Mastering Machine Learning with scikit-learn - Second Edition

Using convolutional neural network activations as features

In recent years, convolutional neural networks (CNNs) have been successfully applied to a variety of tasks, including computer vision tasks such as object recognition and semantic segmentation. We will not discuss CNN in detail in this section; while we will discuss ordinary neural networks like multi-layer perceptrons in a later chapter, scikit-learn is not suitable for deep learning.

As in the word embeddings section, we are interested in CNNs only to use them to extract features to use in other models. We will use the Python bindings for Caffe, a popular deep learning library, and a pre-trained network called CaffeNet to extract features from images. As with word embeddings, we ...