|
|
Named Entity Recognition Exploiting Sub Word Information
Dobrovodský, Patrik ; Egorova, Ekaterina (oponent) ; Kesiraju, Santosh (vedoucí práce)
The aim of this thesis is the creation of a Named Entity Recognition system based on an older state-of-the-art model and studying how subword information can improve the recognition of out-of-vocabulary words. This proposed system besides English has to support two additional Indo-European languages: German and Hungarian. This work features a named entity tagger based on deep learning using pretrained and custom-trained word embeddings, sparse features, and character embeddings extracted by a Convolutional Neural Network. All these features are then processed by sequence-based (bidirectional Long Short-Term Memory) and feature-based (Conditional Random Field) approaches with the goal of achieving a F1-score similar to the work it is based on, and to compare how far present time state-of-the-art systems have evolved. The result is a system that achieves a 90.98% F1-score on the CoNLL 2003 English test dataset using pretrained word embeddings, not far behind the original work's 91.26%. For the other two languages, the model scores 89.34% on the WikiAnn German test dataset and 93.04% on the WikiAnn Hungarian test dataset with the usage of custom-trained embeddings.
|
|
|
Personal Voice Activity Detection
Sedláček, Šimon ; Landini, Federico Nicolás (oponent) ; Švec, Ján (vedoucí práce)
This work aims to implement, test, and evaluate a speaker-conditioned Voice Activity Detection (VAD) method called Personal VAD. The method builds upon an LSTM-based approach to VAD and its purpose is to introduce a system that can reliably detect speech of a target speaker, while retaining the typical characteristics of a VAD system, mainly in terms of small model size, low latency, and low necessary computational resources. The system is trained to distinguish between three classes: non-speech, target speaker speech, and non-target speaker speech. For this purpose, the method utilizes speaker embeddings as a part of the input feature vector to represent the target speaker. Some of the more heavyweight personal VAD variants also make use of speaker verification scores issued to each frame based on the target embedding, resulting in a more robust system. In addition to the one scoring method presented in the original article, two other scoring approaches are introduced, both outperforming the baseline method and improving the performance even for acoustically challenging conditions.
|
|
|
Umělý básník
Bančák, Michal ; Szőke, Igor (oponent) ; Beneš, Karel (vedoucí práce)
Dokument predstavuje prácu na automatickom generovaní poézie, pomocou Long Short-Term Memory rekurentnej neurónovej siete. Cieľom práce je vytvoriť aplikáciu, ktorá imituje písanie básní. Jedná sa o jazykové modelovanie na úrovni znakov v slovenskom jazyku. Model neurónovej siete použitý v práci sa skladá z troch vrstiev LSTM so 400 skrytými jednotkami. K tejto práci bola taktiež vytvorená zbierka básní v slovenskom jazyku vo veľkosti 900k znakov. Výsledkom práce je generovanie textu, ktorý má prvky básne. Dosahovaná presnosť generovania je 41.85%.
|
|
|
Detection of persons and evaluation of gender and age in image data
Dobiš, Lukáš ; Vičar, Tomáš (oponent) ; Kolář, Radim (vedoucí práce)
This master thesis describes an approach for automated human recognition by using convolutional neural networks (CNN) to perform facial analysis of persons face in image data. The predicted biometric indicators are following: age, gender, facial landmarks and facial expression. CNN architectures with pretrained weights for each task are described. Age estimation CNN has new weights trained and freezed, then has added new LSTM layers into its architecture. New LSTM layers are trained and tested on newly created video data set. Test results indicate improved age prediction accuracy. Solution for human recognition inference with single image and time series variants, in form of script with interconnected CNNs is explained, and its inference speed performance supports further proposed expansion plans for live video inference.
|
|
|
Vehicle Classification Using Radar
Gottwald, Vilém ; Zemčík, Pavel (oponent) ; Maršík, Lukáš (vedoucí práce)
The goal of this work is to recognize vehicles from radar point clouds. The radar produces the distance and angle for each target. This representation can be converted into the Cartesian coordinate system to obtain a point cloud 3D representation of the scene. In this thesis, existing approaches to object recognition in point clouds are presented. The method chosen for this thesis consists of object detection using point clustering and subsequent classification using a recurrent neural network. The objects are created from the point clouds using a modified DBSCAN algorithm. Features are extracted from each entity and utilized for classification into different types of vehicles using long short-term memory (LSTM) neural network. A dataset containing 57 345 annotated objects was created to train and evaluate the model. The developed model achieved an F1-score of 83 % on this data.
|
|
|
Time-Series Analysis and Prediction by Means of Neural Networks
Kňažovič, Martin ; Jaroš, Jiří (oponent) ; Bidlo, Michal (vedoucí práce)
This thesis deals with stock price prediction based on the creation of prediction models for selected stocks (BRK-A, GOOG, and MSFT), which can help investors in the creation of their financial decisions or by replacing other stock prediction models in existing prediction systems. Models created in this thesis are presented in two types - univariate model and multivariate model, which are in their final version presented in two architectures, one-layer architecture and two-layer architecture. Discussed models are created by means of neural networks, specifically recurrent neural networks with its extension - Long short-term memory. The output of the presented models is a forecast of the next-day stock price, which can be used for evaluating the right time to buy or sell a given stock. The quality of individual prediction models is evaluated via the mean squared error of the validation or testing dataset or alternatively based on stock price trend prediction.
|
|
|
Named Entity Recognition Exploiting Sub Word Information
Dobrovodský, Patrik ; Egorova, Ekaterina (oponent) ; Kesiraju, Santosh (vedoucí práce)
The aim of this thesis is the creation of a Named Entity Recognition system based on an older state-of-the-art model and studying how subword information can improve the recognition of out-of-vocabulary words. This proposed system besides English has to support two additional Indo-European languages: German and Hungarian. This work features a named entity tagger based on deep learning using pretrained and custom-trained word embeddings, sparse features, and character embeddings extracted by a Convolutional Neural Network. All these features are then processed by sequence-based (bidirectional Long Short-Term Memory) and feature-based (Conditional Random Field) approaches with the goal of achieving a F1-score similar to the work it is based on, and to compare how far present time state-of-the-art systems have evolved. The result is a system that achieves a 90.98% F1-score on the CoNLL 2003 English test dataset using pretrained word embeddings, not far behind the original work's 91.26%. For the other two languages, the model scores 89.34% on the WikiAnn German test dataset and 93.04% on the WikiAnn Hungarian test dataset with the usage of custom-trained embeddings.
|
|
|
Personal Voice Activity Detection
Sedláček, Šimon ; Landini, Federico Nicolás (oponent) ; Švec, Ján (vedoucí práce)
This work aims to implement, test, and evaluate a speaker-conditioned Voice Activity Detection (VAD) method called Personal VAD. The method builds upon an LSTM-based approach to VAD and its purpose is to introduce a system that can reliably detect speech of a target speaker, while retaining the typical characteristics of a VAD system, mainly in terms of small model size, low latency, and low necessary computational resources. The system is trained to distinguish between three classes: non-speech, target speaker speech, and non-target speaker speech. For this purpose, the method utilizes speaker embeddings as a part of the input feature vector to represent the target speaker. Some of the more heavyweight personal VAD variants also make use of speaker verification scores issued to each frame based on the target embedding, resulting in a more robust system. In addition to the one scoring method presented in the original article, two other scoring approaches are introduced, both outperforming the baseline method and improving the performance even for acoustically challenging conditions.
|
|
|
Detection of persons and evaluation of gender and age in image data
Dobiš, Lukáš ; Vičar, Tomáš (oponent) ; Kolář, Radim (vedoucí práce)
This master thesis describes an approach for automated human recognition by using convolutional neural networks (CNN) to perform facial analysis of persons face in image data. The predicted biometric indicators are following: age, gender, facial landmarks and facial expression. CNN architectures with pretrained weights for each task are described. Age estimation CNN has new weights trained and freezed, then has added new LSTM layers into its architecture. New LSTM layers are trained and tested on newly created video data set. Test results indicate improved age prediction accuracy. Solution for human recognition inference with single image and time series variants, in form of script with interconnected CNNs is explained, and its inference speed performance supports further proposed expansion plans for live video inference.
|
|
|
Umělý básník
Bančák, Michal ; Szőke, Igor (oponent) ; Beneš, Karel (vedoucí práce)
Dokument predstavuje prácu na automatickom generovaní poézie, pomocou Long Short-Term Memory rekurentnej neurónovej siete. Cieľom práce je vytvoriť aplikáciu, ktorá imituje písanie básní. Jedná sa o jazykové modelovanie na úrovni znakov v slovenskom jazyku. Model neurónovej siete použitý v práci sa skladá z troch vrstiev LSTM so 400 skrytými jednotkami. K tejto práci bola taktiež vytvorená zbierka básní v slovenskom jazyku vo veľkosti 900k znakov. Výsledkom práce je generovanie textu, ktorý má prvky básne. Dosahovaná presnosť generovania je 41.85%.
|
host ::
RSS.