Browsing by Author "Kaski, Kimmo, Prof., Aalto University, Department of Computer Science, Finland"

Filter results by typing the first few letters
Now showing 1 - 5 of 5
  • Analysis of Cumulative and Temporal Patterns in Science
    (2017) Della Briotta Parolo, Pietro
     School of Science | Doctoral dissertation (article-based)
    The goal of science has always been to investigate the world and its phenomena, by collecting data from all possible events that take place around us, breaking them down into their most simple elements and trying to come up with models able to explain and predict the outcome of these events. For centuries, the primary focus of science was mainly on natural events, but as the new technologies allowed to gather data from human interactions, it was natural for scientists to use this new information in order to apply the same logic to social systems, including science itself.  Since the late 19th century, when the first modern scientific journals were published, science has seen a constant rise in both its size and productivity, thanks to the standardization of research practices and the building of an international community that actively helps to push forward the limits of human knowledge. As science itself went from being a purely intellectual endeavor to a complex social, economical and political system, it is no surprise that a lot of attention has been dedicate in recent years to the study of the underlying mechanisms of science, aided by the explosion of means of communication that allow collaborations and exchange of information at instant speed across the globe, leaving behind digital traces that provide valuable data to study. The continuous exponential growth of science however, causes also difficulties in analyzing objectively the patterns and statistics that scientific data can reveal: for example a paper from the early 20th century would rarely get more than 100 citations, while now it is not uncommon for publications to pass the 10 thousand citation mark.  This thesis follows these attempts in trying to grasp how science works, by investigating the connections, i.e. citations, that exists between scientific publications and how these connections create structures and patterns. It shows that typical patterns in citation count and diffusion of information between fields is heavily influenced by the rate of growth of science, thus suggesting to use the number of publications as a better measure of time. It shows that there is a lag between breakthrough discoveries and the time when they are recognized, thus suggesting that we might be either running out of discoveries or rather having too much of them, in either case an extreme phenomenon. It shows that the community of publications which builds around an original successful paper has a typical life cycle, with an initial clustering, followed by an inevitable breaking down. Finally, it offers a new way of quantifying the impact of publications across time based on their cumulative impact on the overall corpus of scientific material.
  • Applicability and Robustness of Deep Learning in Healthcare
    (2024) Sahlsten, Jaakko
     School of Science | Doctoral dissertation (article-based)
    The worldwide population is aging causing an increased demand for healthcare, motivating a goal to reduce the burden of health professionals to maintain the expected level of care. Deep learning (DL)-based methods, especially convolutional neural networks (CNNs), have achieved state-of-the-art performance in various classification and segmentation tasks on imaging data. Thus, there is an interest in applying these methods to automate routine, laborious or time-consuming clinical tasks based on medical imaging. However, the conventional DL approaches may not be trustworthy to be used in healthcare due to limited explainability combined with overconfidence and their sensitivity to distribution shifts. In this thesis, the applicability of DL approaches in healthcare are investigated in the clinical tasks of screening and medical image segmentation. The approaches are evaluated for robustness to distribution shifts with in-distribution and out-of-distribution datasets including other imaging centers, devices, and under defacing techniques. In order to improve the lack of explainability and overconfidence, approximate Bayesian neural networks with novel uncertainty measures are applied to the tasks and systematically evaluated in terms of performance and uncertainty quantification. The deep learning paradigm and its practical usage in the investigated medical imaging tasks is first introduced. The following part describes uncertainty quantification in deep learning, its downstream utilization in clinical workflow, and the current approaches of approximate Bayesian deep learning. The next part includes the summary for the included publications and related works. The last part includes the conclusion and discussion about the analysis, its limitations, and proposed future research to improve the trustworthiness and applicability of deep learning techniques for imaging in healthcare. The publications demonstrated that CNN-based DL methods have clinically acceptable performance in the evaluated tasks using in-distribution data. However, the robustness to distribution shift varied depending on the task such as robustness to other imaging devices but sensitivity to defacing in segmentation. In terms of explainability and overconfidence, the approximate Bayesian deep learning and the novel uncertainty measures demonstrated improved utility of uncertainty in comparison to conventional approaches in both tasks.
  • Community detection in complex networks: the role of node metadata
    (2017) Hric, Darko
     School of Science | Doctoral dissertation (article-based)
    Recently, it was recognized that the problems lying between the order and chaos require a new scientific language and models to be developed. Network science has emerged as a promising interdisciplinary field studying the properties of all kinds of systems that emerge from interactions of large number of elements or constituents. A particularly interesting feature of complex networks is the presence of communities, or groups of nodes that have more connections between them than to the rest of the network. Communities provide an insight into the structure of the whole system and the immediate environment of each node, like circles of friends, or functionally related genes, and they have also been shown to play a role in various processes on networks. For these reasons numerous community detection algorithms have been proposed that take the network structure as input and return the communities, the nodes belong to. As the field of community detection matured, more scrutiny was applied to old and new algorithms. The researchers were not satisfied any more with good results on simple, almost toy examples, more proofs were sought for the applicability of the algorithms in the real world. At the same time, larger and more complex network datasets were becoming available, in which the need to identify meso-scale structures was even higher. A straightforward way to test the algorithms is to compare the results with the known node community assignments, which are taken to correspond to metadata labels on the nodes. In the first part of this dissertation a large number of algorithms were tested on a large number of labeled networks from different domains. Weak correspondences between metadata and communities indicate that more care has to be taken when using metadata as community labels. The relationship between the node metadata and communities is perhaps more complex than it was earlier assumed, but this does not mean that it is absent. Second part of this dissertation presents a novel approach for incorporating the metadata into community detection without assuming their usefulness. This approach enables to discriminate between metadata that are aligned with community structure and those that are not. The third part of this dissertation proposes the use of the stochastic blockmodel for modeling the citation networks of journals. The model is able to capture rich structures present in the data, while being simple, intuitive and applicable to huge networks (millions of nodes and links). By splitting the data spanning more that a hundred years into separate time windows, it was possible to track the evolution of science in time, and using the model presented in the previous part of the dissertation, the usefulness of journal classification into subject categories as predictors of the citation flows was evaluated.
  • Human behavioural patterns : A reality mining study
    (2018) Monsivais Velázquez, Daniel
     School of Science | Doctoral dissertation (article-based)
    Mobile phone communication is a source of information for studying human behavioural patterns. A mobile phone can collect information of its usage, communication events and data captured by integrated sensors, and this information has been used for studying mobility, epidemics, health and depression, and information diffusion. Particularly, the call detail records have been used to study different social features, like the network structure, people's sleeping patterns, and response to natural disasters. They provides useful insights about the behaviour of the people involved in the calls. This dissertation is based on four research articles in which a huge data set containing call details records of around 3 million users over a 12-month period in 2007 is analysed, to study the dynamics of the human daily resting periods and human social focus over the life course. Each day, the calling activity of the mobile phone users follows two different circadian rhythms, each one synchronised to a different clock. On one hand there is the clock of social time, marked by social activities of the daily routine, in which the working and schooling times, opening times of offices, etc, set a specific social schedule to follow. On the other hand, some human physiological processes, like the human sleep-wake cycle, follow a natural 24h cycle, entrained to a biological clock. The calling pattern shows the struggle of living between these two clocks. It follows a specific schedule (it peaks and decreases twice each day, showing a strong dependence on social time). The location and size of the peaks of activity change over the year, by expanding during the summer and shrinking during the winter, thus indicating a seasonal dependence. Moreover, people living in the same time zone but at different locations, are found to start (or cease) their activity at different times, with a difference given by their local sun transit times, thus people living eastward in the time zone have earlier schedules than those living westward. The emotional closeness between users and their contacted alters can be determined based on their communication pattern. The level of interaction between a mobile phone user and the alters in his/her egocentric network is different, having a dominant interaction with the romantic partner. The features of the ego-centric network and the social focus invested on alters depends on the age and gender of the user, showing clear changes as the users go through different life course stages. Younger people contact more alters and more frequently but this changes noticeably as the egos cross the parenthood stage, in such a way that when egos reach old age, the size of the egocentric network has considerably decreased and it is mainly populated by alters younger then the ego. At the grand-parenting age, an important gender difference appears, when females (probably crossing menopause) show a strong change in social focus towards their daughters, who are in the reproductive stage, whereas males remain focused mainly on their romantic partners, providing supporting arguments for the grand-mothering hypothesis.
  • Training methods for climate and neural network models
    (2018) Abbas, Mudassar
     School of Science | Doctoral dissertation (article-based)
    When modeling complex phenomena in nature and in technological systems, one is often faced with  the task of tuning/calibrating the models. In such cases, there typically exists a need for model  parameter (and/or meta-parameter) value tuning for more effective modeling performance. Often such cannot be done manually, and in the machine learning approach, the tuning is done in an algorithmic and data-driven manner, and is called model training. The thesis presents studies in which such methods are adopted, in the contexts of climate and artificial neural networks, and proposes novel techniques. One of the studies is on the suitability of a well-known machine learning method called Bayesian optimization (BO), for parametric tuning of chaotic systems such as climate and numerical weatherprediction (NWP) models. The obtained results show that BO is a suitable method for such tuning tasks. A major desiderata for a trained machine learning model is the ability to generalize well to unseen data, and thus the phenomena such as (so-called) under- and overfitting are to be avoided. In this context, adopting (so-called) regularization methods as part of the model training process has become a standard procedure. In this thesis, we introduce a regularization framework that is shown to have close connections with many existing state-of-the-art regularization approaches. An adversarial variant, derived from the proposed regularization framework, is used for solving a classification task, and the obtained results are compared to those of other regularization methods.