Banca de QUALIFICAÇÃO: VERNER RAFAEL FERREIRA

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
STUDENT : VERNER RAFAEL FERREIRA
DATE: 01/12/2023
TIME: 10:00
LOCAL: Remoto
TITLE:

FIberNet: A simple and efficient convolutional neural network architecture

KEY WORDS:

Convolutional Neural Network, Image Classification, Agave Sisalana plant.

PAGES: 100
BIG AREA: Ciências Exatas e da Terra
AREA: Ciência da Computação
SUBÁREA: Metodologia e Técnicas da Computação
SPECIALTY: Sistemas de Informação
SUMMARY:

With the ongoing increase in data generation each year, a wide array of technologies has emerged with the aim of transforming this information into valuable insights. However, often the financial and computational costs associated with the software used in this process render them inaccessible to the majority of individuals. A notable example is the requirement for specific hardware, such as Graphics Processing Unit (GPU) and Tensor Processing Unit (TPU), which are highly advanced in technological terms but are also notably expensive.

These challenges of accessibility are reflected in various areas of computing, including the application of Convolutional Neural Networks (CNNs), which play a crucial role in the field of computer vision. CNNs are highly effective in extracting meaningful information from images and identifying objects. However, the high cost associated with cutting-edge resources like GPUs and TPUs can limit the widespread adoption of these powerful networks. This makes it imperative to explore alternatives that allow for the construction and deployment of effective models with more accessible resources, without compromising the quality of results.

In this context, we present our research, in which we have developed an algorithm that sets itself apart from other existing models in terms of size, number of trainable parameters, and inference speed. Despite its compactness, the algorithm maintains high accuracy and the ability to process large volumes of data.

The proposed architecture, named FiberNet in reference to the sisal plant, is a small and straightforward CNN. The primary objective is to offer a financially viable low-cost model for classifying Agave Sisalana images and its fibers. FiberNet features a reduced number of trainable parameters, resulting in high inference speed. To achieve this, we employ a specialized layer that reduces the dimension of input data before the convolution layers.

The main goal of this research is to reduce computational costs without compromising algorithm performance. To assess the viability of the proposed method, we conducted an empirical analysis where our model achieved an accuracy of 96.25% on the Sisal dataset and 74.9% on the CIFAR10 dataset, using only 754,345 trainable parameters. Furthermore, we applied the proposed method to a widely recognized image dataset, obtaining promising results. These outcomes reinforce the effectiveness and applicability of our model in practice.

COMMITTEE MEMBERS:
Interna - 1350250 - ANNE MAGALY DE PAULA CANUTO
Externo ao Programa - 4351681 - JOAO CARLOS XAVIER JUNIOR - UFRNExterno à Instituição - ARAKEN DE MEDEIROS SANTOS - UFERSA