Horário: 14h
Palestrante: Diego Santos Kieckbusch (mestrado)
Orientador: Prof Li Weigang
Title: Towards Intelligent Processing of Electronic Invoices: the General Framework and Case Study of Short Text Deep Learning in Brazil
Abstract: An electronic invoice (E-invoice) is a kind of document that
record goods or service transactions, and store and exchanges them electronically. E-invoice is an emerging practice and presents a valuable source of information for many areas. Dealing with these invoices is usually a very challenging task. Information reported is often incomplete or presents mistakes. Before any meaningful treatment of these invoices, it is necessary to evaluate the product represented in each file. This research puts forward a conceptual framework to explain how to apply machine learning technology to extract meaningful information from invoices at different levels of aggregation. Related work in the field is contextualized within a given framework. A study case based on real data from electronic invoice (NF-e) and Electronic Consumer Invoice (NFC-e) documents in Brazil, related to B2B and retail transactions. We compared traditional term frequency models with the Convolutions sentence classification models. Our experiments show that even if invoice text descriptions are short and there are a lot of errors and typos, simple term frequency models can achieve high baseline results on product code assignment.
Horário: 14h20
Palestrante: Flávia Maria Alves Lopes (doutorado)
Orientadora: Profa Alba Cristina M. A. de Melo
Horário: 14h40
Palestrante: Osmar Luiz Ferreira de Carvalho (mestrado)
Orientador: Prof Dibio Leandro Borges
Title: Deep Learning & Remote Sensing: Pushing the Frontiers in Image Segmentation
Abstract: Image segmentation is a hot computer vision topic that aims to simplify the understanding of digital images enabling the extraction of useful information. Deep learning approaches with convolutional neural networks were game-changing, allowing the exploration of different tasks, in which semantic, instance, and panoptic segmentation are the most important. Semantic segmentation assigns a class to every pixel in an image, instance segmentation classifies objects at a pixel level with a unique identifier for each target, and panoptic segmentation combines instance-level predictions with different backgrounds. Remote sensing data largely benefits from those methods, being very suitable for developing new DL algorithms and creating solutions using top-view images. Nonetheless, some peculiarities prevents this field from growing when compared to traditional images (e.g., camera photos): (1) The images are huge, (2) it presents different characteristics such as many channels and image format, (3) there is a high number of pre-processes and post-processes steps, such as extracting patches and classifying large scenes, and (4) most of the open software for labeling and deep learning applications are not friendly to remote sensing due to the aforementioned reasons. This dissertation aims to advance in all three main categories of image segmentation, presenting solutions that may increase research on this topic. First, we enhanced the box-based instance segmentation approach for classifying large scenes, allowing practical pipelines to be implemented. Besides, we evaluated the impact of the image dimensions in which increasing the image significantly improves results. We developed the first semi-supervised learning procedure using remote sensing and GIS data, allowing for detecting vehicles on a city scale, with more than 120 thousand unique vehicles. Besides, we proposed a new box-free instance segmentation approach by using a simple change in the data preparation step, considering the target borders. Moreover, we presented the first remote sensing panoptic segmentation dataset containing fourteen classes and disposed of software and methodology for converting GIS data into the panoptic segmentation format. Since our first study considered RGB images, we extended this method for multispectral data, in which we used WorldView-3 images in the beach setting, considering 13 classes and eight spectral bands. Then, we extended the box-free method initially designed for instance segmentation to the panoptic segmentation task. Finally, we proposed the first amodal perspective dataset for dealing with occlusion in remote sensing images.
Horário: 15h
Palestrante: Leonardo Rebouças de Carvalho (doutorado)
Orientadora: Profa Aleteia Patrícia Favacho de Araújo
Título: Orama: um framework para execução de benchmarks de Function-as-a-Service
Resumo: O proeminente modelo de serviço em nuvem Function-as-a-Service (FaaS) se posicionou como uma alternativa para resolver vários problemas e, portanto, o interesse em soluções arquitetônicas orientadas à nuvem que usam FaaS cresceu rapidamente. Consequentemente, a importância de conhecer o comportamento de arquiteturas baseadas em FaaS em diferentes cenários de simultaneidade também se tornou significativa, principalmente nos processos de tomada de decisão de implementação. Neste trabalho, é proposto o framework Orama, que auxilia na execução de benchmarks em ambientes baseados em FaaS, orquestrando a implantação de arquiteturas pré-construídas, bem como a execução de testes e análises estatísticas. Foram realizados experimentos com arquiteturas contendo vários serviços em nuvem em conjunto com FaaS em dois provedores de nuvem pública (AWS e GCP). Os resultados foram analisados usando projeto fatorial e teste t e mostraram que os casos de uso executados na AWS obtiveram melhor resultado em tempo de execução em comparação com seus equivalentes no GCP, mas apresentaram taxas de erro consideráveis em situações de concorrência. Vale ressaltar que o framework Orama foi utilizado desde o provisionamento automatizado de casos de uso, execução de benchmarks, análise de resultados e desprovisionamento do ambiente, dando suporte a todo o processo.
Horário: 15h20
Palestrante: Wanderson Jean Conceição Silva (mestrado)
Orientadora: Profa Maristela Holanda
Title: Automatic feedback in the teaching of programming in undergraduate courses
Abstract: Teaching programming in the early years of undergraduate courses has been a challenge for teachers, institutions, and professors. In view of this premise the Learning Management System (LMS) and other teaching platforms have emerged as possible solutions to solve the difficulties encountered in this process.
Horário: 15h40
Palestrante: Guo Ruizhe (mestrado)
Orientador: Prof Li Weigang
Title: Pre-Training+RoBERTa
Resumo: Chinese-Português modelo de tradução
Local: Teams MS - Equipe PPGI-316415 Seminário, Canal Seminários 1-2022
Profa Célia Ghedini Ralha (Este endereço de email está sendo protegido de spambots. Você precisa do JavaScript ativado para vê-lo.)
Coordenadora Seminários de Pós-Graduação em Informática 1-2022