Health care represents an increasing proportion of global consumption. We discuss ways to model health care utilization on an individual basis. We present a probabilistic, generative model of utilization. Leveraging previously observed utilization levels, we learn a latent structure that can be used to accurately understand risk and make predictions. We evaluate the effectiveness of the model using data from a large population.

This thesis analyzes the changing design and planning strategies in the contemporary urban design area. The rapid improvements during the 20th century in complexity sciences and computer technologies have directly affected all the branches of design. In architecture, as in urban design, generative models, evolutionary design attitudes and computer based simulation tools have taken a significant role during the last few decades. In urban design, emerged in a period starting form the second half of the century, non-determinist, dynamic and self-organized design attitudes depending on naturalistic models have emerged as an alternative to determinist, static and reductionist approaches based on linear solutions. In this study, it is aimed to define and evaluate these emerging contemporary approaches with respect to their antecedents and precedents. The study also searches for the conceptual and technical developments and background which support this process. With an analysis of case studies, the paradigm shift is examined in practice. The study intends to clarify whether contemporary urban design approaches, especially naturalistic models could be an alternative to deterministic stances.

As Internet communities such as question-answer (Q&A) forums and online social networks (OSNs) grow in prominence as knowledge sources, traditional editorial filters are unable to scale to their size and pace. This absence hinders the exchange of knowledge online, by creating an understandable lack of trust in information. This mistrust can be partially overcome by a forum by consistently providing reliable information, thus establishing itself as a reliable source. This work investigates how algorithmic approaches can contribute to building such a community of voluntary experts willing to contribute authoritative information. This work identifies two approaches: a) reducing the cost of participation for experts via matching user queries to experts (question recommendation), and b) identifying authoritative contributors for incentivization (authority estimation). The question recommendation problem is addressed by extending existing approaches via a new generative model that augments textual data with expert preference information among different questions. Another contribution to this domain is the introduction of a set of formalized metrics to include the expert's experience besides the questioner's. This is essential for expert retention in a voluntary community, and has not been addressed by previous work. The authority estimation problem is addressed by observing that the global graph structure of user interactions, results from two factors: a user's performance in local one-to-one interactions, and their activity levels. By positing an intrinsic authority 'strength' for each user node in the graph that governs the outcome of individual interactions via the Bradley-Terry model for pairwise comparison, this research establishes a relationship between intrinsic user authority, and global measures of influence. This approach overcomes many drawbacks of current measures of node importance in OSNs by naturally correcting for user activity levels, and providing an explanation for the frequent disconnect between real world reputation and online influence. Also, while existing research has been restricted to node ranking on a single OSN graph, this work demonstrates that co-ranking across multiple endorsement graphs drawn from the same OSN is a highly effective approach for aggregating complementary graph information. A new scalable co-ranking framework is introduced for this task. The resulting algorithms are evaluated on data from various online communities, and empirically shown to outperform existing approaches by a large margin. / text

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

During the last decade, there is an increasing interest in applying deep learning in de novo drug design. In this thesis, a tool is developed to address the specific needs for generating small library for lead optimization. The optimization of small molecules is conducted given an input scaffold with defined attachment points. Various chemical fragments are proposed by the generative model and reinforcement learning is used to guide the generation to produce a library of molecules that satisfy user-defined properties. The generation is also constrained to follow user-defined reactions which makes synthesis controllable. Several experiments are executed to find the optimal hyperparameters, make comparison of different learning strategies, demonstrate the superiority of slicing molecules based on defined reactions compared to RECAP rules, showcase the model’s ability to follow different synthetic routes as well as its capability of decorating scaffolds with various attachment points. Results have shown that DAP learning strategy outperforms all other learning strategies. The use of reaction based slicing is superior than utilising RECAP rules slicing, it helps the model to learn the reaction filter faster. Also, the model was capable of satisfying different reaction filters and decorating scaffolds with various attachment points. In conclusion, the model is able to rapidly generate a molecular library which contains a large number of molecules sharing the same scaffold, with desirable properties and can be synthesised under specified reactions.

Mastitis is the most common disease affecting Swedish milk cows. Automatic image classification can be useful for quickly classifying the bacteria causing this inflammation, in turn making it possible to start treatment more quickly. However, training an automatic classifier relies on the availability of data. Data collection can be a slow process, and GANs are a promising way to generate synthetic data to add plausible samples to an existing data set. The purpose of this thesis is to explore the usefulness of GANs for generating images of bacteria. This was done through researching existing literature on the subject, implementing a GAN, and evaluating the generated images. A cGAN capable of generating class-specific bacteria was implemented and improvements upon it made. The images generated by the cGAN were evaluated using visual examination, rapid scene categorization, and an expert interview regarding the generated images. While the cGAN was able to replicate certain features in the real images, it fails in crucial aspects such as symmetry and detail. It is possible that other GAN variants may be better suited to the task. Lastly, the results highlight the challenges of evaluating GANs with current evaluation methods.

京都大学 / 0048 / 新制・課程博士 / 博士(情報学) / 甲第20735号 / 情博第649号 / 新制||情||112(附属図書館) / 京都大学大学院情報学研究科知能情報学専攻 / (主査)准教授 川嶋 宏彰, 教授 河原 達也, 教授 熊田 孝恒 / 学位規則第4条第1項該当 / Doctor of Informatics / Kyoto University / DFAM

Combitech AB was interested in how to automate the process of annotating aluminum scrap when it was adjacent to other metals. This was to ultimately create an annotated dataset that could be utilized for training a segmentation model. The idea was to make use of generative models to generate samples of general scrap metals. Then, with this model, introduce a small dataset of only aluminum, to try to change the features into a domain suitable for aluminum. Since the contents of the samples were generated separately, the system would know where the aluminum was and could then annotate it.  This master's thesis aimed to investigate whether it was possible to construct generative models to generate these samples and see if they had realistic characteristics. It was also investigated if it was possible to get a meaningful model based on a relatively small dataset (aluminum in this case). The data used were two datasets, one with general scrap metal (excluding aluminum) and the other containing only aluminum scrap. Unconditional diffusion models were utilized as generative models. The scrap model achieved satisfactory results, making it possible to generate samples that carried similar properties as the real scrap dataset. When it came to aluminum, which had a much smaller dataset than the scrap dataset, it was possible to get promising results when utilizing transfer learning. However, the same good quality as the scrap model gave was not achieved. This master's thesis has shown that it is possible to get a model to generate realistic-looking images of scrap metal. Furthermore, this scrap model served as a good base when training other generative models to generate images of metals, even if the provided datasets were small. In this way, a foundation was laid for an investigation of an automatic annotation system.

京都大学 / 新制・課程博士 / 博士(情報学) / 甲第24725号 / 情博第813号 / 新制||情||137(附属図書館) / 京都大学大学院情報学研究科知能情報学専攻 / (主査)教授 鹿島 久嗣, 教授 山本 章博, 教授 吉川 正俊 / 学位規則第4条第1項該当 / Doctor of Informatics / Kyoto University / DFAM

We introduce a Lennard-Jones layer (LJL) to equalize the density across the distribution of 2D and 3D point clouds by systematically rearranging points without destroying their overall structure (distribution normalization). LJL simulates a dissipative process of repulsive and weakly attractive interactions between individual points by solely considering the nearest neighbor of each point at a given moment in time. This pushes the particles into a potential valley, reaching a well-defined stable configuration that approximates an equidistant sampling after the stabilization process. We apply LJLs to redistribute randomly generated point clouds into a randomized uniform distribution over the 2D Euclidean plane and 3D mesh surfaces. Moreover, LJLs are embedded in point cloud generative network architectures by adding them at later stages of the inference process. The improvements coming with LJLs for generating 3D point clouds are evaluated qualitatively and quantitatively. Finally, we apply LJLs to improve the point distribution of a score-based 3D point cloud denoising network. In general, we demonstrate that LJLs are effective for distribution normalization which can be applied at negligible cost without retraining the given neural networks.