Approximate DNN Accelerator

Approximate DNN accelerators aim to improve the energy efficiency and speed of deep neural network inference by employing approximate computing techniques, such as low-precision arithmetic and inexact multipliers, within hardware implementations. Current research focuses on optimizing these accelerators for various architectures, including convolutional neural networks (CNNs) and vision transformers (ViTs), exploring different quantization methods (e.g., power-of-two quantization) and developing frameworks for efficient approximation-aware training and deployment. This research is significant because it addresses the high energy consumption of DNNs, enabling their deployment on resource-constrained edge devices and potentially leading to more sustainable and efficient AI systems.

Papers