This repository contains code for surgically fine-tuning and evaluating an ImageNet-trained ResNet-50 on a data set called Stylized-ImageNet.
Project Motivation:
As society is increasingly finding more novel applications for Convolutional Neural Networks (CNNs) - applications for which there’s a high price to pay for incorrectly classifying images (e.g. autonomous cars and cancer-detection) - it is important to ensure that these models use the same classification strategies as humans. In this work, we show that this can be achieved by fine-tuning an ImageNet-trained ResNet-50 on a data set called Stylized-ImageNet and using a novel fine-tuning technique called ‘surgical fine-tuning’.