Custom transformer implementation for edge

Official URL: https://dx.doi.org/10.1109/FMEC65595.2025.11119349

Although deep learning models have greatly improved image processing capabilities, they can be difficult to implement in resource-constrained edge environments. This is mainly due to their high energy consumption and computational requirements. Transformers implement a hierarchical approach to interpret images, differentiating them from traditional convolutional approaches in computer vision. Locality-Sensitive Hashing (LSH) is a widely-used mechanism to use clustering and exploit computational similarities. In this paper, we implement a Transformer with LSH as a hardware accelerator for edge computing environments. Our preliminary results indicate that it is possible to achieve a 1.35x speedup and a 68% power reduction with a 0.55% accuracy loss.