Learning from Hypervectors: A Survey on Hyperdimensional Encoding

Sercan Aygun; Mehran Shoushtari Moghadam; M. Hassan Najafi; Mohsen Imani

cs.LGAug 1, 2023

Learning from Hypervectors: A Survey on Hyperdimensional Encoding

Authors: Sercan Aygun, Mehran Shoushtari Moghadam, M. Hassan Najafi, Mohsen Imani

Organizations: Sch. Comp. and Info., Univ. of Louisiana at Lafayette, 70503 Lafayette, LA, USA · Comp. Sci. and Eng., Univ. of California Irvine, 92697 Irvine, CA, USA

Abstract

Hyperdimensional computing (HDC) is an emerging computing paradigm that imitates the brain's structure to offer a powerful and efficient processing and learning model. In HDC, the data are encoded with long vectors, called hypervectors, which typically range from 1,000 to 10,000 in length. The literature provides several encoding methods to generate orthogonal or correlated hypervectors, based on the application. Existing surveys focus on overall HDC system aspects like inputs, primary computations, and outputs. However, this study takes a more specific approach. It zeroes in on the HDC system input and the generation of hypervectors, directly influencing the hypervector encoding. This survey brings together various hypervector generation and encoding methods, exploring the limitations, challenges, and potential benefits they entail.

Figures & tables

Panel of Figure 1 — Figure 1. Overview of an HDC system for classification, together with a compact reference for the representation and operation conventions used throughout this survey. The main stages are: I encoding incoming data, II training with the input dataset, and III computing similarity between class-based $\boldsymbol{\mathcal{HV}}$ s ( $C_{\mathrm{class}}$ ) and the query $\boldsymbol{\mathcal{HV}}$ . The right-side quick reference summarizes binary / bipolar domains, binding operators, bundling and thresholding conventions, and permutation notation.

Figure 2. PRISMA-inspired literature search, screening, and inclusion workflow used in this survey.

Figure 3. Comparison of learning : HDC vs. NNs, together with a compact summary of HDC encoding and ML-based representation learning.

Target	SOTA Works	Target	SOTA Works
Image Processing/Computer Vision	( Morris et al., 2022b ; Poduval et al., 2022 ; Imani et al., 2020a ; Kang et al., 2022b ; Imani et al., 2019f ; Mitrokhin et al., 2020 ; Imani et al., 2019d ; Ma et al., 2023a ; Kleyko et al., 2018b ; Duan et al., 2022 ; Ma and Jiao, 2022 ; Zhao et al., 2022 ; Khaleghi et al., 2022 ; Kovalev et al., 2022 ; Montone et al., 2017 ; Kirilenko et al., 2021 ; Hersche et al., 2020 ; Watkinson et al., 2021 ; Duan et al., 2022 ; Miranda and d’Aliberti, 2022 ; Dutta et al., 2022 ; Peter Sutor et al., 2022 ; Ma et al., 2023c ; Hersche et al., 2023b ; Wilson et al., 2023 ; Neubert et al., 2021 ; Menet et al., 2023 ; Hersche et al., 2023a ; Aygun et al., 2024 ; Moghadam et al., 2023 ; Imani et al., 2022 ; Redding et al., 2023 ; Kim et al., 2021a ; Yu et al., 2022 ; Moghadam et al., 2026 ; Masum et al., 2025 )	Language/Text Processing	( Wu et al., 2018a ; Rahimi et al., 2016b ; Räsänen, 2015 ; Kleyko et al., 2018b ; Karunaratne et al., 2021b ; Moraliyage et al., 2022 ; Kovalev et al., 2022 ; Montone et al., 2017 ; Kirilenko et al., 2021 ; Mohammad Navaz et al., 2023 ; Liu et al., 2022 ; Aygun et al., 2023 ; Moghadam et al., 2026 )
Biomedical/Biometrics	( Imani et al., 2018a ; Nazemi et al., 2020 ; Rahimi et al., 2019 ; Zou et al., 2022a ; Burrello et al., 2019 ; Chang et al., 2019 ; Imani et al., 2020a ; Kang et al., 2022b ; Imani et al., 2019f ; Imani et al., 2020b ; Imani et al., 2019d ; Rahimi et al., 2016a ; Burrello et al., 2020 ; Burrello et al., 2018 ; Menon et al., 2022a ; Basaklar et al., 2021 ; Moin et al., 2018 ; Watkinson et al., 2021 ; Hersche et al., 2018 ; Gupta et al., 2022 ; Thomann et al., 2023 ; Rahimi et al., 2020 ; Schindler and Rahimi, 2021 ; Pale et al., 2022 ; Xu et al., 2023a ; Yang et al., 2022 ; Pinge et al., 2023 ; Kim et al., 2021a ; Moghadam et al., 2026 ; Masum et al., 2026 ; Masum et al., 2025 )	Cross-paradigm Computing †	( Poduval et al., 2021b ; Liu et al., 2019 ; Bosch et al., 2019 ; Hao et al., 2021 ; Moghadam et al., 2023 ; Aygun et al., 2023 ; Orchard and Jarvis, 2023 ; Zou et al., 2022b ; Bent et al., 2022 ; Zhang et al., 2023 ; Aygun et al., 2024 ; Imani et al., 2022 ; Fayza et al., 2023 ; Moghadam et al., 2026 ; Masum et al., 2026 ; Yu et al., 2026 )
Security/Reliability Issues	( Amrouch et al., 2022 ; Ma et al., 2023a ; Ma et al., 2023b ; Cai et al., 2022 ; Yang and Ren, 2020b ; Moraliyage et al., 2022 ; Gungor et al., 2022 ; Gungor et al., 2023a ; Wang et al., 2023a ; Khaleghi et al., 2020a ; Yang and Ren, 2020a ; Wang et al., 2022 ; Ma et al., 2021 ; Wang and Jiao, 2022 ; Chen and Li, 2021 ; Gungor et al., 2023b ; Genssler et al., 2024a ; Barkam et al., 2023a ; Barkam et al., 2023b ; Park et al., 2023 ; Kleyko et al., 2018a ; Kumar et al., 2022 ; Nam et al., 2023 ; Yu et al., 2023 ; Guirado et al., 2023 ; Liu et al., 2023 ; Khaleghi et al., 2024 ; Zhang et al., 2025 ; Jeong et al., 2025 ; Barkam et al., 2024 ; Justus Rajappa et al., 2026 )	Graph Processing	( Poduval et al., 2022 ; Nunes et al., 2022 ; Kang et al., 2022c ; Amrouch et al., 2022 ; Zhou et al., 2021 ; Chen et al., 2023 ; Cornell et al., 2022 ; Nickel et al., 2016 ; Yi and Achour, 2023 )
Memory-centric Computing	( Morris et al., 2022a ; Poduval et al., 2022 ; Poduval et al., 2021b ; Imani et al., 2020b ; Karunaratne et al., 2020a ; Kazemi et al., 2021 ; Li and others, 2016 ; Imani et al., 2020 ; Dutta et al., 2022 ; Karunaratne et al., 2021a ; Poduval et al., 2021a ; Karunaratne et al., 2020b ; Langenegger et al., 2023 ; Guirado et al., 2023 ; Karunaratne et al., 2022 ; Guirado et al., 2022 ; Thomann et al., 2022 ; Cai et al., 2022 ; Yang et al., 2023 ; Gupta et al., 2022 ; Liu et al., 2019 ; Teeters et al., 2023 ; Genssler et al., 2023 ; Mohammad Navaz et al., 2023 ; Kazemi et al., 2022b ; Wang et al., 2023e ; Masum et al., 2025 ; Yun et al., 2026b )	DNA - Genome Application	( Kim et al., 2020 ; Imani et al., 2018b ; Cumbo et al., 2020 ; Zou et al., 2022c ; Poduval et al., 2021c ; Chen and Imani, 2022 ; Barkam et al., 2023b )
IoT & Edge AI	( Gungor et al., 2023b ; Khaleghi et al., 2021 ; Datta et al., 2019 ; Kim et al., 2018 ; Gungor et al., 2023a ; Amrouch et al., 2022 ; Zou et al., 2021a ; Khaleghi et al., 2022 ; Basaklar et al., 2021 ; Kandaswamy et al., 2022 ; Genssler et al., 2022 ; Yu et al., 2023 ; Taheri et al., 2022 ; Huang et al., 2025 ; Ding et al., 2026 ; Zhang et al., 2026 ; Yu et al., 2026 ; Yun et al., 2026b ; Tang et al., 2026 ; Yun et al., 2026a ; Justus Rajappa et al., 2026 )	Robotics	( Menon et al., 2022b ; Hassan et al., 2024 ; Mitrokhin et al., 2019 ; Neubert et al., 2019 ; McDonald et al., 2021 ; Neubert et al., 2021 )

Table 1. First Glance of Prior HDC Works Classified based on Prominent Targets

HV Type	Domain	Binding	Unbinding	Bundling	Binarization	Similarity
Binary	$\{0,1\}^{D}$	XOR : $(\mathbf{h}_{a}\oplus\mathbf{h}_{b})_{i}$	XOR again	Majority vote / pop-count	Majority threshold	Hamming Distance: $\frac{1}{D}\\|\mathbf{h}_{a}\oplus\mathbf{h}_{b}\\|_{1}$
Bipolar	$\{-1,+1\}^{D}$	Hadamard: $(\mathbf{h}_{a}\odot\mathbf{h}_{b})_{i}=h_{a,i}h_{b,i}$	Hadamard again	$\sum_{k}\mathbf{h}_{k}$	$\text{sgn}(\cdot)$	Cosine: $\frac{1}{D}\sum_{i}h_{a,i}h_{b,i}$
Real-valued	$\mathbb{R}^{D}$ (unit-norm)	Hadamard	Approximate (re-normalize)	$\frac{1}{n}\sum_{k}\mathbf{h}_{k}$	$\text{sgn}(\cdot)$ or quantize	Cosine: $\mathbf{h}_{a}\cdot\mathbf{h}_{b}/(\\|\mathbf{h}_{a}\\|\\|\mathbf{h}_{b}\\|)$
Complex (phasor)	$\mathbb{C}^{D}$ $(\|h_{j}\|=1)$	Phase add: $(\mathbf{h}_{a}\odot\mathbf{h}_{b})_{j}=e^{i(\theta_{a,j}+\theta_{b,j})}$	Conj. mult.: $\mathbf{h}_{c}\odot\overline{\mathbf{h}_{b}}$	Superposition $\rightarrow$ renormalize: $\frac{\sum_{k}h_{k,j}}{\|\sum_{k}h_{k,j}\|}$	Normalize to unit circle	Inner product: $\frac{1}{D}\text{Re}(\mathbf{h}_{a}\cdot\overline{\mathbf{h}_{b}})$

Table 2. Summary of

\boldsymbol{\mathcal{HV}}

Types and Operations

HDC Surveys	Year	Vector Mapping	Random Source Variety Discussion	Open-source Platform Discussion	Cross-paradigm Computing	Application Classifications
Gritsenko et al. ( Gritsenko et al., 2017 )	2017	• • • • • • •	• • • • • • •	• • • • • • •	• • • • • • •	• • • • • • •
Lulu and Parhi ( Ge and Parhi, 2020 )	2020	• • • • • • •	• • • • • • •	• • • • • • •	• • • • • • •	• • • • • • •
Thomas et al. ( Thomas et al., 2021 )	2021	• • • • • • •	• • • • • • •	• • • • • • •	• • • • • • •	• • • • • • •
Kleyko et al. (part 1) ( Kleyko et al., 2022 )	2022	• • • • • • •	• • • • • • •	• • • • • • •	• • • • • • •	• • • • • • •
Kleyko et al. (part 2) ( Kleyko et al., 2023 )	2022	• • • • • • •	• • • • • • •	• • • • • • •	• • • • • • •	• • • • • • •
Hassan et al. ( Hassan et al., 2022 )	2022	• • • • • • •	• • • • • • •	• • • • • • •	• • • • • • •	• • • • • • •

Table 3. Comparison of Previous SOTA Surveys for Key Contributions of This Survey

Figure 4. An overview of the comprehensive multi-stage encoding pipeline taxonomy.

$\boldsymbol{\mathcal{HV}}$ Source	Dimensionality of $\boldsymbol{\mathcal{HV}}$ s ( $\mathbf{D}$ )	Run-to-run reproducibility	Executions for validation	Security-critical deployment	Hardware generator complexity	High- $\mathbf{D}$ scaling overhead	Current validation gaps
Pseudo-Random	Effective at sufficiently large $D$ ; near-orthogonality holds with high probability	Reproducible with fixed seed; seed-dependent otherwise	Usually needs multiple seeds for statistical confidence	Suitable if seeds are protected and entropy is sufficient	Low–Medium: LFSR+comparator	Usually modest; depends on parallelism and generator period	Standard baseline, but seed- and workload-dependent
True Random	Effective at large $D$ ; task- and technology-dependent	Weak unless stored or replayed	Requires repeated trials or stored realizations	Unpredictable; attractive for security-sensitive settings	Depends on entropy source or emerging device	Depends on generation, storage, and readout cost	Limited HDC-specific validation across workloads and hardware
LD Sequences	Reported benefits mainly at moderate $D$ ; task-dependent	Deterministic and reproducible by construction	May reduce repeated seed trials; validation remains task-dependent	Potentially vulnerable if the generator is known	Medium–High: LUT/counter or deterministic logic	Implementation-dependent; LUT/counter overhead may grow with $D$	Broader validation needed across tasks, datasets, dimensions, and hardware

Table 4. Qualitative Tradeoffs of LD Sequences vs. Pseudo-Random and True-Random Initializations

Figure 5. Text and image processing pipelines in HDC. (a) Text Processing: Letters are converted into orthogonal symbol vectors via a pseudo-random source and a comparator, followed by $N$ -gram encoding: shifting-and-rotating , XOR , and accumulate operations. (b) Image Processing: Pixel values are processed using intensity ( $L$ , level) and symbol-based position ( $P$ ) vector encoding.

Encoding	Accuracy vs. Dimensionality	Hardware Cost	Best-Suited Domain
♣ Record-based ( Imani et al., 2018a ; Datta et al., 2019 ; Rahimi et al., 2016a ; Ge and Parhi, 2020 )	Medium	Low	Multivariate sensor streams
♣ N-gram ( Rahimi et al., 2016b ; Ge and Parhi, 2020 ; Datta et al., 2019 )	High (text)	Medium	Language and other symbol sequences
◆ RBF (cosine) ( Hernández-Cano et al., 2021b ; Zou et al., 2021a ; Wang et al., 2023c ; Zou et al., 2021b )	High (continuous)	High	Regression or continuous-valued data
◆ Random projection ( Cannings and Samworth, 2015 ; Hersche et al., 2018 ; Imani et al., 2019f ; Morris et al., 2022b )	Medium	Medium–High	General-purpose vision tasks
▲ Cellular automaton ( Menon et al., 2022a ; Neubert et al., 2021 )	Medium	Very Low	Spatio-temporal signal processing
◆ Sigmoidal activation function ( Hernández-Cano et al., 2021b ; Zou et al., 2021a )	Medium	Medium	Analog/continuous sensor data

Table 5. Selection Guideline on HDC Encoding Methods with Hardware Cost and Taxonomy Symbols

Representative Studies	Dominant Encoding Stage
	◼ Feature	◆ Mapping	♣ Composition	▲ $\boldsymbol{\mathcal{HV}}$ Source
VoiceHD ( Imani et al., 2017b ) , AdaptHD ( Imani et al., 2019e ) , PRID ( Hernández-Cano et al., 2021a ) , MIMHD ( Kazemi et al., 2021 ) , Dynamic-HDC ( Chuang et al., 2023 ) , HyperGRAF ( Chen et al., 2023 ) , SemiHD ( Imani et al., 2019b ) , BinHD ( Imani et al., 2019d ) , QubitHD ( Bosch et al., 2019 ) , QuantHD ( Imani et al., 2020a ) , HDC-IM ( Liu et al., 2019 ) , EnHDC ( Wang et al., 2023d ) , HDCluster ( Imani et al., 2019c ) , TP-HDC ( Chang et al., 2020 ) , LeHDC ( Duan et al., 2022 ) , XCelHD ( Kang et al., 2022a ) , MulTa-HDC ( Chang et al., 2021 ) , HyperRec ( Guo et al., 2021 ) , ScaleHD ( Zhang et al., 2022 ) , HyperSpec ( Xu et al., 2023a ) , Imani et al. ( Kim et al., 2018 ) , Nazemi et al. ( Nazemi et al., 2020 ) , Montone et al. ( Montone et al., 2017 )	$\ast$	$\ast$	C1, C2	P. Random
ST-HDC ( Karunaratne et al., 2021a ) , RelHD ( Kang et al., 2022c ) , LookHD ( Imani et al., 2021 ) , SHDC ( Huch et al., 2023 ) , DOMINO ( Wang et al., 2023b ) , GENERIC ( Khaleghi et al., 2022 ) , BioHD ( Zou et al., 2022c ) , GrapHD ( Poduval et al., 2022 ) , HAM ( Imani et al., 2017c ) , PartialHD ( Taheri et al., 2023 ) , Rahimi et al. ( Rahimi et al., 2016b ) , Datta et al. ( Datta et al., 2019 ) , Li et al. ( Li and others, 2016 ) , Wu et al. ( Wu et al., 2018b ) , Karunaratne et al. ( Karunaratne et al., 2020a ) , Rahimi et al. ( Rahimi et al., 2017 ; Rahimi et al., 2020 ) , Kovalev et al. ( Kovalev et al., 2022 )	$\ast$	$\ast$	C1, C2, C3	P. Random
FebHD ( Hernández-Cano et al., 2021b ) , DistHD ( Wang et al., 2023c ) , DUAL ( Imani et al., 2020 ) , TempHD ( Ni et al., 2022 ) , NeuralHD ( Zou et al., 2021a ) , RE-FHDC ( Zeulin et al., 2023 ) , ManiHD ( Zou et al., 2021b ) , HD-I-IoT ( Gungor et al., 2023b ) , CascadeHD ( Kim et al., 2021b ) , SupportHD ( Safa et al., 2023 ) , CyberHD ( Wang et al., 2023a ) , Yu et al. ( Yu et al., 2022 )	$\ast$	M1	$\ast$	Gaussian
SparseHD ( Imani et al., 2019h ) , F5-HD ( Salamat et al., 2019 ) , compHD ( Morris et al., 2019 ) , HD-Core ( Salamat et al., 2020 ) , HoloGN ( Kleyko et al., 2017 ) , Kleyko et al. ( Kleyko et al., 2018a ; Kleyko et al., 2018b )	$\ast$	$\ast$	C2, C3	P. Random
SecureHD ( Imani et al., 2019a ) , OnlineHD ( Hernández-Cano et al., 2021c ) , FACH ( Imani et al., 2019g ) , SHEARer ( Khaleghi et al., 2020b ) , SearcHD ( Imani et al., 2020b )	$\ast$	$\ast$	C1, C2	Gaussian

Table 6. Representative Studies Categorized by the Multi-Stage HDC Encoding Taxonomy

HDC Projects	Relevant Taxonomy Stages	Encoding-Relevant Details ( $\boldsymbol{\mathcal{HV}}$ source, mapping, composition, or encoder realization)
TorchHD ( Chandrasekaran and Heddes, 2022 ; Heddes and others, 2023 )	◼ ▲ ◆ ♣	General HDC framework supporting random, level, thermometer, circular, RBF-kernel, N-gram-based , and Record-based encoding.
DistHD ( Wang et al., 2023c )	▲ ◆	Gaussian $\boldsymbol{\mathcal{HV}}$ source and RBF-style mapping for distance-preserving continuous-feature encoding.
OpenHD ( Kang and Kim, 2022 )	◆ ♣ ✸	GPU-oriented framework support for mapping, composition, and encoder execution.
HDTorch ( Simon et al., 2022 )	▲ ◆	Random, sandwich, and scale-with-radius encoding for HDC model construction.
Hyperdimensional Computing Library ( , 2019 )	▲ ♣ ✸	Random $\boldsymbol{\mathcal{HV}}$ source and N-gram-based composition with CPU/GPU-oriented encoder support.
Laelaps ( Burrello et al., 2019 )	⚫ ◼ ♣ ✸	LBP-based feature extraction with composition-oriented encoder realization ( Kaya et al., 2014 ) .

Table 7. Representative HDC Frameworks and Implementations Mapped to the Encoding Taxonomy

Year, Authors	Contribution	Application, Dataset (Accuracy, D)	Encoding	Environment , Model , or Hardware Details	Hardware Efficiency
2015, Räsänen ★ ( Räsänen, 2015 )	Vector generation for non-discrete data	Spoken word classification CAREGIVER Y2 UK corpus ( $>97$ %, $D$ =4,000)	S-WARP Mapping and Weighted Accumulation of Random Projections ◼ ◆	No information given	No hardware implementation
2018, Wu et al. ❆ ( Wu et al., 2018a )	Exploring the opportunities of HDC for next generation devices (nanotubes, RRAM)	21 European languages classification Europarl (98%)	Evenly spaced time-delay encoding ◆ ♣ ✸	Synopsys Design Compiler CNFET and RRAM-based design	# of CNFET (1952) # of RRAM (224)
2018, Imani et al. ♠ ( Imani et al., 2018a )	Decreasing classification cost	Speech recognition ISOLET (95.9%, $D$ =10,000)	Random-Bipolar $\boldsymbol{\mathcal{HV}}$ Bit-flipping applied Hierarchical multi-encoder: I. Record-based II. N-gram-based ▲ ♣ ✸	Synopsys Design Compiler (TSMC 45nm)	Energy (6.6×) Speed (6.3×)
2019, Chang et al. ❃ ( Chang et al., 2019 )	$\boldsymbol{\mathcal{HV}}$ embedding using random non-linearity	Galvanic Skin Response, Electroencephalogram, Electrocardiogram dataset (AMIGOS) binary emotion classification (76.6%, $D$ =10,000)	Multiplexer-based ternary (-1, 0, +1) feature mapping into hyperspace ◼ ◆	No information given	Only system block diagram is given
2019 Liu et al. ✥ ( Liu et al., 2019 )	RRAM-based HDC in-memory architecture	ISOLET (92.1%, $D$ =8,192)	Block generation of $\boldsymbol{\mathcal{HV}}$ s on crossbar, XOR -summation encoding ▲ ♣ ✸	RRAM, CPU, and GPU separately	2.28 J Energy (in-memory) 3960 J Energy (CPU)
2020, Nazemi et al. ❣ ( Nazemi et al., 2020 )	Automatic feature extraction	HAR (96.44%, $D$ =10,240) ISOLET (96.67%, $D$ =10,240)	Random seed $\boldsymbol{\mathcal{HV}}$ Bit-flipping Binding and Bundling ◼ ▲ ♣ ✸	FPGA (Xilinx UltraScale+ VU9P)	BRAMs (1.8%) DSPs (15%) FFs (0.8%) LUTs (5.1%) Latency (23.3 us) Power (5.3 W)

Table 8. Summary of Noticeable SOTA Works Highlighting Applications,

\boldsymbol{\mathcal{HV}}

Mapping, and Hardware

Term	Description
Associative Memory (AM)	A memory for storing learned class hypervectors used for similarity-based classification.
Binary Spatter Coding	A foundational VSA model using high-dimensional dense binary vectors; binding is performed via XOR and bundling via majority rule.
Binarization	Transforms real-valued bundles into bipolar form (e.g., sign() to get $\pm 1$ ) or into binary form (e.g., majority thresholding to get $0$ or $1$ ).
Binding	Combines hypervectors (e.g., via XOR, circular convolution) to represent relationships or associations.
Bit-Flipping Encoding	Generates hypervectors for neighboring scalar values by flipping a small number of bits.
Bundling	Aggregates multiple hypervectors into one, typically by majority vote or summation + sign thresholding.

Table 9. Glossary of Key Terms in Hyperdimensional Computing (HDC)

Hyperdimensional Computing

Hyperdimensional computing uses high-dimensional vector representations and algebraic operations for efficient learning, memory, and reasoning.

See all 39 papers

Superposed Inference for Hyperdimensional Computing
Sep 26, 2026Quanling Zhao, Nilesh Prasad Pandey, Ye Tian +1Hyperdimensional Computing Efficient Inference
University of California San Diego
Bridging the Gap Between Hyperdimensional Computing and Kernel Methods via the Nyström Method
Aug 7, 2026Quanling Zhao, Anthony Hitchcock Thomas, Ari Brin +2Nyström Method Hyperdimensional Computing
Department of Computer Science and Engineering, UC San Diego
Large-Margin Hyperdimensional Computing: A Learning-Theoretical Perspective
Mar 4, 2026Nikita Zeulin, Olga Galinina, Ravikumar Balakrishnan +2Hyperdimensional Computing Energy-Efficient ML
Tampere University, Tampere, Finland · Tampere Institute for Advanced Study, Tampere University, Finland · Intel Corporation, Santa Clara, CA, United States

Sep 26, 2026J/K move · Enter open · S save

Superposed Inference for Hyperdimensional Computing

Quanling Zhao, Nilesh Prasad Pandey, Ye Tian, Tajana Rosing

University of California San Diego

Hyperdimensional Computing Efficient Inference

Open paper AskPDF arXiv

Hyperdimensional computing (HDC) is attractive for efficient and robust learning, but conventional inference still encodes every query independently, repeatedly paying the cost of high-dimensional projection. We introduce SupHDC, a new inference paradigm that processes multiple queries through a shared encoding computation. SupHDC assigns lightweight random slot keys, superposes the keyed queries before encoding, and uses slot-specific classifiers to recover their individual predictions. A random-feature kernel view explains why exact recovery of each hypervector is unnecessary: inference only needs to preserve the class evidence that determines the prediction. Across ten datasets, SupHDC achieves 1.39x analytical speedup with no average accuracy loss, and up to 2.08x speedup with only a 2.67 percentage-point mean accuracy loss. On a Raspberry Pi~5, it delivers 2.01x measured wall-clock speedup with a 2.26 percentage-point loss in mean prediction accuracy. SupHDC shows that high-dimensional redundancy can be used not only for robustness, but also as capacity for shared inference.

Abstract

Hyperdimensional Computing

Explore similar work

Superposed Inference for Hyperdimensional Computing

Bridging the Gap Between Hyperdimensional Computing and Kernel Methods via the Nyström Method

Large-Margin Hyperdimensional Computing: A Learning-Theoretical Perspective