ANAS: Software–hardware co-design of approximate neural network accelerators via neural architecture search

Abstract

Deep Neural Networks (DNNs) are prevalent solutions for perception tasks, with energy efficiency being particularly critical for deployment on edge platforms. Various studies have proposed efficient DNN edge deployment solutions; however, an important aspect – approximate computing – has been overlooked. Current research primarily focuses on designing approximate circuits for specific DNN models, neglecting the influence of DNN architecture design. To address this gap, this paper proposes a software–hardware co-exploration framework for approximate DNN accelerator design that jointly explores approximate multipliers and neural architectures. This framework, termed Approximate Neural Architecture Search (ANAS), tackles two main challenges: (1) efficiently evaluating the impact of approximate multipliers on application performance and accelerator design for each sample, and (2) effectively navigating a large design space to identify optimal configurations. The framework employs a recurrent neural network-based reinforcement learning algorithm to identify an optimal approximate multiplier-DNN architecture pair that balances DNN accuracy and hardware cost. Experimental results demonstrate that ANAS achieves comparable accuracy while reducing energy consumption and latency by up to 40% compared to state-of-the-art NAS-based methods.