Recent large-scale mapping of dust continuum emission from star-forming clouds has revealed their hierarchical nature, which includes web-like filamentary structures that often harbor clumpy over-densities where new stars form. Understanding the motions of these structures and how they interact to form stars, however, can only be learned through observations of emission from their molecular gas. Observations of tracers such as ammonia (NH3), in particular, reveal the stability of dense gas structures against forces such as the inward pull of gravity and the outward push of their internal pressure, thus providing insights into whether or not those structures are likely to form stars in the future. Due to recent large-scale ammonia surveys that have mapped both nearby and distant clouds in the Galaxy, it is finally possible to investigate and compare the stability of star-forming structures in different environments. In this dissertation, we utilize ammonia survey data to provide one of the largest investigations to date into the stability of structures in star-forming regions. Dense gas structures have been identified in a self-consistent manner across a variety of star-forming regions and the environmental factors (e.g., the presence or lack of local filaments and heating by local massive stars) most influential to their stability were investigated. The analysis has revealed that dense gas structures identified by ammonia observations in nearby star-forming clouds tend to be gravitationally bound. In high-mass star-forming clouds, however, bound and unbound ammonia structures are equally likely. This result suggests that either gravity is more important to structure stability at the small scales probed in nearby clouds or ammonia is more widespread in high-mass star-forming regions. In addition, a new method to detect and measure emission with multiple velocity components along the line of sight has been developed. Based on convolutional neural networks and named Convnet Line-fitting Of Emission-line Regions (CLOVER), the method is markedly faster than traditional analysis techniques, requires no input assumptions about the emission, and has demonstrated high classification accuracy. Since high-mass star-forming regions are often plagued by multiple velocity components along the line of sight, CLOVER will improve the accuracy of stability measurements for many clouds of interest to the star formation community. / Graduate

Identifying new potential treatment options for medical conditions that cause human disease burden is a central task of biomedical research. Since all candidate drugs cannot be tested with animal and clinical trials, in vitro approaches are first attempted to identify promising candidates. Likewise, identifying other essential relations (e.g., causation, prevention) between biomedical entities is also critical to understand biomedical processes. Hence, it is crucial to develop automated relation prediction systems that can yield plausible biomedical relations to expedite the discovery process. In this dissertation, we demonstrate three approaches to predict treatment relations between biomedical entities for the drug repositioning task using existing biomedical knowledge bases. Our approaches can be broadly labeled as link prediction or knowledge base completion in computer science literature. Specifically, first we investigate the predictive power of graph paths connecting entities in the publicly available biomedical knowledge base, SemMedDB (the entities and relations constitute a large knowledge graph as a whole). To that end, we build logistic regression models utilizing semantic graph pattern features extracted from the SemMedDB to predict treatment and causative relations in Unified Medical Language System (UMLS) Metathesaurus. Second, we study matrix and tensor factorization algorithms for predicting drug repositioning pairs in repoDB, a general purpose gold standard database of approved and failed drug–disease indications. The idea here is to predict repoDB pairs by approximating the given input matrix/tensor structure where the value of a cell represents the existence of a relation coming from SemMedDB and UMLS knowledge bases. The essential goal is to predict the test pairs that have a blank cell in the input matrix/tensor based on the shared biomedical context among existing non-blank cells. Our final approach involves graph convolutional neural networks where entities and relation types are embedded in a vector space involving neighborhood information. Basically, we minimize an objective function to guide our model to concept/relation embeddings such that distance scores for positive relation pairs are lower than those for the negative ones. Overall, our results demonstrate that recent link prediction methods applied to automatically curated, and hence imprecise, knowledge bases can nevertheless result in high accuracy drug candidate prediction with appropriate configuration of both the methods and datasets used.

The aim of this paper is to investigate if rainfall prediction (nowcasting) can successively be made using a deep learning approach. The input to the networks are different spatiotemporal variables including forecasts from a NWP model. The results indicate that these networks has some predictive power and could be use in real application. Another interesting empirical finding relates to the usage of transfer learning from a domain which is not related instead of random initialization. Using pretrained parameters resulted in better convergence and overall performance than random initialization of the parameters.

Automatic morpheme segmentation as a field has been dominated by unsupervised methods since its inception. Partly due to theoretical motivations, but also due to resource constraints. Given the success neural network methods have shown on a wide variety of field in later years, it would seem compelling to apply these methods to the morpheme segmentation field. This study explores the efficacy of modern neural networks, specifically convolutional neural networks and Bi-directional LSTM networks, on the morpheme segmentation task in a resource low setting to determine their viability as contenders with previous unsupervised, minimally supervised, and semi-supervised systems in the field. One architecture of each type is implemented and trained on a new gold standard data set and the results are compared to previously established methods. A qualitative error analysis of the architectures’ segmentations is also performed. The study demonstrates that a BLSTM system can be trained with minimal effort to produce a proof of concept solution at low levels of training data and suggests that BLSTM methods may be a fruitful direction for further research in this field.

Orientador: Prof. Dr. Fabrício Olivetti de França / Dissertação (mestrado) - Universidade Federal do ABC, Programa de Pós-Graduação em Ciência da Computação, 2017. / Go é um jogo de tabuleiro que chama muita atenção na área de Inteligência Artificial, por ser um problema complexo de resolver e precisar de diferentes estratégias para obter um bom nível de habilidade no jogo. Até 2015, todos os melhores programas de Go precisavam começar a partida com vantagem para poder ganhar de um jogador profissional, mas no final de 2015, o programa AlphaGo foi o primeiro e único até o momento capaz de vencer um jogador profissional sem precisar de vantagem, combinando o uso de redes neurais convolucionais profundas para direcionar as buscas em árvores de Monte-Carlo. Esta dissertação tem como objetivo principal criar um agente inteligente de Go que decide seus próximos movimentoscom base no cenário atual do tabuleiro e em modelos de predição criados para três estratégias específicas do jogo. Para isso, duas hipóteses foram testadas: i) é possívelespecializar agentes inteligentes para o aprendizado de estratégias parciais do jogo de Go, ii) a combinação dessas estratégias permitem a construção de um agente inteligente para o jogo de Go. Para a primeira hipótese um agente foi treinado para aprender, com base em um jogador heurístico e posteriormente com base nos melhores agentes treinados, a posicionar as pedras para permitir a expansão do território, este agente aprendeu a generalizar esta estratégia contra os indivíduos treinados em diferentes estágios e também a capturar pedras. Também foram treinados dois agentes com base na resolução de problemas, com objetivo de aprenderem as estratégias específicas de captura e defesa das pedras. Em ambos os treinamentos foi possível notar que o conhecimento para resolver um problema era propagado para as próximas gerações de indivíduos, mas o nível de aprendizado foi baixo devido ao pouco treinamento. Para a segunda hipótese, um agente foi treinado para decidir qual das três estratégias específicas utilizar de acordo com o estado atual do tabuleiro. Foi possível constatar que este agente, jogando contra outros indivíduos da população, evoluiu na escolha de melhores estratégias, permitindo a dominação de territórios, captura e defensa das pedras. Os agentes foram criados utilizando Redes Neurais Convolucionais, sem qualquer conhecimento prévio sobre como jogar Go, e o treinamento foi feito com Neuroevolução. Como resultado foi possível perceber a evolução dos agentes para aprender as estratégias e comportamentos distintos de forma segmentada. O nível do agente inteligente gerado ainda está distante de um jogador profissional, porém ainda existem opções de melhorias para serem testadas com parametrização, reformulação da função de aptidão, entre outros. Esses resultados propõem novas possibilidades para a criação de agentes inteligentes para jogos complexos. / Go is a board game that draws a lot of attention in the Artificial Intelligence area, because it is a complex problem to solve and needs different strategies in order to obtain a good skill level in the game. By 2015, all the Go¿s best programs must start the match with advantage to win over a professional player, but in the end of 2015, the AlphaGo program was the first and, so far, the only one capable of beating a professional player without needing advantage, combining the use of deep convolutional neural networks to orientate the searches on Monte-Carlo trees. This dissertation has as main objective to create an intelligent agent of Go that decides its next movements based on current scenario of the board and in prediction models created for three specific strategies of the game. For this purpose, two hypothesis were tested: i) whether it is possible to specialize intelligent agents to learn partial strategies of Go game, ii) whether the combination of these strategies allows the construction of an intelligent agent to play Go. For the first hyphotesis, an agent was trained to learn, based on matches again a heuristic player and later based on the best trained agents, to position the stones to allow the expansion of territory, this agent learn to generalize this strategy against individuals trained in different stages and capture stones too. Two agents were also trained based on problem solving, in order to learn the specific strategies of catching and defense of stones. In both trainings were possible to note that the knowledge to solve a problem was propagated to the next generations of individuals, but the level of learning was low due to the short training. For the second hyphotesis, an agent was trained to decide which of the three specific strategies to use according to the current state of the board. It was possible to verify that this agent, playing against other individuals population, evolved in choosing better strategies, allowing territories domination, capture and defend stones. The agents was created using Convolution Neural Networks, without any previous knowledge about how to play Go, and the training was performed using Neuroevolution. As a result, it was possible to perceive the evolution of agents to learn different strategies and behaviours in a segmented way. The intelligent agent generated¿s skill still far from a professional player, however there are still options for improvement to be tested with parameterization, reformulation of fitness function, and others. These results propose new opportunities for the creation of intelligent agents for complex games.

The commercially used automated number plate recognition (ANPR) sys- tems constitute a mature technology which relies on dedicated industrial cam- eras capable of capturing high-quality still images. In contrast, the problem of ANPR from low-quality video sequences has been so far severely under- explored. This thesis proposes a trainable convolutional neural network (CNN) with a novel architecture which can efficiently recognize number plates from low-quality videos of arbitrary length. The proposed network is experimentally shown to outperform several existing approaches dealing with video-sequences, state-of-the-art commercial ANPR system as well as the human ability to recog- nize number plates from low-resolution images. The second contribution of the thesis is a semi-automatic pipeline which was used to create a novel database containing annotated sequences of challenging low-resolution number plate im- ages. The third contribution is a novel CNN based generator of super-resolution number plate images. The generator translates the input low-resolution image into its high-quality counterpart which preserves the structure of the input and depicts the same string which was previously predicted from a video-sequence. 1

A disponibilidade de conteúdo gerado por usuários em sites colaborativos de perguntas e respostas tem impulsionado o avanço de modelos de Question Answering (QA) baseados em reúso. Essa abordagem pode ser implementada por meio da tarefa de seleção de respostas (Answer Selection, AS), que consiste em encontrar a melhor resposta para uma dada pergunta em um conjunto pré-selecionado de respostas candidatas. Nos últimos anos, abordagens baseadas em vetores distribucionais e em redes neurais profundas, em particular em redes neurais convolutivas (CNNs), têm apresentado bons resultados na tarefa de AS. Contudo, a maioria dos modelos é avaliada sobre córpus de perguntas objetivas e bem formadas, contendo poucas palavras. Raramente estruturas textuais complexas são consideradas. Perguntas de consumidores, comuns em sites colaborativos, podem ser bastante complexas. Em geral, são representadas por múltiplas frases inter-relacionadas, que apresentam pouca objetividade, vocabulário leigo e, frequentemente, contêm informações em excesso. Essas características aumentam a dificuldade da tarefa de AS. Neste trabalho, propomos um modelo de seleção de respostas para perguntas de consumidores. São contribuições deste trabalho: (i) uma definição para o objeto de pesquisa perguntas de consumidores; (ii) um novo dataset desse tipo de pergunta, chamado MilkQA; e (iii) um modelo de seleção de respostas, chamado SlimRank. O MilkQA foi criado a partir de um arquivo de perguntas e respostas coletadas pelo serviço de atendimento de uma renomada instituição pública de pesquisa agropecuária (Embrapa). Anotadores guiados pela definição de perguntas de consumidores proposta neste trabalho selecionaram 2,6 mil pares de perguntas e respostas contidas nesse arquivo. A análise dessas perguntas levou ao desenvolvimento do modelo SlimRank, que combina representação de textos na forma de grafos semânticos com arquiteturas de CNNs. O SlimRank foi avaliado no dataset MilkQA e comparado com baselines e dois modelos do estado da arte. Os resultados alcançados pelo SlimRank foram bastante superiores aos resultados dos baselines, e compatíveis com resultados de modelos do estado da arte; porém, com uma significativa redução do tempo computacional. Acreditamos que a representação de textos na forma de grafos semânticos combinada com CNNs seja uma abordagem promissora para o tratamento dos desafios impostos pelas características singulares das perguntas de consumidores. / The increasing availability of user-generated content in community Q&A sites has led to the advancement of Question Answering (QA) models that relies on reuse. Such approach can be implemented by the task of Answer Selection (AS), which consists in finding the best answer for a given question in a pre-selected pool candidate answers. Recently, good results have been achieved by AS models based on distributed word vectors and deep neural networks that are used to rank answers for a given question. Convolutinal Neural Networks (CNNs) are particularly succesful in this task. Most of the AS models are built over datasets that contains only short and objective questions expressed as interrogative sentences containing few words. Complex text structures are rarely considered. However, consumer questions may be really complex. This kind of question is the main form of seeking information in community Q&A sites, forums and customer services. Consumer questions have characteristics that increase the difficulty of the answer selection task. In general, they are composed of multiple interrelated sentences that are usually subjective, and contains laymans terms and excess of details that may be not particulary relevant. In this work, we propose an answer selection model for consumer questions. Specifically the contributions of this work are: (i) a definition for the consumer questions research object; (ii) a new dataset of this kind of question, which we call MilkQA; and (iii) an answer selection model, named SlimRank. MilkQA was created from an archive of questions and answers collected by the customer service of a well-known public agricultural research institution (Embrapa). It contains 2.6 thousand question-answer pairs selected and anonymized by human annotators guided by the definition proposed in this work. The analysis of questions in MilkQA led to the development of SlimRank, which combines semantic textual graphs with CNN architectures. SlimRank was evaluated on MilkQA and compared to baselines and two state-of-the-art answer selection models. The results achieved by our model were much higher than the baselines and comparable to the state of the art, but with significant reduction of computational time. Our results suggest that combining semantic text graphs with convolutional neural networks are a promising approach for dealing with the challenges imposed by consumer questions unique characteristics.

Sentiment analysis is a field within machine learning that focus on determine the contextual polarity of subjective information. It is a technique that can be used to analyze the "voice of the customer" and has been applied with success for the English language for opinionated information such as customer reviews, political opinions and social media data. A major problem regarding machine learning models is that they are domain dependent and will therefore not perform well for other domains. Transfer learning or domain adaption is a research field that study a model's ability of transferring knowledge across domains. In the extreme case a model will train on data from one domain, the source domain, and try to make accurate predictions on data from another domain, the target domain. The deep machine learning model Convolutional Neural Network (CNN) has in recent years gained much attention due to its performance in computer vision both for in-domain classification and transfer learning. It has also performed well for natural language processing problems but has not been investigated to the same extent for transfer learning within this area. The purpose of this thesis has been to investigate how well suited the CNN is for cross-domain sentiment analysis of Swedish reviews. The research has been conducted by investigating how the model perform when trained with data from different domains with varying amount of source and target data. Additionally, the impact on the model’s transferability when using different text representation has also been studied. This study has shown that a CNN without pre-trained word embedding is not that well suited for transfer learning since it performs worse than a traditional logistic regression model. Substituting 20% of source training data with target data can in many of the test cases boost the performance with 7-8% both for the logistic regression and the CNN model. Using pre-trained word embedding produced by a word2vec model increases the CNN's transferability as well as the in-domain performance and outperform the logistic regression model and the CNN model without pre-trained word embedding in the majority of test cases.

Title: Deep neural networks and their application for economic data processing Author: Bc. Tomáš Witzany Department: Department of Theoretical Computer Science and Mathematical Logic Supervisor: Doc. RNDr. Iveta Mrázová, CSc., Department of Theoretical Com- puter Science and Mathematical Logic Abstract: Analysis of macroeconomic time-series is key for the informed decisions of national policy makers. Economic analysis has a rich history, however when considering modeling non-linear dependencies there are many unresolved issues in this field. One of the possible tools for time-series analysis are machine learn- ing methods. Of these methods, neural networks are one of the commonly used methods to model non-linear dependencies. This work studies different types of deep neural networks and their applicability for different analysis tasks, including GDP prediction and country classification. The studied models include multi- layered neural networks, LSTM networks, convolutional networks and Kohonen maps. Historical data of the macroeconomic development across over 190 differ- ent countries over the past fifty years is presented and analysed. This data is then used to train various models using the mentioned machine learning methods. To run the experiments we used the services of the computer center MetaCentrum....

Research in the Arctic is of ever growing importance, and modern technology is used in news ways to map and understand this very complex region and how it is effected by climate change. Here, animals and vegetation are tightly coupled with their environment in a fragile ecosystem, and when the environment undergo rapid changes it risks damaging these ecosystems severely.  Understanding what kind of data that has potential to be used in artificial intelligence, can be of importance as many research stations have data archives from decades of work in the Arctic. In this thesis, a YOLOv4 object detection model has been trained on two classes of images to investigate the performance impacts of disturbances in the training data set. An expanded data set was created by augmenting the initial data to contain various disturbances. A model was successfully trained on the augmented data set and a correlation between worse performance and presence of noise was detected, but changes in saturation and altered colour levels seemed to have less impact than expected. Reducing noise in gathered data is seemingly of greater importance than enhancing images with lacking colour levels. Further investigations with a larger and more thoroughly processed data set is required to gain a clearer picture of the impact of the various disturbances.