Autoencoder is another word for autoassocitve memory, but more often used in recent neural network literature. In the training data the input and output are identical, often the same image as each other. The network thius learns to associate the image with itself. This initially does not seem very intersting except that internal layers typically are substantially smaller than the full input. A small middle layer is effectvely coding the key features of the image for reconstruction. Becasue of this the inner, pinch point layer can sometimes be used as a form of data-reduction for other knds of network, for example a classifier. Alternatively the network may be able to recreate images from partial inputs or even generate new ones by setting the inner layer.
Used on pages 118, 272, 273, 282