[J39] Multi-Centroid Hyperdimensional Computing for Compact IMC Arrays via Dimension Pruning

Abstract

The implementation of Hyperdimensional Computing (HDC) on In-Memory Computing (IMC) architectures faces fundamental challenges due to dimensional mismatches between high-dimensional vectors and IMC array constraints. Traditional HDC approaches utilize vectors that exceed IMC array dimensions, causing computational resource underutilization and leaving most columns unused during similarity calculations. Existing partitioning methods fail to reduce computation cycles, with challenges intensified for compact IMC arrays (64x64 and 128x128). This paper presents MEMHD, a Memory-Efficient Multi-centroid HDC framework that replaces traditional single-vector-per-class with multiple class vectors to achieve full IMC array utilization and enable single-cycle inference through clustering-based initialization and quantization-aware training. To address compact IMC array deployment, we also propose DiP-MEMHD, Dimension Pruned MEMHD, which prunes dimensions while preserving critical information through calibration and variance-based weighting. As a result, MEMHD achieves 10.60x better memory efficiency at comparable accuracy and 14.61% higher accuracy with the same memory usage compared to existing binary HDC baselines. DiP-MEMHD introduces a multi-centroid-optimized pruning metric, showing 2.47% higher accuracy than existing pruning metrics and achieving 22.4% higher accuracy than state-of-the-art binary HDC models on compact IMC arrays with 6-8x better simulated energy efficiency. These properties make the framework suitable for always-on IoT classification on compact edge devices, such as wearable activity recognition and other sensor-based tasks under tight latency and power budgets. To the best of our knowledge, this is the first multi-centroid HDC framework specifically optimized for IMC array constraints, enabling practical deployment for real-time IoT applications.