Skip to content

SavannahMilburnUGA/Mini-Project-4

Repository files navigation

Mini-Project-4: ASL Classifier Neural Networks

This program analyzes different neural network architectures for classifying ASL fingerspelling shapes accurately.
We use a simple neural network, an improved simple neural network w/ improvements of cross-entropy cost function, regularization, and better network weight initialization, and a deep neural network that used the same improved neural network file but is created with multiple hidden layers in the training script.
We also use a CNN; however, due to time constraints, we could not figure out the bug in the learning curves to include all 4 network models so the learning curves only includes the simple, improved, and deep neural network models.
We can analyze how certain changes in neural network architecture impact accuracy via learning curve visualizations along with accuracy and training time visualizations.

Understanding Kaggle ASL data structure

  • Dataset: https://www.kaggle.com/datasets/datamunge/sign-language-mnist
  • Format: label, pixel1, pixel2, ..., pixel784
  • Labels: 0-24 (A-Y. Doesn't have J=9 OR Z=25).
  • 24 classes
    sign_mnist_train.csv = Training data:
  • 27,455 samples w/ 785 columns (1 label + 784 pixels)
    sign_mnist_test.csv = Testing data:
  • 7172 samples w/ 785 columns (1 label + 784 pixels)

AI was used in helping generate visualizations and debugging network examples from the Nielsen textbook.

About

Mini Project 4 for NSF REU - 6/18/2025

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 2

  •  
  •  

Languages