Yeong Hwan Oh 오영환 IRIS Lab, Sungkyunkwan Univ.
Index / Works / 02
Exhibit 02 of 04 Early Access
IEEE IoT-JEarly Access DiP-MEMHD

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

Fitting multi-centroid hyperdimensional classifiers onto compact in-memory computing arrays by pruning dimensions without losing what matters.

Fig. 02 Schematic
DiP-MEMHD: pruning dimensions to fit a compact IMC array Schematic. A long class hypervector of D dimensions is pruned to D-prime, the number of array rows, using calibration and variance-based weighting. Each class occupies several centroid columns, so the whole compact array is used in one search cycle. FIG. 02 / DIMENSION PRUNING FIG. 02 SUBJECT DIP-MEMHD SCALE N.T.S. D DIMENSIONS CLASS HV PRUNE CALIBRATION + VARIANCE WEIGHT D′ A1 A2 B1 B2 C1 C2 D1 D2 COLUMNS = CLASSES × CENTROIDS ROWS = D′ SIMILARITY IN ONE SEARCH CYCLE → ARGMAX COMPACT IMC 64 × 64 / 128 × 128 +22.4% ACC. VS. BINARY HDC 6–8× ENERGY EFFICIENCY +2.47% ACC. VS. PRUNING METRICS
Fig. 02 The hypervector dimension D is pruned down to the array height, and each class is spread over several centroid columns, so every row and column of a compact array does work in a single search cycle.

Specifications

As reported in the paper
Memory efficiency (MEMHD) 10.60× at comparable accuracy
Accuracy at equal memory (MEMHD) +14.61%
vs. existing pruning metrics +2.47% accuracy
vs. state-of-the-art binary HDC +22.4% accuracy
Energy efficiency on compact arrays 6–8× better

Abstract

IEEE IoT-J

This work presents MEMHD, a memory-efficient multi-centroid HDC framework, together with DiP-MEMHD that prunes dimensions while preserving critical information through calibration and variance-based weighting. 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 shows 2.47% higher accuracy than existing pruning metrics and 22.4% higher accuracy than state-of-the-art binary HDC models on compact IMC arrays (64x64, 128x128) with 6–8x better energy efficiency, enabling always-on IoT classification on compact edge devices.

* D. Y. Kang and Y. H. Oh contributed equally as co-first authors. † Co-corresponding authors: K. E. Jeon and J. H. Ko.

Cite

@article{kang2026dipmemhd,
  title = {Multi-Centroid Hyperdimensional Computing for Compact {IMC} Arrays via Dimension Pruning},
  author = {Kang, Do Yeong and Oh, Yeong Hwan and Hwang, Chanwook and Kim, Jinhee and Jeon, Kang Eun and Ko, Jong Hwan},
  journal = {IEEE Internet of Things Journal},
  year = {2026},
  doi = {10.1109/JIOT.2026.3730047},
  note = {Early Access},
}