gSwin: Gated MLP Vision Model with Hierarchical Structure of Shifted Window

TL;DR

gSwin is a novel vision model that combines Swin Transformer and gMLP, achieving improved accuracy and smaller size across multiple vision tasks by integrating hierarchical structure and locality.

Contribution

This paper introduces gSwin, a new model merging Swin Transformer and gMLP, enhancing efficiency and performance in image recognition tasks.

Findings

Outperforms Swin Transformer in accuracy on multiple tasks

Achieves smaller model size while maintaining high performance

Effective integration of transformer and MLP architectures

Abstract

Following the success in language domain, the self-attention mechanism (transformer) is adopted in the vision domain and achieving great success recently. Additionally, as another stream, multi-layer perceptron (MLP) is also explored in the vision domain. These architectures, other than traditional CNNs, have been attracting attention recently, and many methods have been proposed. As one that combines parameter efficiency and performance with locality and hierarchy in image recognition, we propose gSwin, which merges the two streams; Swin Transformer and (multi-head) gMLP. We showed that our gSwin can achieve better accuracy on three vision tasks, image classification, object detection and semantic segmentation, than Swin Transformer, with smaller model size.