ShardNet: One Filter Set to Rule Them All

Saumya Jetley, Tommaso Cavallari, Philip Torr, Stuart Golodetz

TL;DR: We compress deep CNNs by reusing a single convolutional layer in an iterative manner, thereby reducing their parameter counts by a factor proportional to their depth, whilst leaving their accuracies largely unaffected

Abstract: Deep CNNs have achieved state-of-the-art performance for numerous machine learning and computer vision tasks in recent years, but as they have become increasingly deep, the number of parameters they use has also increased, making them hard to deploy in memory-constrained environments and difficult to interpret. Machine learning theory implies that such networks are highly over-parameterised and that it should be possible to reduce their size without sacrificing accuracy, and indeed many recent studies have begun to highlight specific redundancies that can be exploited to achieve this. In this paper, we take a further step in this direction by proposing a filter-sharing approach to compressing deep CNNs that reduces their memory footprint by repeatedly applying a single convolutional mapping of learned filters to simulate a CNN pipeline. We show, via experiments on CIFAR-10, CIFAR-100, Tiny ImageNet, and ImageNet that this allows us to reduce the parameter counts of networks based on common designs such as VGGNet and ResNet by a factor proportional to their depth, whilst leaving their accuracy largely unaffected. At a broader level, our approach also indicates how the scale-space regularities found in visual signals can be leveraged to build neural architectures that are more parsimonious and interpretable.

Keywords: neural network compression, filter sharing, network interpretability