Navegando por Autor "Sampaio, Pablo Azevedo"

Agora exibindo 1 - 7 de 7

Aprendizado profundo com capacidade computacional reduzida: uma aplicação à quebra de CAPTCHAs
(2018-08-16) Melo, Diogo Felipe Félix de; Sampaio, Pablo Azevedo; http://lattes.cnpq.br/8865836949700771; http://lattes.cnpq.br/2213650736070295
During the last decade, Deep Neural Networks has been shown to be a powerfull machine learn technique. Generally, to obtain relevant results, these techniques require high computacional power and large volumes of data, which can be a limiting factor on some cases. Neverthless, a careful project of trainig and archtecture may help to reduce these requirements. In the this work we present a comparative approach to the application of deep neural networks to text based CAPTCHAs as a way to cope with these limitations. We studied models that are capable of learn to segment and identify the text content of images, only based on examples. By experimentation of different hiper-parameters and architectures, we were capable to obtain a final model with 96.06% of token prediction accuracy in approximately 3 hours of training in a simple personal computer.
Classificação de documentos de identificação com redes neurais convolucionais
(2019-01-18) Rocha, Alessandro Nazário da; Sampaio, Pablo Azevedo; http://lattes.cnpq.br/8865836949700771; http://lattes.cnpq.br/2466292990350036
Deep Learning has played an important role in data processing, with one of its main techniques being convolutional neural networks, which has the power to automatically learn features included in images in their layers. However, these deep artificial neural networks need a significant amount of images, for the proposed problem, already separated into categories to perform the training and validations of models that are not always available. In this context, in this work was constructed a dataset with images of Brazilian identification document of National Registry (RG) and National Qualification Card (CNH) separating them into some categories and, since they are documents with sensitive information, add a significant amount of images to obtain good results, it was a step that took time. For this reason, care was taken to preserve this sensitive information. This work presents some architectures of deep artificial neural networks to classify the images for different categories. Experiments were performed using a graphics processing unit (GPU) and using only the central processing unit (CPU). Results above 99% were obtained in some scenarios that were tested in the course of the work for the different proposed architectures.
Comparação de modelos de ia para extração de dados em glicosímetros
(2024-09-25) Carmo, Genivaldo Braynner Teixeira do; Correia, Julyanne Maria dos Santos; Silva Filho, Ronaldo Rodrigues da; Sampaio, Pablo Azevedo; Medeiros, Robson Wagner Albuquerque de
A diabetes é uma condição crônica que requer monitoramento constante dos níveis de glicose no sangue, sendo essencial o uso de glicosímetros para a obtenção dessas informações. Este trabalho tem como objetivo comparar três modelos de Inteligência Artificial, Gemini, GPT-4o e Llava 1.5, para identificar qual deles extrai, de forma mais eficaz, os dados de glicose, data e hora dos glicos´ımetros. Utilizando técnicas de engenharia de prompt, busca-se aprimorar a precisão e eficiência na extração desses dados, otimizando o monitoramento e contribuindo para a melhor gestão da saúde de pacientes diabéticos.
Um currículo de aprendizado por reforço para o cenário “Run to Score with Keeper” do Google Research Football Environment
(2019-12-10) Silva, Jonatan Washington Pereira da; Sampaio, Pablo Azevedo; Macário Filho, Valmir; http://lattes.cnpq.br/4346898674852080; http://lattes.cnpq.br/8865836949700771; http://lattes.cnpq.br/6846637095187550
Reinforcement learning is a group of techniques that allow an agent to interact with a particular environment. Agents observe the state of the environment and perform an action, the action is evaluated through a reward obtained. The agent objective is to maximize this reward. Various issues such as three-dimensional locomotion and electronic games have been addressed by reinforcement learning (KURACH et al., 2019). The Trainament of agents for a soccer game usually has sparse rewards, what slows learning (MATIISEN et al., 2019). One technique that can solve this obstacle is the curriculum learning proposed in (BENGIO et al., 2009). This technique use simplest tasks of the main task and the increase difficult level with the time. In This work we present two curriculum, identified as 5-15-30-50 e 3-10-20-67, for the scenario Run To Score With Keeper of Football Academy. We have shown that curriculums on average achieved better results compared to training only in the main scenario, without curriculum. Curriculum 3-10-20-67 achieved a better result even considering the pattern deviation.
Um currículo de aprendizagem por reforço para recompensas modeladas no Lunar Lander
(2021-07-19) Albuquerque, Renilson da Silva; Sampaio, Pablo Azevedo; http://lattes.cnpq.br/8865836949700771; http://lattes.cnpq.br/3364503614448061
Reinforcement learning is a machine learning paradigm where the agent learns to solve problems interacting with an environment, executing actions in a trial and error sequence. For each action performed, the agent receives a reward from the environment indicating how effective it was in solving the whole problem. The agent’s objective is to maximize the total reward received. However, in some reinforcement learning problems, the agent needs to learn complex tasks receiving uninformative rewards, leading to the credit assignment problem that slows the agent’s training process. Reward shaping and curriculum learning are techniques that can speed up agent training time by separating the problem into smaller tasks to be solved sequentially, applying smaller and informative rewards for each action performed. Lunar Lander is a simplified 2D simulator used as a benchmark for reinforcement learning solutions to the optimization problem on landing control of a lunar module. However, its standard rewards system assigns much more punitive rewards for the use of the engines, not being very constructive for the agent, which can lead to the credit assignment problem. Hence, this work proposes a curriculum using two additional shaped reward models and runs experiments that aim to minimize the Lunar Lander learning time. This work found that both the new models and the curriculum were more effective in training the Lunar Lander agent compared to the standard rewards model.
Programinó: um jogo para auxílio ao aprendizado do assunto de tipos de dados na programação
(2019-12-13) Nascimento, Gabriele Pessoa do; Falcão, Taciana Pontual da Rocha; Sampaio, Pablo Azevedo; http://lattes.cnpq.br/8865836949700771; http://lattes.cnpq.br/5706959249737319; http://lattes.cnpq.br/9978319013197863
The digital age we live in means that we are always immersed in ever more ubiquitous technologies. For this contact with technology to remain healthy, it is necessary to learn to consume it consciously, and to learn to develop it in different contexts; therefore, we will have increasingly inclusive solutions. Regarding solution development, as much as we have several facilitating artifacts, the process of programming teaching and learning is still a challenge, especially for beginning students. Dealing with so many competing and constant stimuli and still having the ability to abstract and assimilate programming concepts that is not trivial and has not been worked on since childhood, so playful artifacts such as digital games are essential to facilitate first contacts with the schedule. In this context, this work brings to society a digital educational game that deals with the subject of data types in programming, Programinó, so that beginning students can practice and consolidate the content through a playful tool. The game was developed with three levels of difficulty, one easy, one medium and one hard. The hard one applies the adapted minimax algorithm, while the easy one uses the same minimax adapted in an inverted way. The middle level uses a random algorithm. As a way to validate the difficulty levels, comparative experiments were performed that showed that the minimax lost only 5.6% of the time; winning at 49.7% or drawing in the remaining matches.
Teacher-Student aplicado a diferentes modelos de recompensa do ambiente Lunar Lander
(2021-07-20) Silva, Kenedy Felipe dos Santos da; Sampaio, Pablo Azevedo; http://lattes.cnpq.br/8865836949700771; http://lattes.cnpq.br/7479192156880225
The techniques studied related to learning by reinforcement are becoming more and more common in real world challenges, but one challenge is to reduce the learning time. Currently, the learning time and/or amount of interactions performed by the reinforcement learning agent can result in high costs in applications, as the training of models can consume a lot of time, requiring many interactions between the agent and the task environment. This work seeks to improve learning using a new combination of techniques, the Teacher-Student technique with Reward Shaping. The Teacher-Student technique aims to choose among a set of similar tasks that train for a main task, according to the student’s learning. The Reward Shaping technique, altering the reward to try to accelerate learning, provides more frequent feedback on appropriate behaviors, that is, reports rewards more often. We adapted Teacher-Student algorithms for this combination of techniques, and used the Lunar Lander environment as a case study, using four reward models designed in (ALBUQUERQUE, 2021) for this environment. Experiments were performed running different trainings to compare this approach with training only on the original Lunar Lander (no rewards change), and with the results obtained in (ALBUQUERQUE, 2021) by adopting each of the rewards models individually. The combination of Teacher-Student techniques with Reward Shaping contributed to a new experience in the reinforcement learning area, managing to accelerate the agent’s learning, considering the duration of 600 thousand training steps, reaching the target performance in 2 out of 5 proposals, in addition to of being able to learn better than the original Lunar Lander approach with PPO algorithm.