NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:1100
Title:Neural networks grown and self-organized by noise


		
The authors propose a developmental algorithm that grows a spiking neural architecture in a self-organized manner from local rules. A CNN-like retinotopic connectivity structure is emerging. The model is applied to several setups, including the MNIST dataset. They show that the grown network extracts useful features and can performs as well as hand-crafted networks. The work is very original and the results are interesting. The paper is well-written and technically sound. On the negative side, the terminology is not always clear and the results do not seem overly surprising. The authors submitted a strong Author's response which clarified many points and was well-received by the reviewers. The authors are urged to fix their vocabulary: describing the emergent receptive fields as "convolutional" (and, quite likely, "pooling") is inappropriate due to lack of weight-sharing; the correct term that should be used is "retinotopic".