Search for: All records

An experiment was performed to investigate a modified pooling method for use in convolutional neural networks for image recognition. This algorithm–Variable Stride–allows the user to segment an image and change the amount of subsampling in each region. This control allows for the user to maintain a higher amount of data retention in more important regions of the image, while more aggressively subsampling the less important regions to increase training speed. Three Variable Stride methods were compared to the preexisting pooling algorithms, Maximum Pool and Average Pool, in three different network configurations tasked with classifying Diabetic Retinopathy images between its early and advanced stages. Each combination was run multiple times and the AUC, Validation Loss, Validation Accuracy, and number of training epochs until convergence of each run was all collected. Maximum Pool and Average Pool were both found to be superior to Variable Stride when deployed in these scenarios.