The field of edge AI is moving towards developing more efficient and scalable solutions for real-time applications. Researchers are exploring innovative methods to reduce computational costs and memory usage, enabling the deployment of accurate models on resource-limited edge devices. Notably, quantization techniques, sparse modeling, and optimized training algorithms are being developed to improve performance and efficiency. These advancements have the potential to enhance various applications, including depth estimation, face recognition, and computer vision.
Some noteworthy papers in this area include: QuartDepth, which proposes a post-training quantization method for real-time depth estimation on edge devices, achieving competitive accuracy while enabling fast inference and higher energy efficiency. PRIOT, which introduces a pruning-based integer-only transfer learning method for embedded systems, improving accuracy by up to 33.75 percentage points over existing methods. HOT, which presents a Hadamard-based optimized training approach, achieving up to 75% memory savings and a 2.6 times acceleration on real GPUs with negligible accuracy loss.