This convolutional neural network for MNIST Handwritten Digit Classification reached a 99.4% accuracy. I helped assemble a CNN backbone, with 3 convolution and activation layers for feature extraction and 2 dense layers for classification, and experiment with the batch size hyperparameter tuning. Other hyperparameters experimented by Haijian Shao, a graduate student, included learning rates, dropout layers, kernel sizes, batch normalization, activation function, etc., which allowed for increased accuracy rates.
