Is there any research on neural networks with multiple outputs for hierarchical label classification?

Question

I had this idea of training for example a CNN on images, and having output branches at several of its intermediate layers. The early layers' output branch might then predict high-level class of detected objects (supposedly able to do this because less info is needed for a high-level classification than a very specialised one), and the later layers giving more detailed labels of the sub-class of the earlier high level class.

I have been searching for research on this type of setup but couldn't really find anything. Is there a name for this idea, or is this an open question/idea?

Answer 1

Sounds like a very interesting idea! I dont know existing work on the idea, but the implementation should be pretty simple with separate training with tied weights. It should be noted however that such behaviour occurs naturally in CNN's already. (See: http://www.cs.toronto.edu/~guerzhoy/321/lec/W07/HowConvNetsSee.pdf)