Vision Transformer for Small-Size Datasets

TL;DR

This paper introduces SPT and LSA modules that enhance Vision Transformers' ability to learn from small datasets by improving locality bias, leading to significant performance gains.

Contribution

The paper proposes generic SPT and LSA modules that enable ViTs to learn effectively from scratch on small datasets, a capability previously reliant on large pre-training datasets.

Findings

Performance improved by 2.96% on Tiny-ImageNet

Swin Transformer performance increased by 4.08%

Modules are easily applicable to various ViTs

Abstract

Recently, the Vision Transformer (ViT), which applied the transformer structure to the image classification task, has outperformed convolutional neural networks. However, the high performance of the ViT results from pre-training using a large-size dataset such as JFT-300M, and its dependence on a large dataset is interpreted as due to low locality inductive bias. This paper proposes Shifted Patch Tokenization (SPT) and Locality Self-Attention (LSA), which effectively solve the lack of locality inductive bias and enable it to learn from scratch even on small-size datasets. Moreover, SPT and LSA are generic and effective add-on modules that are easily applicable to various ViTs. Experimental results show that when both SPT and LSA were applied to the ViTs, the performance improved by an average of 2.96% in Tiny-ImageNet, which is a representative small-size dataset. Especially, Swin Transformer achieved an overwhelming performance improvement of 4.08% thanks to the proposed SPT and LSA.