A lot of development has happened within the deep learning domain in recent years, to enhance algorithmic efficacy and computational efficiency across different domains such as text, images, audio, and video. However, when it comes to training on new datasets, machine learning usually rebuilds the model from scratch, as is done in traditional data science problem solving. This becomes challenging when a new big dataset need to be trained as it will require very high computation power a lot of and time to reach the desired model efficacy.

Transfer Learning is a mechanism to learn new scenarios from existing models. This approach is very useful to train on big datasets, not necessarily from a similar domain or problem statement. For example, researchers have shown examples of Transfer Learning where they have trained Transfer Learning for completely different problem scenarios, such as when a model built using classifications of cat and dog is used for classifying objects such as aeroplane vs automobile.

In terms of analogy, it's more about passing the learned relationship to new architecture in order to fine-tune weights. An example of how Transfer Learning is used is shown in the following figure:

The figure shows the steps of Transfer Learning, where the weights/architectures from a predeveloped deep learning model are reused to predict a new problem statement. Transfer Learning helps provide a good starting point for deep learning architectures. There are different open source projects going on in different domains, which facilitate Transfer Learning, for example, ImageNet (http://image-net.org/index) is an open source project for image classification where a lot of different architectures such as Alexnet, VGG16, and VGG19 have been developed. Similarly, in text mining, there is a Word2Vec representation of Google News trained using three billion running words.