Learned Features

Deep neural networks learn high-level features in the hidden layers. This is one of their greatest strengths and reduces the need for feature engineering. Assume you want to build an image classifier with a support vector machine. The raw pixel matrices are not the best input for training your SVM, so you create new features based on color, frequency domain, edge detectors and so on. With convolutional neural networks, the image is fed into the network in its raw form (pixels). The network transforms the image many times. First, the image goes through many convolutional layers. In those convolutional layers, the network learns new and increasingly complex features in its layers. Then the transformed image information goes through the fully connected layers and turns into a classification or prediction.

• The first convolutional layer(s) learn features such as edges and simple textures.
• Later convolutional layers learn features such as more complex textures and patterns.
• The last convolutional layers learn features such as objects or parts of objects.
• The fully connected layers learn to connect the activations from the high-level features to the individual classes to be predicted.

Cool. But how do we actually get those hallucinatory images?