|
|
Expert classification and retrieval
Nascimento Vianna, Felipe ; Pecina, Pavel (advisor) ; Vomlelová, Marta (referee)
Searching for experts is a common demand, especially within organizations. A general task called expertise retrieval relates people to topics and, therefore, can be used for expert finding and/or profiling experts. Currently, most approaches used to solve this task are based on traditional document retrieval methods and do not consider prior profiling information available. In this thesis, it is proposed to map people to topics by training a multi-class classifier using available profile data. The inputs (documents associated to candidates by authorship or other relations) and target data (profile information) were prepared by unsupervised document classification methods and were used to train a neural network. The effects of feature ordering and a convolutional layer are also evaluated. The experiments show that profiling the experts is not only suitable for a recommender system, but also an effective way for expert finding, achieving a performance comparable to state of the art in benchmark tasks such as TREC Enterprise. 1
|
|
|
Bayesian Optimization of Hyperparameters Using Gaussian Processes
Arnold, Jakub ; Straka, Milan (advisor) ; Vomlelová, Marta (referee)
The goal of this thesis was to implement a practical tool for optimizing hy- perparameters of neural networks using Bayesian optimization. We show the theoretical foundations of Bayesian optimization, including the necessary math- ematical background for Gaussian Process regression, and some extensions to Bayesian optimization. In order to evaluate the performance of Bayesian op- timization, we performed multiple real-world experiments with different neural network architectures. In our comparison to a random search, Bayesian opti- mization usually obtained a higher objective function value, and achieved lower variance in repeated experiments. Furthermore, in three out of four experi- ments, the hyperparameters discovered by Bayesian optimization outperformed the manually designed ones. We also show how the underlying Gaussian Process regression can be a useful tool for visualizing the effects of each hyperparameter, as well as possible relationships between multiple hyperparameters. 1
|
|
|
Theoretical aspect of modeling of user decision
Lacký, Peter ; Vojtáš, Peter (advisor) ; Vomlelová, Marta (referee)
In this thesis we address to the problematics of modelling user preferences. We discuss different views on user preferences as well as we give an overview of known models of user preferences and compare them. In more detail we introduce Fuzzy Logic Programming, Bayesian Logic Programming, Probabilistic Relational Models and Markov Logic Networks. For each model we propose transformations to other models and we show possible utilizations in real world. Finally we present our suggestions how to extend and improve these models. Powered by TCPDF (www.tcpdf.org)
|
|
|
Genres classification by means of machine learning
Bílek, Jan ; Neruda, Roman (advisor) ; Vomlelová, Marta (referee)
In this thesis, we compare the bag of words approach with doc2vec doc- ument embeddings on the task of classification of book genres. We cre- ate 3 datasets with different text lengths by extracting short snippets from books in Project Gutenberg repository. Each dataset comprises of more than 200000 documents and 14 different genres. For 3200-character documents, we achieve F1-score of 0.862 when stacking models trained on both bag of words and doc2vec representations. We also explore the relationships be- tween documents, genres and words using similarity metrics on their vector representations and report typical words for each genre. As part of the thesis, we also present an online webapp for book genre classification. 1
|
|
|
DRESS & GO: Deep belief networks and Rule Extraction Supported by Simple Genetic Optimization
Švaralová, Monika ; Mrázová, Iveta (advisor) ; Vomlelová, Marta (referee)
Recent developments in social media and web technologies offer new opportunities to access, analyze and process ever-increasing amounts of fashion-related data. In the appealing context of design and fashion, our main goal is to automatically suggest fashionable outfits based on the preferences extracted from real-world data provided either by individual users or gathered from the internet. In our case, the clothing items have the form of 2D-images. Especially for visual data processing tasks, recent models of deep neural networks are known to surpass human performance. This fact inspired us to apply the idea of transfer learning to understand the actual variability in clothing items. The principle of transfer learning consists in extracting the internal representa- tions formed in large convolutional networks pre-trained on general datasets, e.g., ImageNet, and visualizing its (similarity) structure. Together with transfer learn- ing, clustering algorithms and the image color schemes can be, namely, utilized when searching for related outfit items. Viable means applicable to generating new out- fits include deep belief networks and genetic algorithms enhanced by a convolutional network that models the outfit fitness. Although fashion-related recommendations remain highly subjective, the results we have achieved...
|
|
|
Transformation of Logic Programs
Vyskočil, Jiří ; Štěpánek, Petr (advisor) ; Vomlelová, Marta (referee) ; Mařík, Radek (referee)
This paper is a contribution to improving computational e fficiency of de nite Prolog programs using Unfold/Fold (U/F) strategy with homeomorphic embedding as a control heuristic. Unfold/Fold strategy is an alternative to so called conjunctive partial deduction (CPD). The ECCE system is one of the best system for program transformations based on CPD. In this thesis is presented a new fully automated system of program transformations based on U/F strategy. The experimental results, namely CPU times, the number of inferences, and the size of the transformed programs are included. These results are compared to the ECCE system and indicate that in many cases both systems have produced programs with similar or complementary e fficiency. Moreover, a new method based on a simple combination of both systems is presented. This combination represents, to our best knowledge, the most effective transformation program for normal logic programs. In most cases, the combination signi cantly exceeds both the Unfold/Fold algorithm presented here and the results of the ECCE system. The experimental results with a complete comparison among these algorithms are included.
|
|
|
Computational Intelligence Methods in Metalearning
Šmíd, Jakub ; Neruda, Roman (advisor) ; Vanschoren, Joaquin (referee) ; Vomlelová, Marta (referee)
This thesis focuses on the algorithm selection problem, in which the goal is to recommend machine learning algorithms to a new dataset. The idea behind solving this issue is that algorithm performs similarly on similar datasets. The usual approach is to base the similarity measure on the fixed vector of metafeatures extracted out of each dataset. However, as the number of attributes among datasets varies, we may be loosing important information. Herein, we propose a family of algorithms able to handle even the non-propositional representations of datasets. Our methods use the idea of attribute assignment that builds the distance measure between datasets as a sum of distance given by the optimal assignment and an attribute distance measure. Furthermore, we prove that under certain conditions, we can guarantee the resulting dataset distance to be a metric. We carry out a series of metalearning experiments on the data extracted from the OpenML repository. We build up attribute distance using Genetic Algorithms, Genetic Programming and several regularization techniques such as multi-objectivization, coevolution, and bootstrapping. The experiment indicates that the resulting dataset distance can be successfully applied on the algorithm selection problem. Although we use the proposed distance measures exclusively...
|
|
|
Summarization of gene expression arrays from free living species
Tuma, Vojtěch ; Mořkovský, Libor (advisor) ; Vomlelová, Marta (referee)
Gene expression arrays are used to assess expression of exons and genes of orga- nisms. The design of expression arrays is based on a genome of laboratory strains of model organisms. The most frequent summarization algorithms used to pro- cess data from measurements are gcRMA, PLER and IterPLIER. When using expression arrays to research free living species, the measured values are influen- ced by differences in genomes of free living and model organisms. We propose a method to improve the results by removing parts of genomes influenced by known differences between species from the summarization. Removing influenced parts can improve summarization, especially on exon level. 1
|
|
|
Pravděpodobnostní modely pro lokalizaci bezpilotního letounu testované na reálných datech
Figura, Juraj ; Vomlelová, Marta (advisor) ; Obdržálek, David (referee)
The thesis addresses the dynamic state estimation problem for the field of robotics, particularly for unmanned aerial vehicles (UAVs). Based on data collected from an UAV, we design several probabilistic models for estimation of its state (mainly speed and rotation angles), including the configurations where one of the sensors is not available. We use Kalman filter and Particle filter and focus on learning the model parameters using EM algorithm. The EM algorithm is then adjusted with respect to non-Gaussian density of some sensor errors and modified using model complexity penalization terms for better generalization. We implement these methods in MATLAB environment and evaluate on separate datasets. We also analyze data from a ground robot and use our implementation of Particle filter for estimation of its position. Powered by TCPDF (www.tcpdf.org)
|
|
|
Gobblet game from the point of artificial intelligence
Kotrč, Pavel ; Majerech, Vladan (referee) ; Vomlelová, Marta (advisor)
Gobblet is a new abstract board game, rules of which are based on the classic 4-in-arow game played on 4×4 board. However, the ability to gobble up and move the pieces on the board greatly increases its complexity and Gobblet is thus comparable to games like Checkers or Othello. That makes it interesting from the artificial intelligence point of view. This thesis explores the possibilities of classic and more recent methods for searching the Gobblet game tree - the minimax algorithm, alpha-beta pruning, a heuristic for move ordering, iterative deepening and others. The resulting algorithm is compared to the computer players on the Boardspace game server where it plays above-average with the best-playing robot. Implementation of all described algorithms and a graphical user interface for testing them in the Java programming language is an inseparable part of this thesis.
|
guest ::
RSS feed.