Hyperdimensional Computing Versus Convolutional Neural Network: Architecture, Performance Analysis, and Hardware Complexity

The interest in brain-inspired computing architectures has been growing, particularly in the context of edge devices with constrained resources for executing cognitive functions. One such approach is hyperdimensional computing (HDC), a novel concept that draws inspiration from the large representation of human neuronal activity. HDC has demonstrated effectiveness in one-dimensional tasks, like text identification and activity recognition, offering advantages in power consumption and response time over convolutional neural networks (CNNs). This paper compares HDC and CNN regarding architecture, accuracy, and hardware complexity, explicitly focusing on image classification tasks. Our findings indicate that CNNs generally outperform HDC in two-dimensional tasks but require significantly more computational resources. In contrast, HDC offers adequate results using just 16% of the data needed for training. Additionally, experiments conducted using a Raspberry Pi 4 show that HDC can enhance inference speed and energy efficiency by approximately 2.5 times relative to CNNs.