We present a novel approach which is able to explore the configuration of grouped convolutions within neural networks. Group-size Series (GroSS) decomposition is a mathematical formulation of tensor factorisation into a series of approximations of increasing rank terms. GroSS allows for dynamic and differentiable selection of factorisation rank, which is analogous to a grouped convolution. Therefore, to the best of our knowledge, GroSS is the first method to enable simultaneous training of differing numbers of groups within a single layer, as well as all possible combinations between layers. In doing so, GroSS is able to train an entire grouped convolution architecture search-space concurrently. We demonstrate this through architecture searches with performance objectives on multiple datasets and networks. GroSS enables more effective and efficient search for grouped convolutional architectures.
@article{arxiv.1912.00673,
title = {GroSS: Group-Size Series Decomposition for Grouped Architecture Search},
author = {Henry Howard-Jenkins and Yiwen Li and Victor A. Prisacariu},
journal= {arXiv preprint arXiv:1912.00673},
year = {2020}
}