Randomized Search on a Grid of CNN Networks with Simplified Search Space

Abstract

One of the prime issues in Convolutional Neural Networks (CNN) is the design of the architecture, which is mainly human crafted, requiring significant time and resources, including expert knowledge, as the number of design choices for CNN is quite large given the number of choices in the parameters of the CNN. In this paper, we analyze the different neural architecture search (NAS) approaches that have been used in recent times, and their issues, and propose a novel method of performing neural architecture search. Our proposed model uses a simplified search space, with a randomized search strategy. We utilize a cell-based architecture search method, with a cell having multiple CNN operations, along with the multiple link options within the operation nodes of a cell. The proposed model is then tested on the MNIST dataset, with significant comparable performance with state of art architecture for MNIST.