Students’ Research Works – Autumn 2017: Neural Networks (PeWe.NN)

Learning Video Representations for Generating Descriptions

Patrik Gajdošík

Abstract: Eye-tracking is a great way to enhance the user experience. That can be either in the direct way, when using it as a new way for users to control applications, or in an indirect way, when eye-tracking is used by interface designers and application creators that use it for usability testing to increase the usability and efficiency of their applications. The problem with eye-tracking is that it requires specialized devices that capture the gaze but these devices are not easily accessible to ordinary users but are used in only specialized environments. However, web cameras are present in almost every mobile device.
In our work, we propose a solution that utilizes the ordinary web-cams for eye-tracking. In order to achieve that we are using neural networks that are good with data containing noise or lacking quality. The architecture of the neural network that we designed is based on existing techniques and models used in the field of AI, namely the Inception modules used in convolutional neural networks. We train and evaluate our solution on one of the UX datasets created as a part of the projects done at ÚISI.

to the topto the main

Text Generation with Neural Networks

Adam Kňaze
bachelor study, supervised by Matúš Pikuliak

Abstract: In my bachelor thesis I’m looking into ways to generate text with deep neural networks. To be more specific, I’m trying to create non-goal-driven conversational agent trained on corpus of data from microblogging service like Twitter.
There are several reasons for development of such agent (apart from having to cope with very interesting and challenging problem). It may be used for tasks without directly measurable goal (such as language learning), for fun (as a character in computer game) or more meta stuff (like human simulator that trains other goal-driven system).
The fastest-developing field in domain of natural language processing (with neural nets) is right now machine translation. I’m going to try utilize some of the current state-of-the-art approaches from machine translation in the task of response generation by conversational agent.

to the topto the main

Application of Neural Networks for Data Generation

Michal Maňak
bachelor study, supervised by Jakub Ševcech

Abstract: In this work, we address neural networks as a generative model. We analyze neural networks in general and focus on the use of neural networks to generate data similar to already existing data. Based on the analysis, we decided to generate handwritten digits using an autoencoder and a generative-adversarial network.
The basis is to teach the autoencoder to display data from the MNIST dataset. Subsequently, we generate data that should represent data similar to those from the MNIST dataset, but in a reduced dimensionality. This data should be similar to data we get from the encoder part, which is the essence of this work. The decoder part will then be used to visually display digits from generated data.

to the topto the main

Automatic Text Summarization of Customer Reviews

Samuel Pecár
doctoral study

Abstract: In recent years there were a lot of advances in natural language processing. Some of them, especially base techniques of processing text are mostly solved like part-of-speech tagging or named entity recognition. On the other hand, there is also many tasks, that has a lot work to do like creating open dialog systems, paraphrase generation or text summarization. In last few years a new problems, connected to era of internet, were introduced like mining social media. With exponential growth of data and texts on internet tasks like text summarization become significantly more important.
Task of text summarization is know for a very long period. In late 50s Luhn tried to automatically create abstract of documents. Over decades there have been many summarization systems dealing both with extractive and abstractive summarization. In recent years there have been significant progress in abstractive summarization of news articles.
Interesting part of text summarization is summarization of opinion in comments or customer reviews. With increasing number of comments on social networks and reviews on services and product is still more difficult to get right opinion or choose product wisely. Summarization of customer reviews can help either customers to decide, which product buy, but also can help producers and owners of services to improve their services or product in future.

to the topto the main

Cross-Lingual Learning

Matúš Pikuliak
doctoral study

Abstract: Deep learning is revolutionizing natural language processing for English or Chinese, but other less developed languages (such as Slovak) simply can not use these new exciting technologies simply because of the lack of data. Most of the world languages lack the necessary amount of supervised (and unsupervised) data to successfully train more advanced and complex neural network architectures. Cross-lingual learning is one solution for this problem, as it tries to transfer the knowledge learned from rich languages to others. This knowledge can be transferred on various levels (label-level, model-level, parameter-level) that differ in how hard the transfer is and what resources we need to do it. In my work I focus on creating a multilingual language independent representations for sentences or higher lexical units. Such representations encode the semantic information present in the text into short vector. This vector is in fact a projection of sentences into feature space shared across languages. Such representations can then be used to transfer knowledge that models in one language learned about this space to other languages.

to the topto the main

Generative Adversarial Networks

Adam Rafajdus
master study, supervised by Mária Bieliková

Abstract: Generative adversarial networks (GANs) are the new architecture of neural networks, in which two models are trained simultaneously and their adversarial relationship (playing min-max game against each other) helps producing better results on set tasks.
Although this framework is still pretty new, it showed its potential on tasks like generating quality image or understanding features from images.
Time series represents a series of data points in time order, which contains multi-level information about the domain. Weather data is perfect example of such data, containing multiple levels of data information.
In this work, we will be utilizing GANs architecture to implement a method whose main purpose is to forecast weather. We will try to fully take advantage of all variations of this architecture, which could help the neural network to understand the problems of weather domain and thanks to that, generate forecasting from the past data.

to the topto the main

Modelling the appropriateness of text posts

Andrej Švec
master study, supervised by Mária Bieliková

Abstract: The volume of user generated content on the web grows every year. Web discussions are also affected by this trend, called the rise of Web 2.0. It is impossible to keep the discussions clean by using only human resources. That is why we want to create an automated solution to help the moderators.
The goal of this thesis is to create a model that can detect potentially inappropriate text posts. The thesis mainly focuses on insulting and offensive posts. This is the reason for working primarily with artificial neural network models, because these models have proved the best performance in tasks similar to modelling the appropriateness of text posts, such as sentiment analysis.
Detecting potentially inappropriate posts is still not enough sometimes. To help the
moderators even more, we also want to generate a rationale for which the model thinks that a post is inappropriate. This way the moderator does not have to read the whole post when deciding whether it is really inappropriate. He just needs to read the provided rationale.

to the topto the main