Over the years, many organizations, research groups, and individuals within the open source community have perfected some complex models trained using gigantic amounts of data for generic use cases. In some cases, they have invested years of effort in optimizing these models. Some of these open source models can be used for the following applications:

• Object detection in video
• Object detection in images
• Transcription for audio
• Sentiment analysis for text

Whenever we start working on training a new machine learning model, the question to ask ourselves is this: instead of starting from scratch, can we simply customize a well-established pre-trained model for our purposes? In other words, can we transfer the learning of existing models to our custom model so that we can answer our business question? If we can do that, it will provide three benefits:

• Our model training efforts will be given a jump-start.
• By using a well-tested and well-established model, the overall quality of our model is likely to improve.
• If we do not have enough data for the problem we are working on, using a pre-trained model through transfer learning may help.

Let's look into two actual examples where this would be useful:

• When training a robot, we could first use a simulation game to train a neural network model. In that simulation, we could create all those rare events that are very hard to find in the real world. Once trained, we could use transfer learning ;to train the model for the real world.
• Let's assume that we want to train a model that can classify Apple And Windows laptops from a video feed. There are already well-established object detection models available as open source that can accurately classify various objects in a video feed. We can use these models as a starting point and identify objects as laptops. Once we have identified the objects as laptops, we can further train the model to differentiate between Apple and Windows laptops.

In the next section, we will apply the concepts that we have covered in this chapter to building a fraudulent document classifying neural network.