Objectives
-
Become familiar with ordinary fully-connected, feed-forward neural nets
-
Understand the fundamental differences between CNNs and ordinary NNs
-
Try various models, including LeNet-5 because of its history with the MNIST classification problem
-
Free to use the PyTorch Library to simplify the code
-
Classification Task: Identify the number from 1 to 9, among characters from unknown languages
-
Tamper with different neural net architectures
Methods/Features/Results
-
Pre-Processing: brought the pixel values into the range [-1, +1]
-
Experimented with adding 50% dropout; helped some models and hurt others
-
Model trained on 95% of the dataset, tested on the remaining 5%
-
Should experiment with hyperparameter tuning methodologies such as grid search in the future
-
Although LeNet-5 performed well, it was beaten by a custom model of mine that used dropout
-
Accuracy on validation set did not exceed 95% among the models experimented with
-
This work is associated to a project submitted for the McGill course: ECSE 551