Breaking New Ground in AI with Posit Arithmetic and Vision Transformer

In the dynamic realm of artificial intelligence, the quest for more efficient and precise deep learning models stands as an enduring challenge. Low-bit encoding presents a transformative approach to neural network representation by reducing their precision to lower-bit formats, thereby substantially curtailing memory usage and computational demands. Vision Transformer (ViT) have garnered substantial attention at the intersection of Transformers and Computer Vision. However, ViTs are computationally expensive when dealing with large-scale models and datasets, and encoding ViT with low-bit arithmetic has significant advantages in terms of computational efficiency and resource requirements. This paper introduces a low-bit encoding of Vision Transformer using Posit Arithmetic, a novel number system known for its precision, reduced execution time, storage economy, and lower power consumption compared to traditional floating-point numbers. We conducted a comparative study involving low-precision Posit empowered Stacked Convolutional Neural Networks (Stacked CNNs), conventional ViT, and Stacked Vision Transformer (Stacked ViT). Our findings reveal that the Stacked Vision Transformer competes effectively with the Stacked Convolutional Neural Network baseline with respect to training time and accuracy. This underscores the potential of low-bit encoding using Posit as a compelling solution for AI applications that must adhere to stringent computational constraints. Through meticulous experimentation and comprehensive evaluation, this paper unveils the remarkable advantages of low-bit encoding, offering new insights into the future of efficient and accurate deep-learning models.