A Novel Encoding Scheme for Complex Neural Architecture Search

Recent boom in Image Processing can be attributed mostly to Artificial Neural networks, enormity of data and availability of high-performance hardware. ANNs need to be designed with care and a lot of tweaking is required according to the problem at hand. However, they can be designed automatically by deploying search algorithms given a set of possible values for hyperparameters. Previously, network design has been limited because of the inherent complexity of network diversity in terms of their style (e.g. VGG-style or ResNet-style). Aforementioned two network styles require different kinds of parameters hence limiting the search algorithm to output only a specific style of architecture. We propose a methodology which will allow the system to search for the best architecture with least possible constraints.