Artificial Intelligence in Progress

Home   Artificial Intell.   Bio-exact   Chemistry   Computer Sci. BSc   Earth Sciences   Mathematics   Physics & Astr.   Science Educ.   Submit   Webmaster        
bachelors   masters   all  

Displaying theses 1-10 of 428 total
Previous  1  2  3  4  5  6  7  8  9  10  11  Next  Search:

S. Negrijn
Master programme: Artificial Intelligence December 19th, 2017
Institute: Informatics Institute Research group: Computer Vision Graduation thesis Supervisor: Dr T.E.J. Mensink
Exploiting Symmetries to Relocalise in RoboCup Soccer
In this thesis we are interested in exploiting symmetries in order to more accurately determine the location of a robot while playing robotic soccer. For many autonomous robots, localisation is an essential task as without it the next action can not reliably be determined. For example, a robot can not run clear if it has no understanding of its own position in the field, the position of the ball and the position of its teammates and opponents. When the robot no longer knows where it is at it needs to redetermine, or relocalise, its location without being able to use its previously known position. By exploiting the perfect symmetries of a soccer field and extending PoseNet, a relocalisation model, we use a single color image to predict the position of the robot. Instead of generating these images with a real robot, a simulator is created to easily be able to change conditions such as the lighting, field colour and even the surroundings of the field. After showing that using these symmetries in a soccer field increases the accuracy of the location predictions, we also show an increase when compared to the original PoseNet architecture and its Street Scenes data sets.
picture that illustrates the research done
Scientific abstract (pdf 28K)   Full text (pdf 5795K)

C. Zavou
Master programme: Artificial Intelligence December 18th, 2017
Institute: UvA / Other Research group: UvA / Other Graduation thesis Supervisor: dr. Maarten de Rijke
Question Retrieval in Community Question Answering Enhanced by Tags Information in a Deep Neural Network Framework
Community Question Answering (CQA) platforms need to be easy and fast in question or answer exploration. It is common to use tags to categorize items in these platforms, and create taxonomies that assist exploration, indexing and searching. The focus of this thesis lies in recommending similar questions (Question Retrieval) by simultaneously deciding whether the contexts of two questions are similar, and which tags are applicable for each question. Current methods targeted for Question Retrieval in CQA either consider deep learning approaches (Lei et. al. 2016, Bogdanova et al. 2015), or conventional approaches that utilize the available information on questions' tags (Cao et. al. 2010, Zhou et al. 2013). The former framework is proved to be more powerful –especially in the case with loads of available data–, while the later is faster and successful in cases with few data. In this thesis, a deep learning approach for both question retrieval and tags recommendation is proposed, and their joint learning is found successful for transferring knowledge in the Question Retrieval task, after applying it on the AskUbuntu forum data. Additionally the neural network based Tag Recommendation performs better than the existing conventional methods.
picture that illustrates the research done
Scientific abstract (pdf 2K)   Full text (pdf 1281K)

M.J. Scheepers
Master programme: Artificial Intelligence December 14th, 2017
Institute: Informatics Institute Research group: Information and Language Processing Systems Graduation thesis Supervisor: Evangelos Kanoulas
Improving the Compositionality of Word Embeddings
Word embedddings are mathematical representations of word meaning. Combining these representations allows us to create meaning for statements, sentences or even entire paragraphs. In this thesis we attempt to find better word embeddings which are easier to combine, i.e. compose. Our results indicate that we can improve the compositionality for four popular and widely used word embeddings.
picture that illustrates the research done
Scientific abstract (pdf 79K)   Full text (pdf 2303K)

K.T. Keune
Master programme: Artificial Intelligence December 13th, 2017
Institute: UvA / Other Research group: UvA / Other Graduation thesis Supervisor: Maarten van Someren
Robust Detection of Anomaly Types in Energy Consumption Data
Improving the energy consumption in buildings is an ongoing problem. One of the ways to improve the energy consumption in buildings is by detecting anomalies. This thesis makes a clear distinction between anomaly types, while other approaches do not. This paper introduces different anomaly types based on the anomalies that occur in energy consumption data and presents a robust method to detect them. The anomaly types occur in the context of the outside temperature and the time in hours. They are detected with a robust method that consists of finding a regression model of the data with the context and an anomaly detection rule based on the robust residuals of the regression model. The results show that the general anomaly detection approach could be used for the detection of anomaly types in energy consumption data.
picture that illustrates the research done
Scientific abstract (pdf 2K)   Full text (pdf 336K)

J. Sander
Master programme: Artificial Intelligence November 24th, 2017
Institute: Informatics Institute Research group: Amsterdam Machine Learning Lab Graduation thesis Supervisor: Max Welling
photo of the author
Combining adaptive-computation-time and learning-to-learn approaches for optimizing loss functions of base-learners
Learning to learn approaches can be used to train a recurrent neural network (RNN) that learns an optimization algorithm. Adaptive computation time for RNNs allows a network to adjust its depth at each time step to the input received so far. In this work we developed two adaptive meta-learners that combine the learning to learn and adaptive computation time approaches in order to optimize loss functions of base-learners. Integrating both approaches is motivated by the idea that a learned, iterative optimization algorithm benefits from being able to adjust the number of optimization steps to the input, by learning to weight the time step losses. Our newly proposed models could be trained with less computational effort on convex and non-convex optimization tasks compared to the baseline optimizer introduced by Andrychowicz et al. (2016). Moreover the adaptive optimizers developed their own training regime that trades off computational effort against accuracy based on a prior injected preference.
picture that illustrates the research done
Scientific abstract (pdf 1K)   Full text (pdf 8204K)

Y.M.A.A. Galama
Master programme: Artificial Intelligence November 21st, 2017
Institute: Informatics Institute Research group: Computer Vision Graduation thesis Supervisor: Thomas Mensink
Learning implicit 3D models; Generating images from a different viewpoint using a single image
3D manipulation of objects requires a full understanding of their geometry and appearance. Previous research therefore often required explicit 3D models to extract and manipulate objects in a scene. In this thesis, we explore the possibility of transforming images with Deep Learning, and in this way rotating the object within the image in 3D. Thus, this method no longer needs an explicit 3D model, it is shown that it is still possible to rotate the objects. It is also demonstrated that the implicit 3D model is able to rotate objects never seen before, thus indicating the models contain generic visual knowledge.
picture that illustrates the research done
Scientific abstract (pdf 1K)   For more info or full text, mail to:

J. Baptist
Master programme: Artificial Intelligence October 24th, 2017
Institute: ILLC Research group: Language and Computation Graduation thesis Supervisor: Joost Bastings
photo of the author
Re-encoding in Neural Machine Translation
Currently, most approaches to neural machine translations use an encoder that produces continuous representations of the words in the input sentence. Next, another component decodes the representations into output words. This approach comes with two potential problems. First, the input representations remain constant during the prediction process, which may not provide enough variance for the model to effectively discriminate between them. Second, a high burden is put on the component responsible for predicting output words (decoder), as translations involves many subtasks: knowing which words to translate next, remembering which words have been translated, producing a fluent and grammatically correct sentence, and more. As a result, many neural models suffer from over-translation (generating unnecessary words), under-translation (forgetting to translate words) and repetition. By extending such models with a re-encoding component, which allows the model to update the input representations depending on the previous predictions, these tasks can naturally be moved away from the decoding component. This thesis investigates two architectures that use re-encoding and compares them to multiple baselines. The qualitative and quantitative results show that re-encoding can potentially improve performance of neural models, especially on longer sentences.
picture that illustrates the research done
Scientific abstract (pdf 1K)   Full text (pdf 1243K)

L.C. Herstix
Master programme: Artificial Intelligence September 29th, 2017
Institute: VU / Other Research group: Computational Intelligence Group Graduation thesis Supervisor: Mark Hoogendoorn
Predicting Risks of Readmission and Mortality for Releasing ICU Patients - A Comparison of various Machine Learning Techniques
This thesis compares the performance of four machine learning algorithms for predicting readmission and mortality risks. In total, eight predictive models are investigated. The four algorithms used are the logistic regression, random forest, multilayer perceptron (MLP) and support vector machines (SVM). The underlying dataset used is the Medical Information Mart for Intensive Care III (MIMIC-III) dataset. A daily model was used for data aggregation and the high imbalance in the dataset was handled by weighing the classes. The outcome of the predictions showed high discrepancies depending on the machine learning technique used. The best performance was achieved by SVM (mortality prediction: AUC of 0.994, readmission prediction AUC of 0.984) followed by random forest (mortality prediction: AUC of 0.931, readmission prediction AUC of 0.940). Therefore, both SVM and random forest produced extremely good results with medical significance (AUC over 0.80) for both predictive cases. Multilayer perceptron was not able to produce any reasonable results since the minority class was not recognized. Logistic regression performed poorly for the prediction of mortality (AUC of 0.62) and only slightly better than random for the prediction of readmission.
picture that illustrates the research done
Scientific abstract (pdf 2K)   For more info or full text, mail to:

B.H.L. Vredebregt
Master programme: Artificial Intelligence September 22nd, 2017
Institute: Informatics Institute Research group: Computer Vision Graduation thesis Supervisor: Theo Gevers
Crowdsourcing Part Annotations for Visual Veri fication
Visual Verification involves first detecting the object, followed by locating each of the parts and finally judging the state of each part. Many popular datasets already contain object annotations, datasets containing part annotations are rare and there is no dataset that provides part state judgments. As a result state-of-the-art object detection algorithms are only evaluated on detecting relatively large objects and not on the often much smaller parts. Thus there is a need for a new dataset. In this thesis we created an unique crowdsourced dataset consisting of 10.000 images of bicycles in various settings to fill this gap. For each bicycle 22 parts were annotated using the crowdsourcing platform CrowdFlower. Resulting in a total of 220.000 bounding box annotations. Additionally each part in the dataset also was judged to determine its state (intact, broken, occluded or absent) allowing future research into Visual Verification. For this purpose 220.000 state judgments are made available in addition to the bounding boxes. In our experiments we show that, unlike most crowd sourcing campaign, only a single judgment is sufficient to create annotations of sufficient quality. We show under which condition this is a reasonable trade-off between cost and quality.
picture that illustrates the research done
Scientific abstract (pdf 2K)   Full text (pdf 3180K)

M.A. ter Hoeve
Master programme: Artificial Intelligence September 15th, 2017
Institute: Informatics Institute Research group: Information and Language Processing Systems Graduation thesis Supervisor: Maarten de Rijke
photo of the author
Explaining Rankings
Machine learning algorithms are increasingly popular for all kinds of tasks. Often, we do not have a clear understanding of how these algorithms really work and what they base their decisions on. Understanding this is important for the user of the system and for the developer of the system. In this study we look into the explainability of ranking algorithms. Ranking algorithms are used to score and order items in a list, for example the search results that are returned after querying a search engine. We explain the decisions of a ranker by changing the input to the ranker. If this changes the output of the ranking, this part of the input is assumed to be important. If not, this part of the input is assumed to be unimportant. This process is very time consuming. Therefore we train a neural network that learns how to find the most important parts of the input. We test our approach on users of online news kiosk Blendle. We find that the large majority of users wants to see explanations for the articles in their personalized selection, yet that their reading behaviour does not change depending on the type of explanation they get.
picture that illustrates the research done
Scientific abstract (pdf 3K)   Full text (pdf 12586K)

Previous  1  2  3  4  5  6  7  8  9  10  11  Next  

This page is maintained by