Browsing by master's degree program "Tietojenkäsittelytieteen maisteriohjelma"

Computing the permanent of a matrix is a famous #P-hard problem with a wide range of applications. The fastest known exact algorithms for the problem require an exponential number of operations, and all known fully polynomial randomized approximation schemes are rather complicated to implement and have impractical time complexities. The most promising recent advancements on approximating the permanent are based on rejection sampling and upper bounds for the permanent. In this thesis, we improve the current state of the art by developing the deep rejection sampling method, which combines an exact algorithm with the rejection sampling method. The algorithm precomputes a dynamic programming table that tightens the initial upper bound used by the rejection sampling method. In a sense, the table is used to jump-start the sampling process. We give a high probability upper bound for the time complexity of the deep rejection sampling method for random (0, 1)-matrices in which each entry is 1 with probability p. For matrices with p < 1/5, our high probability bound is stronger than in previous work. In addition to that, we empirically observe that our algorithm outperforms earlier rejection sampling methods by testing it with different parameters against other algorithms on multiple classes of matrices. The improvements in sampling times are especially notable in cases in which the ratios of the permanental upper bounds and the exact value of the permanent are huge.

The multi-armed bandit is a sequential decision making problem where an agent chooses actions and receives rewards. The agent faces an explore-exploit dilemma: it has to balance exploring its options to find the optimal actions, and exploiting choosing the empirically best actions. This problem can also be solved by multiple agents who collaborate in a federated learning setting, where agents do not share their raw data samples. Instead, small updates containing learned parameters are shared. In this setting, the learning process can happen with a central server that coordinates the agents to learn the global model, or in a fully decentralized fashion where agents communicate with each other to collaborate. The distribution of rewards may be heterogeneous, meaning that the agents face distributions with local biases. Depending on the context, this can be handled by cancelling the biases by averaging, or by personalizing the global model to fit each individual agent’s local biases. Another common characteristic of federated multi-armed bandits is preserving privacy. Even though only parameter updates are shared, they can be used to infer the original data. To privatize the data, a method known as differential privacy is applied by adding enough random noise to mask the effect of a single contribution. The newest area of interest for federated multi-armed bandits is security. Collaboration between multiple agents means more opportunities for attacks. Achieving robust security means defending against Byzantine attacks that inject arbitrary data into the learning process to affect the model accuracy in an undesirable way. This thesis is a literature review that explores how the federated multi-armed bandit problem is solved, and how privacy and security for it is achieved.

Automatic readability assessment is considered as a challenging task in NLP due to its high degree of subjectivity. The majority prior work in assessing readability has focused on identifying the level of education necessary for comprehension without the consideration of text quality, i.e., how naturally the text flows from the perspective of a native speaker. Therefore, in this thesis, we aim to use language models, trained on well-written prose, to measure not only text readability in terms of comprehension but text quality. In this thesis, we developed two word-level metrics based on the concordance of article text with predictions made using language models to assess text readability and quality. We evaluate both metrics on a set of corpora used for readability assessment or automated essay scoring (AES) by measuring the correlation between scores assigned by our metrics and human raters. According to the experimental results, our metrics are strongly correlated with text quality, which achieve 0.4-0.6 correlations on 7 out of 9 datasets. We demonstrate that GPT-2 surpasses other language models, including the bigram model, LSTM, and bidirectional LSTM, on the task of estimating text quality in a zero-shot setting, and GPT-2 perplexity-based measure is a reasonable indicator for text quality evaluation.

Context: Over the past years, the development of machine learning (ML) enabled software has seen a rise in popularity. Alongside this trend, new challenges have been identified, such as growing concerns about the use, including the ethical concerns, of ML models, as misuse can lead to severe consequences for human beings. To alleviate this problem, more comprehensive model documentation has been suggested, but how can that documentation be made part of a modern, continuous development process? Objective: We design and develop a solution, which consists of a software artefact and its surrounding process, which enables and moderates continuous documentation of ML models. The solution needs to comply with the modern way-of-working of software development. Method: We apply the design science research methodology to divide the design and development into six separate tasks, i.e., problem identification, objective definition, design and development, demonstration, evaluation, and communication. Results: The solution uses model cards for storing model details. These model cards are tested automatically and manually, forming a quality gate and ensuring integrity of the documentation. The software artefact is implemented in the form of a GitHub Action. Conclusion: We conclude that the software artefact supports and assures proper model documentation in the form of a model card. The artefact allows for customization by the user, thereby supporting domain-specific use cases.

Artificial Intelligence (AI) has revolutionized various domains of software development, promising solutions that can adapt and learn. However, the rise of AI systems has also been accompanied by ethical concerns, primarily related to the unintentional biases these systems can inherit during the development process. This thesis presents a thematic literature review aiming to identify and examine the existing methodologies and strategies for preventing bias in iterative AI software development. Methods employed for this review include a formal search strategy using defined inclusion and exclusion criteria, and a systematic process for article sourcing, quality assessment, and data collection. 29 articles were analyzed, resulting in the identification of eight major themes concerning AI bias mitigation within iterative software development, ranging from bias in data and algorithmic processes to fairness and equity in algorithmic design. Findings indicate that while various approaches for bias mitigation exist, gaps remain. These include the need for adapting strategies to agile or iterative frameworks, resolving the trade-off between effectiveness and fairness, understanding the complexities of bias for tailored solutions, and assessing the real-world applicability of these techniques. This synthesis of key trends and insights highlights these specific areas requiring further research.

Anomaly detection in images is the machine learning task of classifying inputs as normal or anomalous. Anomaly localization is the related task of segmenting input images into normal and anomalous regions. The output of an anomaly localization model is a 2D array, called an anomaly map, of pixel-level anomaly scores. For example, an anomaly localization model trained on images of non-defective industrial products should output high anomaly scores in image regions corresponding to visible defects. In unsupervised anomaly localization the model is trained solely on normal data, i.e. without labelled training observations that contain anomalies. This is often necessary as anomalous observations may be hard to obtain in sufficient quantities and labelling them is time-consuming and costly. Student-teacher feature pyramid matching (STFPM) is a recent and powerful method for unsupervised anomaly detection and localization that uses a pair of convolutional neural networks of identical architecture. In this thesis we propose two methods of augmenting STFPM to produce better segmentations. Our first method, discrepancy scaling, significantly improves the segmentation performance of STFPM by leveraging pre-calculated statistics containing information about the model’s behaviour on normal data. Our second method, student-teacher model assisted segmentation, uses a frozen STFPM model as a feature detector for a segmentation model which is then trained on data with artificially generated anomalies. Using this second method we are able to produce sharper anomaly maps for which it is easier to set a threshold value that produces good segmentations. Finally, we propose the concept of expected goodness of segmentation, a way of assessing the performance of unsupervised anomaly localization models that, in contrast to current metrics, explicitly takes into account the fact that a segmentation threshold needs to be set. Our primary method, discrepancy scaling, improves segmentation AUROC on the MVTec AD dataset over the base model by 13%, measured in the shrinkage of the residual (1.0 − AUROC). On the image-level anomaly detection task, a variant of the discrepancy scaling method improves performance by 12%.

Fault management in mobile networks is required for detecting, analysing, and fixing problems appearing in the mobile network. When a large problem appears in the mobile network, multiple alarms are generated from the network elements. Traditionally Network Operations Center (NOC) process the reported failures, create trouble tickets for problems, and perform a root cause analysis. However, alarms do not reveal the root cause of the failure, and the correlation of alarms is often complicated to determine. If the network operator can correlate alarms and manage clustered groups of alarms instead of separate ones, it saves costs, preserves the availability of the mobile network, and improves the quality of service. Operators may have several electricity providers and the network topology is not correlated with the electricity topology. Additionally, network sites and other network elements are not evenly distributed across the network. Hence, we investigate the suitability of a density-based clustering methods to detect mass outages and perform alarm correlation to reduce the amount of created trouble tickets. This thesis focuses on assisting the root cause analysis and detecting correlated power and transmission failures in the mobile network. We implement a Mass Outage Detection Service and form a custom density-based algorithm. Our service performs alarm correlation and creates clusters of possible power and transmission mass outage alarms. We have filed a patent application based on the work done in this thesis. Our results show that we are able to detect mass outages in real time from the data streams. The results also show that detected clusters reduce the number of created trouble tickets and help reduce of the costs of running the network. The number of trouble tickets decreases by 4.7-9.3% for the alarms we process in the service in the tested networks. When we consider only alarms included in the mass outage groups, the reduction is over 75%. Therefore continuing to use, test, and develop implemented Mass Outage Detection Service is beneficial for operators and automated NOC.

Isomorphic web applications combine the best parts of static Hypertext Markup Language (HTML) pages and single-page applications. An isomorphic web application shares code between the server and the client. However, there is not much existing research on isomorphic web applications. Improving the performance, user experience and development experience of web applications are popular research topics in computer science. This thesis studies the benefits and challenges of isomorphism in single-page applications. To study the benefits and challenges of isomorphism in single-page applications, a gray literature review and a case study were conducted. The articles used in the gray literature review were searched from four different websites. To make sure the gray literature could be used in this study, a quality assessment process was conducted. The case study was conducted as a developer survey, where developers familiar with isomorphic web applications were interviewed. The results of both studies are then compared and the key findings are compared together. The results of this study show that isomorphism in single-page applications brings benefits to both the developers and the end-users. Isomorphism in single-page applications is challenging to implement and has some downsides, but they mostly affect developers. The performance and search engine optimization of the application are improved. Implementing isomorphism makes it possible to share code between the server and the client, but it increases the complexity of the application. Framework and library compatibility are issues that must be addressed by the developers. The findings of this thesis give motivation for developers to implement isomorphism when starting a new project or transforming existing single-page applications to use isomorphism.

In this thesis we investigate the feasibility of machine learning methods for estimating the type and the weight of individual food items from images taken of customers’ plates at a buffet- style restaurant. The images were collected in collaboration with the University of Turku and Flavoria, a public lunch-line restaurant, where a camera was mounted above the cashier to automatically take a photo of the foods chosen by the customer when they went to pay. For each image, an existing system of scales at the restaurant provided the weights for each individual food item. We describe suitable model architectures and training setups for the weight estimation and food identification tasks and explain the models’ theoretical background. Furthermore we propose and compare two methods for utilizing a restaurant’s daily menu information for improving model performance in both tasks. We show that the models perform well in comparison to baseline methods and reach accuracy on par with other similar work. Additionally, as the images were captured automatically, in some of the images the food was occluded or blurry, or the image contained sensitive customer information. To address this we present computer vision techniques for preprocessing and filtering the images. We publish the dataset containing the preprocessed images along with the corresponding individual food weights for use in future research. The main results of the project have been published as a peer-reviewed article in the International Conference in Pattern Recognition Systems 2022. The article received the best paper award of the conference.

A case study that studied the performance impact of a node.js component when it was refactored from monolith environment into independent service. The performance study studied the response time of the blocking part of JavaScript code in the component. The non blocking part of the code and the added network overhead from the refactoring were excluded from the performance review. Literature review didn’t show any related research that studied the performance impact of a node.js component when it was refactored from monolith into microservices. Many found studies were found that studied the response time and throughput of REST API build with node.js with comparisons to other programming languages. A study were found that related to refactoring an application from monolith into microservices. None of the found studies were directly related to the studied case. It was noted that the response time of the component improved by 46.5% when it was refactored from monolith into microservice. It is possible that when a node.js monolith application grows it starts to affect the throughput of the event loop affecting performance critical components. For the case component it was beneficial to refactor it into independent service in order to gain the 92.6ms in the mean response time.