Redefining Deep Learning Across NLP, Image Classification and Beyond - Vision Transformers

The self-attention technique was first used to develop transformers for natural language processing. The groundbreaking work “Attention Is All You Need” (2017) for Natural Language Processing(NLP) introduces transformers. They can perform well as set-based systems and require few inductive biases in their design. Furthermore, transformers are superior to traditional models, including long short-term memory (LSTM) networks, on sequence-based tasks because they are excellent at capturing dependencies of long-range and facilitating parallel processing. Transformers are widely used in computer vision in recent years, leading to impressive breakthroughs in the field. Transformer tools for a range of tasks of computer vision, including object detection, activity recognition, and image augmentation, have been summarized in earlier studies and researches. After the transformers are introduced and the successful Vision Transformer (ViT) is highlighted, more enhancements and optimizations are looked into for image classification jobs. The specific topic of this paper is image classification using the Vision Transformer (ViT) which introduces a novel approach to computer vision by leveraging the transformer architecture, previously dominant in NLP, for visual data interpretation. Comparing to Convolutional Neural Networks(CNNs), ViT sub-divides images into fixed size patches, embeds them in sequences, and processes these using multi-head self-attention mechanisms. Then, utilizing experiments using Pytorch of Python, the implementation of ViT has been done successfully. Comparison of ViT with CNNs (Convolutional Neural Networks) for dataset of large size has also been examined. Notable Results are highly competitive on benchmarks with 91.5% accuracy with Vision Transformer.