In the Mind's Eye
Researchers have estimated that the number of objects categories our eyes can recognize in a brief instant is around 50,000.
Given the computational "limitations" of the human brain our eyes are probably not scanning all 50,000 templates of known objects and picking the best match of what we are seeing.
The brain must have a way to categories objects in a smart way, for instance in a tree structure, so the human brain can search much more efficiently and figure out what it is looking at.
Other advantages to this hierarchy are new object categories can be learned from very few examples by merely drawing information from visually similar objects (so called transfer learning or one-shot learning).
For instance, a cat looks just like a dog, but is a bit bigger, less fluffy and has much sharper nails.
Professor Max Welling’s research group, in collaboration with Pietro Perona's group at Caltech, are using new statistical techniques to infer these representations in an unsupervised setting from image data to learn more about these visual object class taxonomies.
The statistical techniques are based on the Dirichlet process and are able to induce model structure from data: the more data is presented to the system, the more complex the representation is allowed to grow.
