2019 Dynamic Block Sparse Reparameterization of Convolutional Neural Networks(India)

1. Abstract

1.1. In this work, we focus on block sparsity and generate efficient block sparse convolutional neural networks using DBSR(dynamic block sparse reparameterization)

2. Introduction

2.1. Sparsification of neural networks 可以reduce flops，但是对GPU这类硬件不适用。

2.1.1. Can we generate sparse neural networks whose compute is amenable to parallel hardware like GPU?

2.1.1.1. inducing structure in to sparsity