Euclidean Distance based Loss Function for Eye-Gaze Estimation

The Loss function is an integral component in a Neural network. It affects the performance of CNN network in its classification. In this paper, we propose a Euclidean distance based Loss function for the CNN model, in an eye-gaze memory card game. We compared the Euclidean distance loss function with the well-known cross-entropy loss function. The performance parameters used in our comparison are prediction accuracy and average Euclidean distance prediction error. The results show that cross-entropy has better prediction accuracy. However, the Euclidean distance loss function provides a better average Euclidean distance prediction error resulting in better user experience. This is because the wrongly predicted eye gaze cards are near to the user intended card. In the case of cross-entropy, the predicted card error is quite evenly spread across the screen.