Uma abordagem espaço-temporal para aprendizado em conjunto usando redes ConvLSTM

• Autor(a) principal: Souto, Yania Molina
• Data de Publicação: 2018
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações do LNCC
• Texto Completo: https://tede.lncc.br/handle/tede/296
• topic: Redes neurais (Computação); Aprendizagem; Aprendizagem em conjunto; Ensembles; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO

The number of sensors, radars and objects of all kinds that constantly provide space-time data at different resolutions, ranging from global satellites to regional ones made available by meteorological or cellular stations, grow daily. This sheer volume of data has enabled the development of a number of models to solve diverse phenomena of spatio-temporal nature. Among the benefited areas we can cite the meteorology, where advances in the models of weather forecast have been considerable. However, this breakthrough is accompanied by major challenges. The data are collected with an heterogeneous spatial distribution, because it depends on the geographical location of the devices that feed the processes of data assimilation. Thus there are regions where information is abundant while others are analyzed with little data. Each of these sources have varying qualities and uncertainties. Finally, these sets serve as inputs to construct the models of phenomena under study or as initial conditions when they are initialized to generate new predictions. We arrive at a scenario where we have a large number of models, constructed under different criteria, that generate simulations with different predictive qualities, whose variations can be interpreted as uncertainty in the prediction of the phenomenon. Ensemble forecasting strategies (ensembles) arise to solve problems of this nature. Using the combination of predictions resulting from applying multiple models attempts to improve the predictive quality of the set. Existing collaborative learning methods are limited to exploring temporal patterns without exploiting spatial dependencies, which could lead to more precise quality criteria or metrics. This work proposes a space-time clustering approach based on deep neural networks that allows the simultaneous exploration of patterns in space and time dimensions. For this, we adopted a Convolutional Long Short-Term Memory network architecture, which models spatio-temporal series as regular arrays.