Explaining multiple causes using epitomes:

We can learn the epitome of a collection of images and use it as a summary and a retrieval tool.

For illustration, we computed an epitome of a set of 295 facial images. The dataset contained various people with various facial expressions. We can then identify a feature in the epitome and retrieve all images whose epitome mapping included the feature with high probability.  For instance, bellow we show the images with the highest and the the lowest posterior of using the smiling point in the epitome.

 

 

We can also search for a particular combination of features in the epitome. Below, we show images with high posterior at the features corresponding to a closed eye, a smile, and dark hair. Then we modify the query slightly by changing only the first feature to a certain type of an open eye.