In the previous lessons, you looked at building a binary classifier that predicted cats versus dogs or horses versus humans. You also saw how overfitting can occur and explored some practices for avoiding it. The problem with these of course, is that the training data was very small, and there's only so many common features that can be extracted, even if we do some tricks like image augmentation. But in both these cases, you built the model from scratch. What if you could take an existing model that's trained on far more data, and use the features that that model learned? That's the concept of transfer learning, and we'll explore that in this lesson. So for example, if you visualize your model like this with a series of convolutional layers before dense layer leads your output layer, you feed your data into the top layer, the network learns the convolutions that identify the features in your data and all that. But consider somebody else's model, perhaps one that's far more sophisticated than yours, trained on a lot more data. They have convolutional layers and they're here intact with features that have already been learned. So you can lock them instead of retraining them on your data, and have those just extract the features from your data using the convolutions that they've already learned. Then you can take a model that has been trained on a very large datasets and use the convolutions that it learned when classifying its data. If you recall how convolutions are created and used to identify particular features, and the journey of a feature through the network, it makes sense to just use those, and then retrain the dense layers from that model with your data. Of course, well, it's typical that you might lock all the convolutions. You don't have to. You can choose to retrain some of the lower ones too because they may be too specialized for the images at hand. It takes some trial and error to discover the right combination. So let's take a well-trained state of the art model. There's one called Inception, which you can learn more about at his site. This has been pre-trained on a dataset from ImageNet, which has 1.4 million images in a 1000 different classes.