pre-trained convolutional neural networks (CNNs) are powerful off-the-shelf
feature generators and have been shown to perform very well on a variety of
tasks. Unfortunately, the generated features are high dimensional and expensive
to store: potentially hundreds of thousands of floats