Optimal Layer Selection on Deep Convolutional Neural Networks Using Backward Freezing and Binary Search

Herrera Del Aguila H.M.; Herrera Quispe J.A. · Communications in Computer and Information Science (2023) · DOI: 10.1007/978-3-031-35445-8_13

Resumen

Transfer Learning in Deep Convolutional Neural Networks is a highly used method for image classification, the performance depends on the selection of the layers to be frozen; the search for the optimal layer where the backward freezing will be applied becomes increasingly complicated as the number of layers increases. We found an approximation of the optimal layer to apply backward freezing using a modified Binary Search method. We experimented with InceptionV3, Xception, DenseNet121, NasNetMobile, and ResNet50; the case of study is the classification of melanomas of benign and malignant class from Kaggle. We use partitions with four-folds and measure the performance with the ROC AUC score. Our results show an improvement in the ROC AUC score for ResNet50 and no significant change for InceptionV3. Code: https://www.kaggle.com/code/hherrera007/binary-search.

Instituciones: Universidad Nacional Mayor de San Marcos