Involution: Inverting the Inherence of Convolution for Visual Recognition

TL;DR

This paper introduces involution, a novel neural network operation that inverts convolution principles, unifies self-attention, and improves visual recognition performance across multiple benchmarks with reduced computational cost.

Contribution

The paper proposes involution, a new operation that replaces convolution by inverting its principles, and demonstrates its effectiveness and efficiency in various vision tasks.

Findings

Involution improves ResNet-50 top-1 accuracy by up to 1.6%.

Involution reduces computational cost to about 65-72% of convolution.

Involution unifies self-attention as a special case.

Abstract

Convolution has been the core ingredient of modern neural networks, triggering the surge of deep learning in vision. In this work, we rethink the inherent principles of standard convolution for vision tasks, specifically spatial-agnostic and channel-specific. Instead, we present a novel atomic operation for deep neural networks by inverting the aforementioned design principles of convolution, coined as involution. We additionally demystify the recent popular self-attention operator and subsume it into our involution family as an over-complicated instantiation. The proposed involution operator could be leveraged as fundamental bricks to build the new generation of neural networks for visual recognition, powering different deep learning models on several prevalent benchmarks, including ImageNet classification, COCO detection and segmentation, together with Cityscapes segmentation. Our involution-based models improve the performance of convolutional baselines using ResNet-50 by up to 1.6% top-1 accuracy, 2.5% and 2.4% bounding box AP, and 4.7% mean IoU absolutely while compressing the computational cost to 66%, 65%, 72%, and 57% on the above benchmarks, respectively. Code and pre-trained models for all the tasks are available at https://github.com/d-li14/involution.