One of the few disadvantages deep learning has over traditional machine learning is that it requires lots of data. Transfer learning is one way to overcome this, by using the weights of a previously trained model (usually trained on ImageNet data) and then applying them to a new problem set.
The ImageNet dataset consists of 15 million images in 1,000 classes. Since we can reuse parts of a model that has been trained on this amount of data, it may be possible to train the new model with just a few hundred images per category. This would depend on the images being somewhat related to the data used in the original model. For example, trying to use transfer learning from ImageNet models (which is trained on photographs) on data ...
