Faculty Mentor(s)

Bryson Payne, Ph. D

Campus

Dahlonega

Proposal Type

Poster

Subject Area

Computer Science/GIS

Location

Library Technology Center 3rd Floor Common Area

Start Date

24-3-2017 12:45 PM

End Date

24-3-2017 2:00 PM

Description/Abstract

The goal of this paper will be to analyze and present the discrepancies in performance of different implementations of neural networks. The paper aims to compare basic feed-forward neural networks, feed-forward neural networks with convolutional layers and lastly a recurrent convolutional neural network in the task of character recognition. Performance will be measured in terms of maximum accuracy achieved for the MNIST character dataset (with similar training times), training speed, and accuracy in recognizing handwritten digits outside of the MNIST dataset; for this purpose a custom dataset with handwriting samples will be created. To implement these neural networks, Python and TensorFlow will be utilized. The collected data will be used as a framework to make predictions regarding solutions for more elaborate deep learning utilizations, for instance object recognition. A conclusion about the potential held by different implementations for presenting viable solutions to problems the deep learning research community is currently concerned with will be presented at the end.

ARC_Poster_YG.pdf (631 kB)
Poster File

Share

COinS
 
Mar 24th, 12:45 PM Mar 24th, 2:00 PM

12. Approaches to Implementing Neural Networks in TensorFlow for the Task of Character Recognition, How They Differ, and What Inferences Can Be Drawn from the Results Regarding More Complex Problems

Library Technology Center 3rd Floor Common Area

The goal of this paper will be to analyze and present the discrepancies in performance of different implementations of neural networks. The paper aims to compare basic feed-forward neural networks, feed-forward neural networks with convolutional layers and lastly a recurrent convolutional neural network in the task of character recognition. Performance will be measured in terms of maximum accuracy achieved for the MNIST character dataset (with similar training times), training speed, and accuracy in recognizing handwritten digits outside of the MNIST dataset; for this purpose a custom dataset with handwriting samples will be created. To implement these neural networks, Python and TensorFlow will be utilized. The collected data will be used as a framework to make predictions regarding solutions for more elaborate deep learning utilizations, for instance object recognition. A conclusion about the potential held by different implementations for presenting viable solutions to problems the deep learning research community is currently concerned with will be presented at the end.