Browsing by master's degree program "Master's Programme in Data Science"

Planning and decision making are active areas of research in cognitive neuroscience that strive to explain how the brain makes decisions in complex scenarios. Research in this field has traditionally been restricted to simplistic experiments such as two-alternative forced choice situations, and has relatively recently broken into more naturalistic settings with the help of computational modeling and games. Importantly, these computational models aim to be interpretable, meaning that they are crafted in a way that what each parameter means has a clear meaning, perhaps in contrast to massive neural networks. However, the latter may better capture more complex behaviors that the hand-crafted model could miss, so it may be desirable to use a neural network as a guide or ``oracle'' to study and improve the parameters to include in the interpretable model. In this thesis, we present GPT-4IAR, a transformer neural network architecture for modeling and predicting human behavior in the board game four-in-a-row (4IAR). Building upon previous studies that use fully connected neural networks to improve models around 4IAR, and the excellent capabilities of the GPT architecture in tasks where data is sequential, we train a transformer on millions of games of 4IAR to study biases that arise in human decision making. Experiments show that conditioning action predictions on longer histories of previous moves leads to improved accuracy over prior state-of-the-art models, hinting at longer-term strategic biases in human gameplay. Reaction time prediction is also explored, showing promise in capturing meaningful gameplay statistics beyond raw actions.

Semantic textual similarity (STS), the procedure of determining how similar pieces of text are in terms of their meaning, is an important problem in the rapidly evolving field of natural language processing (NLP). STS accelerates major information retrieval applications dealing with natural language text, such as web search engines. For computational efficiency reasons, text pieces are often encoded into semantically meaningful real-valued vectors, sentence embeddings, that can be compared with similarity metrics. Majority of recent NLP research has focused on a small set of largest Indo-European languages and Chinese. Although much of the research is machine learning oriented and is thus often applicable across languages, languages with lesser speaker population, such as Finnish, often lack annotated data required to train, or even evaluate, complex models. BERT, a language representation framework building on transfer learning, is one of the recent quantum leaps in NLP research. BERT-type models take advantage of unsupervised pre-training reducing annotated data demands for supervised tasks. Furthermore, a BERT modification called Sentence-BERT enables us to extend and train BERT-type models to derive semantically meaningful sentence embeddings. However, yet the annotated data demands for conventional training of a Sentence-BERT is relatively low, often such data is unavailable for low-resourced languages. Multilingual knowledge distillation has been shown to be a working strategy for extending mono- lingual Sentence-BERT models to new languages. This technique allows transferring and merging desired properties of two language models, and, instead of annotated data, consumes bilingual parallel samples. In this thesis we study using knowledge distillation to transfer STS properties learnt from English into a model pre-trained on Finnish while bypassing the lack of annotated Finnish data. Further, we experiment distillation with different types of data, English-Finnish bilingual, English monolingual and random pseudo samples, to observe which properties of training data are really necessary. We acquire a bilingual English-Finnish test dataset by translating an existing annotated English dataset and use this set to evaluate the fit of our resulting models. We evaluate the performance of the models in different tasks, English, Finnish and English-Finnish cross-lingual STS, to observe how well the properties being transferred are captured, and how well the models retain the desired properties they already have. We find that knowledge distillation is indeed a feasible approach for obtaining a relatively high quality Sentence-BERT for Finnish. Surprisingly, in all setups large portion of desired properties are transferred to the Finnish model, and, training with English-Finnish bilingual data yields best Finnish sentence embedding model we are aware of.

Two factors define the success of a deep neural network (DNN) based application; the training data and the model. Nowadays, many state-of-the-art DNN models are available free of charge, and training and deploying these models is easier than ever before. As a result, anyone can set up a state-of-the-art DNN algorithm within days or even hours. In the past, most of the focus has been given to the model when researchers were building faster and more accurate deep learning architectures. These research groups commonly use large and high-quality datasets in their work, which is not the case when one wants to train a new model for a specific use case. Training a DNN algorithm for a specific task requires collecting a vast amount of unlabelled data and then labeling the training data. To train a high-performance model, the labeled training dataset must be large and diverse to cover all relevant scenarios of the intended use case. This thesis will present an efficient and straightforward active learning method to sample the most informative images to train a powerful anchor-free Intersection over Union (IoU) predicting objector detector. Our method only uses classification confidences and IoU predictions to estimate the image informativeness. By collecting the most informative images, we can cover the whole diversity of the images with fewer human-annotated training images. This will save time and resources, as we avoid labeling images that would not be beneficial.

Arvopaperin tulevan hinnanmuodostuksen ennustaminen on mielenkiintoista niin sijoittajan kuin aktiivisesti kauppaa tekevän markkinatoimijan näkökulmasta. Tarpeeksi hyvällä tarkkuudella arvopaperin tulevaa hinnanmuodostusta ennustamalla voi markkinatoimija ostaa arvopaperia ennen sen mahdollista markkinahinnan nousua, tai suojata salkkuaan sitä jo omistaessaan, mikäli on vaara, että arvopaperin markkinahinta laskee ajan mittaan merkittävästi. Tutkielmassani käsittelen koneoppimisen soveltamista tekniseen analyysiin. Tutkin, voidaanko tekniseen analyysiin pohjautuen markkinan tai arvopaperin tulevaa hinnanmuodostusta ennustaa lyhyellä aikavälillä tarpeeksi hyvällä tarkkuudella. Selvitän arvopaperimarkkinoiden toimintaa ja käyn läpi, miten tarkasteltavan markkinan tulevaa kysynnän ja tarjonnan suhdetta voidaan teknistä analyysiä hyödyntäen pyrkiä ennustamaan. Taustoitan myös omassa tutkimuksessa käyttämieni teknisen analyysin indikaattorien sekä koneoppimisen menetelmien toimintaa ja esitän aiempaa tutkimusta ongelman parissa. Havaitsin, että markkinoiden tulevan hinnanmuodostuksen ennustaminen on haastavaa. Käyttämilläni ohjatun oppimisen menetelmillä en onnistunut generoimaan mallia, joka olisi osannut ennustaa S&P 500-osakeindeksille, onko tarkasteltavaa ajanhetkeä seuraavan lyhyen aikavälin päätteeksi markkinahinta korkeammalla vai enintään yhtä korkealla kuin tarkasteluajankohtana. Opetetut mallit saavuttivat parhaimmillaan vain 50.8 − 51.4 % ennustetarkkuuden, kun taas naiivi luokittelija, joka ennustaa jokaisen aikavälin päätteeksi markkinahinnan kohonneen saavuttaa 53.0 %:n tarkkuuden. Vehnäfutuurisopimusmarkkinalle saamani tulokset olivat lupaavampia ja opetetut mallit saavuttivat edellä mainitun ongelmanratkaisuun parhaimmillaan 51.7 − 52.5 % ennustetarkkuuden, joka ylitti naiivin luokittelijan 50.9 % tarkkuuden. Analysoin saamiani tuloksia ja esitin jatkotutkimusmahdollisuuksia mallien tehostamiseksi.

In recent years, significant progress has been made in computer vision regarding object detection and tracking which has allowed the emergence of various applications. These often focus on identifying and tracking people in different environments such as buildings. Detecting people allows us to get a more comprehensive view of people flow as traditional IoT data from elevators cannot track individual people and their trajectories. In this thesis, we concentrate on people detection in elevator lobbies which we can use to improve the efficiency of the elevators and the convenience of the building. We compare the performance and speed of various object detection algorithms. Additionally, we research an edge device's capability to run an object detection model on multiple cameras and whether a single device can cover the target building. We were able to train an object detection algorithm suitable for our application. This allowed accurate people detection that can be used for people counting. We found that out of the three object detection algorithms we trained, YOLOv3 was the only one capable of generalizing to unseen environments, which is essential for general purpose application. The performances of the other two models (SSD and Faster R-CNN) were poor in terms of either accuracy or speed. Based on these, we chose to deploy YOLOv3 to the edge device. We found that the edge device's inference time is linearly dependent on the number of cameras. Therefore, we can conclude that one edge device should be sufficient for our target building, allowing two cameras for each floor. We also demonstrated that the edge device allows easy addition of an object tracking layer, which is required for the solution to work in a real-life office building.

This work is focused on Bayesian hierarchical modeling of geographical distribution of marine species Coregonus lavaretus L. s.l. along the Gulf of Bothnia. Spatial dependences are modeled by Gaussian processes. The main modeling objective is to predict whitefish larvae distribution for previously unobserved spatial locations along the Gulf of Bothnia. In order to achieve this objective, we have to solve two main tasks: to investigate the sensitivity of posterior parameters estimates with respect to different parameter priors, and to solve model selection task. In model selection, among all candidate models, we have to choose the model with best predictive performance. The candidate models were divided into two main groups: models that describe spatial effects, and models without such description. The candidates in each group involved different number (6 or 8) and expressions of environmental variables. In the group describing spatial effects, we analyzed four different models of Gaussian mean, and for every mean model we used four different prior parameters combinations. The same four models of latent function were used in the candidates where spatial dependences were not described. For every such model we assigned four different priors of overdispersion parameter. Thus, all at all, 32 candidate models were analyzed. All candidate models were estimated with Hamiltonian Monte Carlo MCMC algorithm. Model checks were conducted using the posterior predictive distributions. The predictive distributions were evaluated using the logarithmic score with 10 fold cross validation. The analysis of posterior estimates in models describing spatial effects revealed, that these estimates were very sensitive to prior parameters choices. The provided sensitivity analysis helped us to choose the most suitable priors combination. The results from model selection showed that the model, which showed best predictive performance, does not need to be very complicated and to involve description of spatial effects when the data are not informative enough to detect well the spatial effects. Although the selected model was simpler, the corresponding predictive maps of log larvae intensity correctly predicted the larvae distribution along the Gulf of Bothnia.

How to store data is an enduring topic in the computer science field, and traditional relational databases have done this well and are still widely used today. However, with the growth of non-relational data and the challenges in the big data era, a series of NoSQL databases have come into view. Thus, comparing, evaluating, and choosing a better database has become a worthy topic of research. In this thesis, an experiment that can store the same data set and execute the same tasks or workload on the relational, graph and multi-model databases is designed. The investigation proposes how to adapt relational data, tables on a graph database and, conversely, store graph data on a relational database. Similarly, the tasks performed are unified across query languages. We conducted exhaustive experiments to compare and report the performance of the three databases. In addition, we propose a workload classification method to analyze the performance of the databases and compare multiple aspects of the database from an end-user perspective. We have selected PostgreSQL, ArangoDB, Neo4j as representatives. The comparison in terms of task execution time does not have any database that completely wins. The results show that relational databases have performance advantages for tasks such as data import, but the execution of multi-table join tasks is slow and graph algorithm support is lacking. The multi-model databases have impressive support for simultaneous storage of multiple data formats and unified language queries, but the performance is not outstanding. The graph database has strong graph algorithm support and intuitive support for graph query language, but it is also important to consider whether the format and interrelationships of the original data, etc. can be well converted into graph format.

Recently it has been shown that sparse neural networks perform better than dense networks with similar number of parameters. In addition, large overparameterized networks have been shown to contain sparse networks which, while trained in isolation, reach or exceed the performance of the large model. However, the methods to explain the success of sparse networks are still lacking. In this work I study the performance of sparse networks using network’s activation regions and patterns, concepts from the neural network expressivity literature. I define network specialization, a novel concept that considers how distinctly a feed forward neural network (FFNN) has learned to processes high level features in the data. I propose Minimal Blanket Hypervolume (MBH) algorithm to measure the specialization of a FFNN. It finds parts of the input space that the network associates with some user-defined high level feature, and compares their hypervolume to the hypervolume of the input space. My hypothesis is that sparse networks specialize more to high level features than dense networks with the same number of hidden network parameters. Network specialization and MBH also contribute to the interpretability of deep neural networks (DNNs). The capability to learn representations on several levels of abstraction is at the core of deep learning, and MBH enables numerical evaluation of how specialized a FFNN is w.r.t. any abstract concept (a high level feature) that can be embodied in an input. MBH can be applied to FFNNs in any problem domain, e.g. visual object recognition, natural language processing, or speech recognition. It also enables comparison between FFNNs with different architectures, since the metric is calculated in the common input space. I test different pruning and initialization scenarios on the MNIST Digits and Fashion datasets. I find that sparse networks approximate more complex functions, exploit redundancy in the data, and specialize to high level features better than dense, fully parameterized networks with the same number of hidden network parameters.

This thesis regards non-Lambertian surfaces and their challenges, solutions and study in computer vision. The physical theory for understanding the phenomenon is built first, using the Lambertian reflectance model, which defines Lambertian surfaces as ideally diffuse surfaces, whose luminance is isotropic and the luminous intensity obeys Lambert's cosine law. From these two assumptions, non-Lambertian surfaces violate at least the cosine law and are consequently specularly reflecting surfaces, whose perceived brightness is dependent from the viewpoint. Thus non-Lambertian surfaces violate also brightness and colour constancies, which assume that the brightness and colour of same real-world points stays constant across images. These assumptions are used, for example, in tracking and feature matching and thus non-Lambertian surfaces pose complications for object reconstruction and navigation among other tasks in the field of computer vision. After formulating the theoretical foundation of necessary physics and a more general reflectance model called the bi-directional reflectance distribution function, a comprehensive literature review into significant studies regarding non-Lambertian surfaces is conducted. The primary topics of the survey include photometric stereo and navigation systems, while considering other potential fields, such as fusion methods and illumination invariance. The goal of the survey is to formulate a detailed and in-depth answer to what methods can be used to solve the challenges posed by non-Lambertian surfaces, what are these methods' strengths and weaknesses, what are the used datasets and what remains to be answered by further research. After the survey, a dataset is collected and presented, and an outline of another dataset to be published in an upcoming paper is presented. Then a general discussion about the survey and the study is undertaken and conclusions along with proposed future steps are introduced.

Recently, a neural network based approach to automatic generation of image descriptions has become popular. Originally introduced as neural image captioning, it refers to a family of models where several neural network components are connected end-to-end to infer the most likely caption given an input image. Neural image captioning models usually comprise a Convolutional Neural Network (CNN) based image encoder and a Recurrent Neural Network (RNN) language model for generating image captions based on the output of the CNN. Generating long image captions – commonly referred to as paragraph captions – is more challenging than producing shorter, sentence-length captions. When generating paragraph captions, the model has more degrees of freedom, due to a larger total number of combinations of possible sentences that can be produced. In this thesis, we describe a combination of two approaches to improve paragraph captioning: using a hierarchical RNN model that adds a top-level RNN to keep track of the sentence context, and using richer visual features obtained from dense captioning networks. In addition to the standard MS-COCO Captions dataset used for image captioning, we also utilize the Stanford-Paragraph dataset specifically designed for paragraph captioning. This thesis describes experiments performed on three variants of RNNs for generating paragraph captions. The flat model uses a non-hierarchical RNN, the hierarchical model implements a two-level, hierarchical RNN, and the hierarchical-coherent model improves the hierarchical model by optimizing the coherence between sentences. In the experiments, the flat model outperforms the published non-hierarchical baseline and reaches similar results to our hierarchical model. The hierarchical model performs similarly to the corresponding published model, thus validating it. The hierarchical-coherent model gives us inconclusive results – it outperforms our hierarchical model but does not reach the same scores as the corresponding published model. With our flat model implementation, we have shown that with minor improvements to a simple image captioning model, one can obtain much higher scores on standard metrics than previously reported. However, it is yet unclear whether a hierarchical RNN is required to model the paragraph captions, or whether a single RNN layer on its own can be powerful enough. Our initial human evaluation indicates that the captions produced by a hierarchical RNN may in fact be more fluent, however the standard automatic evaluation metrics do not capture this.