Transfer Learning

Assume that you have been assigned the responsibility of developing an app that can classify ten new musical instruments developed recently by renowned musicians. You only have a few hundred images of these instruments. To classify these images, you decide to use GoogLeNet architecture, which has around 6.8 million parameters. If you decide to train the model using the given images, you will realize that you have an insufficient number of images. However, if you train the model using the pictures of known instruments, you might be able to train it, but it will take a lot of time and computational resources. So the challenge is to train a sufficiently complex model on a dataset that does not have a sufficient number of images, and you probably do not have GPU as well.