Improving the Accuracy of Neural Networks
Brief Description
This project aims to answer the question: what are effective ways to improve a neural network’s accuracy? A neural network was created to classify an individual’s race according to photos of their faces. This neural network’s performance was recorded after changing its parameters to test how its accuracy is affected. This project also compares the results of normal neural networks vs convolutional neural networks (CNNs) and using black and white images vs RGB images. The results indicate that networks tend to perform better when there are more epochs, more layers, and/or more training data. There appears to be neither much of a difference in accuracy between normal neural nets and CNNs nor between black and white and RGB images. The lack of improvement is most likely because the dataset is not very diverse. However, this study did not cover all possible changes in parameter, so future studies should attempt more possible permutations.
Contributors
Emily Nguyen & Tan Bui-Thanh