Understanding Vision Transformers through Intuition and Simple Visual Examples

Vision Transformers are normally used in image recognition and are considered as an alternative to Convolutional Neural Networks. Unlike traditional convolution-based models, a Vision Transformer does not rely on local filters Instead, it segments an image into smaller patches and understands the relationships among individual patches. This technique is powerful, but it can be challenging for beginners to understand when described only with equations or programming concepts. Hence, this paper showcases the process of Vision Transformers using simple examples and visual illustrations. This paper is intended for students, teachers, and beginners who are interested in learning how transformer models are used in computer vision. The tutorial explains how an image is divided into patches, how each patch is transformed into a token, how self-attention helps different image regions share information, and how the transformer encoder helps the final classification process. The foremost purpose of this tutorial is to support readers in developing a clear understanding of Vision Transformers. After gaining this intuitive understanding, readers can more easily study the mathematical particulars and implementation techniques later.