Unsupervised Dynamic Routing Via Competition Over Network Loss

Lei Zhao, Lin Wang, Jiuqiao Yu

Keywords: unsupervised learning, deep learning, self organizing map, back propagation

Abstract: This paper proposes a novel neural network architecture that can simultaneously do normal network optimizing while attaining the ability of unsupervised learning. Almost all existing unsupervised learning algorithms are based on doing calculations on the input space or feature space, this paper proposes a new possibility to discover a structure in the functional space without supervision. Using the self-organizing map over the competition of the loss of individual neural column, we route the input to the most appropriate modules dynamically, by doing this we separate the input functional space into different sub spaces which are represented by each individual neural column. At the end of the paper, we propose several possible architectures based on the philosophy of this paper that could build a neural network system block by block.

Code Of Conduct: I certify that all co-authors of this work have read and commit to adhering to the NeurIPS Statement on Ethics, Fairness, Inclusivity, and Code of Conduct.