Ponencias Especializadas

Resumen:

Motor imagery is a mental process that consists in performing a movement only mentally, that is, without executing any fisical movement. It has been applied in rehabilitation of people with motor impairments and navigation in virtual environments through Brain-Computer Interfaces (BCIs). When executed, motor imagery produces two phenomena on cerebral electrical activity, called Event Related Desynchronization and Event Related Synchronization, that can be used to distinguish different types motor imagery tasks, such as left and right hand motor imagery. If we use an electroencephalogram to record these events, we can see that, usually, these events happen in frequencies that are specific for each person and, due to many factors, can change through time in the same person. Also, most of the time, long training sessions are required so that the BCI systems can learn to identify motor imagery tasks for each person, which makes the usage of motor imagery impractical for many applications. These three issues, make motor imagery classification a challenging task. In this presentation, I will show an end-to-end deep learning model built upon a convolutional recurrent neural network able to solve those issues.

Resumen:

In this paper, we proposed a model to learn both clusters and representations of our data in an end-to-end manner. Our model is a modification of the stacked generative model M1+M2 applied to semi-supervised learning, in which, we model our clusters with the Gumbel-Softmax distribution and we consider the use of a deterministic autoencoder to learn latent features, avoiding the problem of hierarchical stochastic variables. Experimental results on three datasets show the effectiveness of our model achieving competitive results with the state-of-the-art. Moreover, we show that our model generates realistic samples.

Resumen:

The stochastic streamflow models (SSMS) are time series models for precise prediction of hydrological data. They could generate ensembles of synthetic time series traces useful for hydrologic risk management. Nowadays, deep learning networks get many considerations in time series prediction. However, despite their theoretical benefits, they fail due to their architecture, defects of the backpropagation method, such as slow convergence and the vanishing gradient problem. In order to cover these requirements, we propose a new stochastic model applied in problems that involve phenomena of stochastic behavior and periodic characteristics. In the new model two components were used, the first one, a type of recurrent neural network embedding an echo-state (ESN) learning mechanism instead of conventional backpropagation method, an interesting feature of ESN is that from certain algebraic properties, training only the output of the network is often sufficient to achieve excellent performance in practical applications. The last part consists of the uncertainty associated with stationary processes, the model is finally called stochastic streamflow model ESN (SSNESN). This model was calibrated with time series of monthly discharge data from different river basins of MOPEX data set. We interpret our experimental findings by comparison with two feedforward neural networks and the traditional Thomas–Fiering model. The results show that the SSNESN can achieve a significant enhancement in the prediction performance, learning speed, and short-term memory capacity, with interesting potential in the context of hydrometeorological resources. This model, along with their simplicity and ease of training, can be considered a first attempt that applies the echo state network methodology to stochastic process.

Resumen:

Unmanned aerial vehicles (UAVs) are becoming increasingly popular. Researchers are trying to use them in various tasks, such as, surveillance of environments, persecution, collection of images, etc. In many cases, it may be interesting that they have the ability to analyze the images that are being collected in real time. In this work, we propose a vehicle tracking system to turn UAVs able to recognize a vehicle and monitor it in highways. It is based on a combination of bio-inspired algorithms: VOCUS2, CNN and LSTM.

Resumen:

Recommender systems are tools whose objective is to filter relevant content to users according to their preferences. Recently, due to the new demands of electronic businesses where most of users are not authenticated, session-based recommender systems emerged. This approach models session data (e.g. sequences of interactions, item metadata) to predict which items will be relevant for the user during the current session. In recent session-based approaches user long lasting preferences are not considered to make predictions which for some domains may bring improvements. Hierarchical recurrent neural networks were introduced to address this issue, based on the concatenation of session-level and user-level signals. Nevertheless, this concatenation does not discriminate between long and short term interests of users. In this research proposal is presented a mechanism for handling this concatenation by analyzing the output of both session and personalization level signals. It is expected to design a mechanism that can handle the influence of personalization according to the current session context.

Resumen:

Most of the state of art results in Natural Language Processing (NLP) tasks were obtained using deep learning approaches. Deep learning relies on the quality and quantity of labeled data and usually get better results using larger amounts of them. In some tasks and domains is difficult to get those resources, the problem gets worse on most the languages. To address this problem, we propose to use data augmentation, that is a technique widely used in Computer Vision and Image Processing, but not common in NLP because of text is discrete, and small changes to text can change the meaning of a sentence. This work reviews existing methods used to address this problem, and it propose a generative based model to make data augmentation in NLP tasks.

Resumen:

Common problems in climatological time series data are high dimensionality, correlation between the sequential values and noise due to calibration of meteorological stations influencing dramatically in the quality of clustering, classification, climate pattern finding and data processing. One way to deal with this problem is through feature extraction technique. In order to extract features from large climatological time series data, we propose a feature extraction method based on autoencoder neural network (AUTOE). As a first step, time series is standardized. Then, different architectures of autoencoder is applied on it to reduce dimensionality. Finally, k-means clustering algorithm are used to evaluate them through quality measures. As a result, autoencoder performs well and is competitive with other feature extraction techniques over Synthetic Control Chart Time Series.

Resumen:

Discutiremos algunas ideas de segmentación semántica usando deep learning.

Resumen:

El diseño e Implementación de un sistema de monitoreo en apartamentos usando Deep Learning y Computer Vision para la seguridad en casas o parqueos.

Resumen:

Advances in digital technology have increased event recognition capabilities through the development of devices with high resolution, small physical dimensions and high sampling rates. The recognition of complex events in videos has several relevant applications, particularly due to the large availability of digital cameras in environments such as airports, banks, roads, among others. The large amount of data produced is the ideal scenario for the development of automatic methods based on deep learning. Despite the significant progress achieved through image-based deep networks, video understanding still faces challenges in modeling spatio-temporal relations. In this work, we address the problem of human action recognition in videos. A multi-stream network is our architecture of choice to incorporate temporal information, since it may benefit from pre-trained deep networks for images and from handcrafted features for initialization. Furthermore, its training cost is usually lower than video-based networks. We explore visual rhythm images since they encode longer-term information when compared to still frames and optical flow. We propose a novel method based on point tracking for deciding the best visual rhythm direction for each video. Experiments conducted on the challenging UCF101 and HMDB51 data sets indicate that our proposed stream improves network performance, achieving accuracy rates comparable to the state-of-the-art approaches.

Resumen:

Over the years, the scientific interest towards 3D models analysis has become more popular. Problems such as classification, retrieval and matching are studied with the idea to offer robust solutions. This paper introduces a 3D object classification method for non-rigid shapes, based on the detection of key points, the use of spectral descriptors and deep learning techniques.

We adopt an approach of converting the models into a "spectral image". By extracting interest points and calculating three types of spectral descriptors (HKS, WKS and GISIF), we generate a three-channel input to a convolutional neural network. This CNN is trained to automatically learn features such as topology of 3D models. The results are evaluated and analyzed using the Non-Rigid Classification Benchmark SHREC 2011. Our proposal shows promising results in classification tasks compared to other methods, and also it is robust under several types of transformations.

Resumen:

A neural network is a mathematical model that is able to perform a task automatically or semi-automatically after learning the human knowledge that we provided. Moreover, a Convolutional Neural Network (CNN) is a type of sophisticated neural network that has shown to efficiently learn tasks related to the area of image analysis (among other areas). One example of these tasks is image segmentation, which aims to find regions or separable objects within an image. A more specific type of segmentation called semantic segmentation, makes sure that each region has a semantic meaning by giving it a label or class. Since neural networks can automate the task of semantic segmentation of images, they have been very useful for the medical area, applying them to the segmentation of organs or abnormalities (tumors). This project seeks to address the task of segmentation of volumetric medical images obtained by magnetic resonance: Magnetic Resonance Imaging (MRI). Volumetric images are composed of a set of 2D images that altogether represent a volume. We will use a pre-existing Three-dimensional Convolutional Neural Network (3D CNN) architecture, for the binary semantic segmentation of organs in volumetric images. We will talk about the data preprocessing process, as well as specific aspects of the 3D CNN architecture. We propose a variation in the formulation of the loss function used for training the 3D CNN, also called objective function, for the improvement of pixel-wise segmentation results. Finally we present the comparisons in performance we made between the proposed loss function and other pre-existing loss functions.