Energy-Efficient Sparse Hyperdimensional Computing for Speech Recognition

This study proposes an energy-efficient hyperdimensional computing (HDC) architecture for speech recognition. Energy efficiency is achieved through the use of 1) sparse hypervectors, 2) architecture co-optimized for optimal accuracy-energy trade-off, and 3) redundancy pruning. Sparse hypervectors contribute to computational efficiency by reducing the number of 1s in the hypervectors. Implementing and optimizing segmentation in the architecture enables processing only a portion of the hypervector at a time, promoting hardware reuse, thus reducing both area and energy usage. The integration of redundancy pruning significantly conserves energy by disabling bit positions in hypervectors that do not contribute useful information during inference. The effectiveness of these techniques is demonstrated using the ISOLET dataset. The base HDC model, employing 4096-bit hypervectors with a hypervector density of 2% and a segmentation size of 128, achieves a classification accuracy of 83.71%. With redundancy pruning, the proposed model maintains the same accuracy while eliminating 45.43% redundant bit positions, resulting in 2.26% energy savings per query and 1.02x enhanced energy efficiency at the cost of 11.68% increase in area.