Topographical Internal Representation in Deep Neural Networks

The development of Deep Neural Networks has been making significant impacts for the fields of Artificial Intelligence over the last few years. While most of the Deep Neural Networks research concentrate on improving the performances of the networks, less effort is dedicated for understanding the reasons of those outstanding performances. This short paper attempts to explain that topographical internal representations in hierarchical neural networks can emerge from a supervised learning mechanism. Topographical representations are of interest because of its relevance with biological neural networks, and thus potentially offers new insights for the deeper understanding on artificial neural networks.