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 490 total
Previous  1  2  3  4  5  6  7  8  9  10  11  Next  Search:

C. Thuis
Master programme: Artificial Intelligence August 31st, 2018
Institute: Informatics Institute Research group: Computer Vision Graduation thesis Supervisor: Thomas Mensink
photo of the author
SSD-Sface: Single shot multibox detector for small faces
In this thesis, we present an approach to adapt the object detection method SSD for face detection. Face detection is a fundamental problem in AI since faces hold our identity, tell our age and gender and express our emotions. We test our method on the WIDER face dataset which contains faces in an uncontrolled setting, meaning that faces have a large variation in scale, pose, occlusion, blur, and illumination. Firstly, we experiment with a number of different resolutions to validate how each resolution performs on different face scales. Additionally, we adapt the SSD model to an image pyramid structure to combine the predictive power of each resolution. Finally, adding a selection criterion on each branch of the image pyramid further increases performance for all face scales, because the selection criteria negate the competing behavior of the image pyramid. We conclude that our approach not only increases performance on the small/hard subset of the WIDER dataset but keeps on performing well on the large subset. Multi-resolution
picture that illustrates the research done
Scientific abstract (pdf 1K)   Full text (pdf 4865K)

R.F. Guevara Melendez
Master programme: Artificial Intelligence August 30th, 2018
Institute: Informatics Institute Research group: Information and Language Processing Systems Graduation thesis Supervisor: Maarten Marx
photo of the author
Task-Oriented Dialog Agents Using Memory-Networks and Ensemble Learning
Task-oriented dialog agents are ever more relevant systems to engage with an user through voice or text naturallanguage input to fulfill domain-specific tasks. After recent encouraging results, neural based approaches are gaining popularity over traditional rule based systems. Hybrid Code Networks are one such prominent example that achieved the state of the art in the bAbI dialog tasks. This thesis goal is to modify the HCN architecture in 3 different ways to study the effect of using Memory Networks instead of an LSTM to predict the actions from the dialog agent. A second goal is to combine the Memory Network with an LSTM to produce an ensemble that predicts the actions with even higher accuracy. As a third goal, HCN is modified to include a Natural Language Understanding module to classify the intent behind each user utterance and detect the relevant entities.
picture that illustrates the research done
Scientific abstract (pdf 2K)   Full text (pdf 976K)

D. Kianfar
Master programme: Artificial Intelligence August 28th, 2018
Institute: Informatics Institute Research group: Amsterdam Machine Learning Lab Graduation thesis Supervisor: Herke van Hoof
Uncertainty-based Exploration for Model-based Reinforcement Learning
Efficient exploration is a fundamental problem in reinforcement learning -- how and when should an agent aim to maximize his long-term vs. short-term rewards in the absence of perfect knowledge about the environment? This work addresses this question through the lens of statistical sample bias -- an under-exploring agent does not have a representative view of the environment. We introduce a method for using tractable uncertainty estimation as a means for guiding exploration towards underrepresented regions. We leverage advances in Bayesian deep learning and uncertainty estimation to learn a representation of the environment from scratch using Bayesian neural networks, and explore how the uncertainty of the learned environment can be used an auxiliary reward to guide exploration over multiple steps. Unlike traditional planning-based approaches for model-based reinforcement learning, our method uses the learned environment uniquely for guiding exploration and can easily be applied to any model-free algorithm.
picture that illustrates the research done
Scientific abstract (pdf 1K)   For more info or full text, mail to: h.c.vanhoof@uva.nl

V. Isakov
Master programme: Artificial Intelligence August 28th, 2018
Institute: UvA / Other Research group: Information and Language Processing Systems (ILPS) Graduation thesis Supervisor: Dr. Ilya Markov
Visual Features for Information Retrieval
A web search engine is a software system whose main purpose is to search for relevant information on the World Wide Web. Most modern search engines rely on machine learning algorithms in order to find documents, which are relevant to the user query. LTR is an application of machine learning in the construction of ranking models for information retrieval systems. Most web search involve ranking, and many web search technologies can be potentially enhanced by using LTR techniques. Traditionally, most of LTR approaches rely on text analysis methods. Visual features have been recently introduced as a new development of LTR. A web page is rendered into a snapshot as it is shown in web browsers to users. This allows to understand how users see a web page given their current information need from a visual perspective. Text analysis, however, requires a number of handcrafted algorithms. A more natural approach would be to create such a representation of a document, which would allow to reproduce the text features in a visual manner. Since the introduction of visual features proved to be effective in LTR, using a pure visual approach can be an effective solution of this problem.
picture that illustrates the research done
Scientific abstract (pdf 1K)   Full text (pdf 910K)

S. Herrero Villarroya
Master programme: Artificial Intelligence August 27th, 2018
Institute: Informatics Institute Research group: UvA Bosch Delta Lab Graduation thesis Supervisor: Dr. Giorgio Patrini
photo of the author
Interpretability in sequence tagging models for Named Entity Recognition
The field of Explainable Artificial Intelligence has taken steps towards increasing transparency in the decision-making process of machine learning models for classification tasks. Understanding the reasons behind the predictions of models increases our trust in them and lowers the risks of using them. In an effort to extend this to other tasks apart from classification, this thesis explores the interpretability aspect for sequence tagging models for the task of Named Entity Recognition (NER). This work proposes two approaches for adapting LIME, an interpretation method for classification, to sequence tagging and NER. Given the challenges in the evaluation of the interpretation method, this work proposes an extensive evaluation from different angles. The evaluation has discovered patterns and characteristics to take into account when explaining NER models.
picture that illustrates the research done
Scientific abstract (pdf 1K)   Full text (pdf 1347K)

J.D. Gallego Posada
Master programme: Artificial Intelligence August 24th, 2018
Institute: Informatics Institute Research group: Amsterdam Machine Learning Lab Graduation thesis Supervisor: Patrick Forré
Simplicial AutoEncoders: A connection between Algebraic Topology and Probabilistic Modelling
Within representation learning and dimensionality reduction, there are two main theoretical frameworks: probability and geometry. Unfortunately, there is a lack of a formal definition of a statistical model in most geometry-based dimension reduction works, which perpetuates the division. We introduce a statistical model parameterized by geometric simplicial complexes, which allows us to interpret the construction of an embedding proposed by UMAP as an approximate maximum a posteriori estimator. This is a step towards a theory of unsupervised learning which unifies geometric and probabilistic methods. Finally, based on the the notion of structure preservation between simplicial complexes we define Simplicial AutoEncoders. Along with the construction of a probabilistic model for the codes in the latent space, Simplicial AutoEncoders provide a parametric extension of UMAP to a generative model.
picture that illustrates the research done
Scientific abstract (pdf 1K)   Full text (pdf 4180K)

F. Stablum
Master programme: Artificial Intelligence August 23rd, 2018
Institute: Informatics Institute Research group: Amsterdam Machine Learning Lab Graduation thesis Supervisor: Christos Louizos
photo of the author
Collaborative Filtering with Variational Autoencoders and Normalizing Flows
In this work we integrate collaborative filtering models that make use of Stochastic Gradient Variational Bayes with more recent posterior distribution approximation improvements, such as Planar and RealNVP Normalizing Flows. A model based on the AutoRec collaborative filtering autoencoder model is used as baseline in order to compare it to our Variational Autoencoder-based, named VaeRec and its variant VaeRec-NF which makes use of Normalizing Flows. Modifications to gradient-based parameter update algorithms are introduced in order to take into account the sparsity of the data tensors. Extensive hyperparameter search is performed and regularizing techniques have been investigated, such as soft free bits, which employs an adaptive coefficient to the Kullback-Leibler divergence of the variational lower bound. Methods to prevent gradient explosion are also utilized. A novel collaborative filtering input schema that makes use of the concatenation of user and item vectors has been tried, alongside inputs that make use of solely the item or user vectors.
picture that illustrates the research done
Scientific abstract (pdf 1K)   Full text (pdf 1139K)

K.L. van der Veen
Master programme: Artificial Intelligence August 23rd, 2018
Institute: UvA / Other Research group: UvA / Other Graduation thesis Supervisor: E. Gavves
photo of the author
A Practical Approach to Differential Private Learning
Applying differential private learning to real-world data is currently unpractical. Differential privacy (DP) introduces extra hyper-parameters for which no thorough good practices exist, while manually tuning these hyper-parameters on private data results in low privacy guarantees. Furthermore, the exact guarantees provided by differential privacy for machine learning models are not well understood. Current approaches use undesirable post-hoc privacy attacks on models to assess privacy guarantees. To improve this situation, we introduce three tools to make DP machine learning more practical. First, two sanity checks for differential private learning are proposed. These sanity checks can be carried out in a centralized manner before training, do not involve training on the actual data and are easy to implement. Additionally, methods are proposed to reduce the effective number of tuneable privacy parameters by making use of an adaptive clipping bound. Lastly, existing methods regarding large batch training and differential private learning are combined. It is demonstrated that this combination improves model performance within a constant privacy budget.
picture that illustrates the research done
Scientific abstract (pdf 1K)   Full text (pdf 694K)

U. Sharma
Master programme: Artificial Intelligence August 17th, 2018
Institute: ILLC Research group: Dialogue Modelling Group Graduation thesis Supervisor: Dieuwke Hupkes
Interpreting Decision-Making in Interactive Visual Dialogue
The order and underlying strategy of questions is important in dialogue-driven visual object discovery systems. In this work, we analyze the human game-play strategy with regular-expression based strategy-labeling on the GuessWhat?! data-set to understand human-strategy in an object-discovery setup. We also examine variational inference based controlled text generation models and their applicability to generating question with designated semantics. Finally, we propose a strategy-conditioning framework that generates interpretable conditioning signals for a multi-modal question generation mechanism in GuessWhat?! game. These signals improve the performance of the current module with stronger conditioning while additionally enforcing a stricter compliance with human game-play strategy. The discrete nature of these labels provides an interpretable interface into the model's current questioning strategy. As a part of this conditioning setup, we also introduce a strategy-prediction module trained on a multi-task learning setup that can generate these conditioning signals autonomously and ahead-of-time.
picture that illustrates the research done
Scientific abstract (pdf 2K)   Full text (pdf 1955K)

R.P. van der Weerdt
Master programme: Artificial Intelligence August 15th, 2018
Institute: VU / Other Research group: Knowledge Representation and Reasoning Graduation thesis Supervisor: Prof. Dr. F.A.H. van Harmelen
Generating Music from Text: Mapping Embeddings to a VAE’s Latent Space
Music has always been used to elevate the mood in movies and poetry, adding emotions which might not have been without the music. Unfortunately only the most musical people are capable of creating music, let alone the appropriate music. This paper proposes a system that takes as input a piece of text, the representation of that text is consequently transformed into the latent space of a VAE capable of generating music. The latent space of the VAE contains representations of songs and the transformed vector can be decoded from it as a song. An experiment was performed to test this system by presenting a text to seven experts, along with two pieces of music from which one was created from the text. On average the music generated from the text was only recognized in half of the examples, but the poems gave significant results in their recognition, showing a relation between the poems and the generated music.
picture that illustrates the research done
Scientific abstract (pdf 1K)   Full text (pdf 528K)

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

This page is maintained by thesis@science.uva.nl