HSViT: Horizontally Scalable Vision Transformer

While the Vision Transformer (ViT) architecture gains prominence in computer vision and attracts significant attention from multimedia communities, its deficiency in prior knowledge (inductive bias) regarding shift, scale, and rotational invariance necessitates pre-training on large-scale datasets. Furthermore, the growing layers and parameters in both ViT and convolutional neural networks (CNNs) impede their applicability to mobile multimedia services, primarily owing to the constrained computational resources on edge devices. To mitigate the aforementioned challenges, this paper introduces a novel horizontally scalable vision transformer (HSViT). Specifically, a novel image-level feature embedding allows ViT to better leverage the inductive bias inherent in the convolutional layers. Based on this, an innovative horizontally scalable architecture is designed, which reduces the number of layers and parameters of the models while facilitating collaborative training and inference of ViT models across multiple nodes. The experimental results depict that, without pre-training on large-scale datasets, HSViT achieves up to 10% higher top-1 accuracy than state-of-the-art schemes, ascertaining its superior preservation of inductive bias. The code is available at https://github.com/xuchenhao001/HSViT.