Dynamic Clone Transformer for Efficient Convolutional Neural Netwoks

TL;DR

This paper introduces the dynamic clone transformer, a novel dual-branch module for convolutional neural networks that enhances efficiency and capacity by generating and reforming clones of input features, aiming to improve performance on resource-limited devices.

Contribution

It proposes the dynamic clone transformer (DCT), a new module inspired by MPFC, that improves ConvNet efficiency and capacity by self-expanding channel-wise information with low computational cost.

Findings

DCT effectively replaces pointwise convolutions in bottleneck structures.

DCT improves accuracy-efficiency trade-off in ConvNets.

DCT demonstrates competitive performance on vision tasks.

Abstract

Convolutional networks (ConvNets) have shown impressive capability to solve various vision tasks. Nevertheless, the trade-off between performance and efficiency is still a challenge for a feasible model deployment on resource-constrained platforms. In this paper, we introduce a novel concept termed multi-path fully connected pattern (MPFC) to rethink the interdependencies of topology pattern, accuracy and efficiency for ConvNets. Inspired by MPFC, we further propose a dual-branch module named dynamic clone transformer (DCT) where one branch generates multiple replicas from inputs and another branch reforms those clones through a series of difference vectors conditional on inputs itself to produce more variants. This operation allows the self-expansion of channel-wise information in a data-driven way with little computational cost while providing sufficient learning capacity, which is a potential unit to replace computationally expensive pointwise convolution as an expansion layer in the bottleneck structure.