stationary version of the multiclass process when N > 1.

Every day most people are using applications and services that are utilising machine learning, in some way, without even knowing it. Some of these applications and services could, for example, be Google’s search engine, Netflix’s recommendations, or Spotify’s music tips. For machine learning to work it needs data, and often a large amount of it. Roughly 2,5 quintillion bytes of data are created every day in the modern information society. This huge amount of data can be utilised to make applications and systems smarter and automated. Time logging systems today are usually not smart since users of these systems still must enter data manually. This bachelor thesis will explore the possibility of applying machine learning to task logging systems, to make it smarter and automated. The machine learning algorithm that is used to predict the user’s task, is called multiclass logistic regression, which is categorical. When a small amount of training data was used in the machine learning process the predictions of a task had a success rate of about 91%.

No presente trabalho provamos resultados sobre as flutuações dos fluxos de partículas e das partículas marcadas no processo de Hammersley multiclasse. Os métodos das demonstrações são robustos, formulados de modo a serem aplicados em outros processos, em particular se aplicam ao processo de exclusão totalmente assimétrico multiclasse (TASEP multiclasse) e à seu respectivo modelo de percolação de última passagem. Os principais teoremas obtidos são um teorema central do limite para o choque, seu coeficiente de difusão e uma fórmula exata para a variância do fluxo de partículas de classe N >1 para o processo em equilíbrio multiclasse. / We prove fluctuations results concerning fluxes of particles and tagged particles on multiclass Hammersley process. The methods used are robust and apply to other processes, in particular all the proofs can be adapted to the Multiclass totally asymmetric simple exclusion process (Multiclass TASEP) and its respective last passage percolation model. The main theorems obtained are a central limit theorem for the shock, its diffusion coefficient and an exact formula for the variance of the $N$-th class particle flux in a stationary version of the multiclass process when N > 1.

By combining computed tomography data with computational fluid dynamics, the cardiac hemodynamics of a patient can be assessed for diagnosis and treatment of cardiac disease. The advantage of computed tomography over other medical imaging modalities is its capability of producing detailed high resolution images containing geometric measurements relevant to the simulation of cardiac blood flow. To extract these geometries from computed tomography data, segmentation of 4D cardiac computed tomography (CT) data has been performed using two deep learning frameworks that combine methods which have previously shown success in other research. The aim of this thesis work was to develop and evaluate a deep learning based technique to segment the left ventricle, ascending aorta, left atrium, left atrial appendage and the proximal pulmonary vein inlets. Two frameworks have been studied where both utilise a 2D multi-axis implementation to segment a single CT volume by examining it in three perpendicular planes, while one of them has also employed a 3D binary model to extract and crop the foreground from surrounding background. Both frameworks determine a segmentation prediction by reconstructing three volumes after 2D segmentation in each plane and combining their probabilities in an ensemble for a 3D output.  The results of both frameworks show similarities in their performance and ability to properly segment 3D CT data. While the framework that examines 2D slices of full size volumes produces an overall higher Dice score, it is less successful than the cropping framework at segmenting the smaller left atrial appendage. Since the full size 2D slices also contain background information in each slice, it is believed that this is the main reason for better segmentation performance. While the cropping framework provides a higher proportion of each foreground label, making it easier for the model to identify smaller structures. Both frameworks show success for use in 3D cardiac CT segmentation, and with further research and tuning of each network, even better results can be achieved.

Prostate cancer is the second most diagnosed cancer worldwide and its diagnosis is done through visual inspection of biopsy tissue by a pathologist, who assigns a score used by doctors to decide on the treatment. However, the scoring system, the Gleason score, is affected by a high inter and intra-observer variability, lack of standardization, and overestimation. Therefore, there is a need for new solutions that can reduce these issues and provide a more accurate diagnosis. Nowadays, high-resolution digital images of biopsy tissues can be obtained and stored. The availability of such images, called Whole Slide Images (WSI) allows the implementation of Machine and Deep learning models to assist pathologists in diagnosing prostate cancer. Multiple-Instance Learning (MIL) has been shown to reach very promising results in digital pathology and binary classification of prostate cancer slides. However, such models require large datasets to ensure good performances. This project wants to investigate the use of small sets of strongly annotated images to create new large datasets to train a MIL model. To evaluate the performance of this approach, the standard dataset is used to obtain baselines for both binary and multiclass classification tasks. For multiclassification, the International Society of Urological Pathology (ISUP) score is used, which is derived from the Gleason score. The dataset used is the publicly available PANDA. In this project, only the slides from RadboudUniversity Medical Center are used, which consists of 5160 images. The MIL model chosen is the Clustering-constrained Attention Multiple instance learning (CLAM) model, which is publicly available. The standard approach reaches a Cohen’s kappa (κ) of 0.78 and 0.59 for binary and multiclass classification respectively. To evaluate the new approach, large datasets are created starting from different set sizes. Using 500 images, the model reaches a κ of 0.72 and 0.38 respectively. While for the binary the results of the two approaches are comparable, the new approach is not beneficial for multiclass classification tasks.

[pt] Modelos de Aprendizado Profundo para classificação de imagens alcançaram o estado da arte em um vasto campo de aplicações. Entretanto, é frequente deparar-se com amostras ruidosas, isto é, amostras contendo rótulos incorretos, nos conjuntos de dados provenientes de aplicações do mundo real. Quando modelos de Aprendizado Profundo são treinados nestes conjuntos de dados, a sua performance é prejudicada. Modelos do estado da arte, como Co-teaching+ e Jocor, utilizam a técnica Small Loss Approach (SLA) para lidar com amostras ruidosas no cenário multiclasse. Nesse trabalho, foi desenvolvido uma nova técnica para lidar com amostras ruidosas, chamada Recovering Discarded Samples (RDS), que atua em conjunto com a SLA. Para demostrar a eficácia da técnica, aplicou-se o RDS nos modelos Co-teaching+ e Jocor resultando em dois novos modelos RDS-C e RDS-J. Os resultados indicam ganhos de até 6 por cento nas métricas de teste para ambos os modelos. Um terceiro modelo chamado RDS-Contrastive também foi desenvolvido, este modelo superou o estado da arte em até 4 por cento na acurácia de teste. Além disso, nesse trabalho, expandiu-se a técnica SLA para o cenário multilabel, sendo desenvolvido a técnica SLA Multilabel (SLAM). Com essa técnica foi desenvolvido mais dois modelos para cenário multilabel com amostras ruidosas. Os modelos desenvolvidos nesse trabalho para multiclasse foram utilizados em um problema real de cunho ambiental. Os modelos desenvolvidos para o cenário multilabel foram aplicados como solução para um problema real na área de óleo e gás. / [en] Deep Learning models designed for image classification have consistently achieved state-of-the-art performance across a plethora of applications. However, the presence of noisy samples, i.e., instances with incorrect labels, is a prevalent challenge in datasets derived from real-world applications. The training of Deep Learning models on such datasets inevitably compromises their performance. State-of-the-art models, such as Co-teaching+ and Jocor, utilize the Small Loss Approach (SLA) technique to handle noisy samples in a multi-class scenario. In this work, a new technique named Recovering Discarded Samples (RDS) was developed to address noisy samples, working with SLA. To demonstrate the effectiveness of the technique, RDS was applied to the Co-teaching+ and Jocor models, resulting in two new models, RDS-C and RDS-J. The results indicate gains of up to 6 percent in test metrics for both models. A third model, named RDS-Contrastive, was also developed, surpassing the state-of-the-art by up to 4 percent in test accuracy. Furthermore, this work extended the SLA technique to the multilabel scenario, leading to the development of the SLA Multilabel (SLAM) technique. With this technique, two additional models for the multilabel scenario with noisy samples were developed. The models proposed in this work for the multiclass scenario were applied in a real-world environmental solution, while the models developed for the multilabel scenario were implemented as a solution for a real problem in the oil and gas industry.

Insolvenzprognosen und Ratings sind wichtige Aufgaben der Finanzbranche und dienen der Kreditwürdigkeitsprüfung von Unternehmen. Eine Möglichkeit dieses Aufgabenfeld anzugehen, ist maschinelles Lernen. Dabei werden Vorhersagemodelle aufgrund von Beispieldaten aufgestellt. Methoden aus diesem Bereich sind aufgrund Ihrer Automatisierbarkeit vorteilhaft. Dies macht menschliche Expertise in den meisten Fällen überflüssig und bietet dadurch einen höheren Grad an Objektivität. Allerdings sind auch diese Ansätze nicht perfekt und können deshalb menschliche Expertise nicht gänzlich ersetzen. Sie bieten sich aber als Entscheidungshilfen an und können als solche von Experten genutzt werden, weshalb interpretierbare Modelle wünschenswert sind. Leider bieten nur wenige Lernalgorithmen interpretierbare Modelle. Darüber hinaus sind einige Aufgaben wie z.B. Rating häufig Mehrklassenprobleme. Mehrklassenklassifikationen werden häufig durch Meta-Algorithmen erreicht, welche mehrere binäre Algorithmen trainieren. Die meisten der üblicherweise verwendeten Meta-Algorithmen eliminieren jedoch eine gegebenenfalls vorhandene Interpretierbarkeit. In dieser Dissertation untersuchen wir die Vorhersagegenauigkeit von interpretierbaren Modellen im Vergleich zu nicht interpretierbaren Modellen für Insolvenzprognosen und Ratings. Wir verwenden disjunktive Normalformen und Entscheidungsbäume mit Schwellwerten von Finanzkennzahlen als interpretierbare Modelle. Als nicht interpretierbare Modelle werden Random Forests, künstliche Neuronale Netze und Support Vector Machines verwendet. Darüber hinaus haben wir einen eigenen Lernalgorithmus Thresholder entwickelt, welcher disjunktive Normalformen und interpretierbare Mehrklassenmodelle generiert. Für die Aufgabe der Insolvenzprognose zeigen wir, dass interpretierbare Modelle den nicht interpretierbaren Modellen nicht unterlegen sind. Dazu wird in einer ersten Fallstudie eine in der Praxis verwendete Datenbank mit Jahresabschlüssen von 5152 Unternehmen verwendet, um die Vorhersagegenauigkeit aller oben genannter Modelle zu messen. In einer zweiten Fallstudie zur Vorhersage von Ratings demonstrieren wir, dass interpretierbare Modelle den nicht interpretierbaren Modellen sogar überlegen sind. Die Vorhersagegenauigkeit aller Modelle wird anhand von drei in der Praxis verwendeten Datensätzen bestimmt, welche jeweils drei Ratingklassen aufweisen. In den Fallstudien vergleichen wir verschiedene interpretierbare Ansätze bezüglich deren Modellgrößen und der Form der Interpretierbarkeit. Wir präsentieren exemplarische Modelle, welche auf den entsprechenden Datensätzen basieren und bieten dafür Interpretationsansätze an. Unsere Ergebnisse zeigen, dass interpretierbare, schwellwertbasierte Modelle den Klassifikationsproblemen in der Finanzbranche angemessen sind. In diesem Bereich sind sie komplexeren Modellen, wie z.B. den Support Vector Machines, nicht unterlegen. Unser Algorithmus Thresholder erzeugt die kleinsten Modelle während seine Vorhersagegenauigkeit vergleichbar mit den anderen interpretierbaren Modellen bleibt. In unserer Fallstudie zu Rating liefern die interpretierbaren Modelle deutlich bessere Ergebnisse als bei der zur Insolvenzprognose (s. o.). Eine mögliche Erklärung dieser Ergebnisse bietet die Tatsache, dass Ratings im Gegensatz zu Insolvenzen menschengemacht sind. Das bedeutet, dass Ratings auf Entscheidungen von Menschen beruhen, welche in interpretierbaren Regeln, z.B. logischen Verknüpfungen von Schwellwerten, denken. Daher gehen wir davon aus, dass interpretierbare Modelle zu den Problemstellungen passen und diese interpretierbaren Regeln erkennen und abbilden.

No presente trabalho, estudamos alguns sistemas de partículas interagentes que podem ser vistos como modelos simples de fluxo de tráfego, a saber: O Processo de Hammersley-Aldous-Diaconis e o Processo de Exclusão. Exploramos suas representações como modelos de crescimento no plano. Ênfase é dada aos casos em que há mais de um tipo de partícula, aos processos multiclasses e às suas relações com modelos de filas. Analogia entre os modelos é usada para provar os resultados. Por fim, damos uma nova prova para o cálculo da variância assintótica reescalonada do fluxo de partículas de segunda classe no processo de Hammersley multiclasse em equilíbrio. / In the present work we study the following interacting particle systems which can be seen as simple models of traffic flow: The Hammersley-Aldous-Diaconis Process and the Exclusion Process. We explore the related growth models in the plane. Focus is given to cases where there are more than one kind of particles, to the multitype processes and to their relations with queue models. Analogy between the models is used to prove the results. At last, we give a new proof for the calculation of the asimptotic flux of second class particles in the Multiclass Hammersley process in equilibrium.