Influence of Different Convolutional Neural Network Settings on the Performance of MNIST Handwritten Digits Recognition

Abstract

Artificial Neural network as a power tool is now working in more and more areas such as recognition of handwritten digits and faces, sentence classification and audio recognition. This project aims at using convolutional neural network (CNN) to recognize handwritten digits provided by MNIST datasets with TensorFlow tool. Several variables such as dropout rate, epoch, fully connected layer with different neuron amount and filter amount were investigated in the experiments to investigate the performance of the model. Both training and validation accuracy were detected and gave us a direct reflection of the influence of different settings. With appropriate 0.2 dropout rate and 50 epochs, the model can accomplish 99.81% training accuracy and 99.38% validation accuracy. By adding fully connected layers with reasonable neuron amount, both training and validation accuracy were improved, and this modification had no extra running time required in our work. The filter number in the convolutional layers was also an important factor in the model performance. More filters can extract more features from the raw images and thus increase the accuracy of 99.95% for training accuracy and 99.47% for validation accuracy. However, this modification was found to have a longer running time.