Hardware-aware Automated Architecture Search for Brain-inspired Hyperdimensional Computing

Brain-inspired neural network, a.k.a., hyperdimensional computing (HDC), has been becoming a promising candidate for resource-limited edge computing, due to its small size and robustness. However, the previous HDC architecture exploration only considers the software aspects, like HDC operations. This paper presents a hardware-aware automated architecture search framework, namely HwAwHDC, for HDC, which can consider both hardware and software characters in a uniform reinforcement learning based optimization loop. It fills the gap in the automatic design of HDC architectures for given applications and hardware constraints. We do a thorough analysis to formulate the search spaces for HwAwHDC. On top of this, we design a hardware-friendly reward and employ reinforcement learning (RL) to explore the HDC architectures. The encouraging experimental evidence shows the effectiveness of our framework. The identified model achieves a leading score on each task with lower power consumption and higher energy efficiency compared with the deep learning approach and HDC approach.