WorksComputer Science
Vector-Space Projection and Unified Complex-Valued Acceleration for Scaling Matrix-Based Brain-Inspired Representations
Abstract
Symbolic reasoning at scale requires representations that are both expressive and hardware-efficient. Hyperdimensional Computing (HDC) provides a promising brain-inspired framework, encoding concepts as high-dimensional vectors composed through efficient algebraic operations. However, the widely used Fourier Holographic Reduced Representation (FHRR) relies on commutative binding, which cannot encode order or directionality. Prior FPGA/ASIC accelerators recover this through costly permutation logic, incurring irregular memory access and high control overhead. Generalized Holographic Reduced Representations (GHRR) instead redefine binding as non-commutative matrix multiplication, naturally encoding sequences, graphs, and hierarchies without permutations-but at the cost of turning every element into a matrix, introducing compute and storage overhead that limits scalability. In this work, we propose a flattening methodology that preserves GHRR's matrix-based encoding while executing training and inference directly in vector space, functionally equivalent to FHRR inference yet free of permutation logic. We further design a custom 28 nm ASIC that fuses binding and similarity into a unified complex-valued datapath, with dual-DMA streaming and runtime normalization for accurate inference. Against a PyTorch baseline on an NVIDIA RTX 4090 GPU, the prototype delivers 1.36x-1.56x higher throughput and 16.2x-18.6x better energy efficiency.
CCS Concepts • Computing methodologies → Knowledge representation and reasoning; • Computer systems organization → Neural networks; • Hardware → Application specific integrated circuits.
ACM Reference Format: William Youngwoo Chung, Hyunwoo Oh, Calvin Yeung, Hansen Jin Lillemark, Hamza Errahmouni Barkam, and Mohsen Imani. 2026. Vector-Space Projection and Unified Complex-Valued Acceleration for Scaling Matrix-Based Brain-Inspired Representations. In ACM/IEEE International Symposium on Low Power Electronics and Design (ISLPED ’26), August 05–07, 2026, Evanston, IL, USA. ACM, New York, NY, USA, 7 pages. https://doi.org/10.1145/3816440.3818598
Introduction
Designing architectures for symbolic reasoning requires representations to be expressive enough to capture complex structures, while also being efficient enough to meet the power and latency budgets of modern accelerators. Machine learning models have seen excellent results across various domains but often entangle objects and attributes into dense activations, limiting their transparency and compositional reasoning ability. For deployment in neuromorphic and embedded systems, an alternative framework is needed, one that preserves symbolic structure while working within hardware limitations.
Hyperdimensional Computing (HDC), or Vector Symbolic Architectures (VSA), offers such a framework by encoding concepts as high-dimensional vectors and manipulating them through algebraic operations such as binding, unbinding, and superposition [11]. These operations are inherently parallel and noise-tolerant, making them attractive for hardware acceleration. Yet the most widely used implementation, Fourier Holographic Reduced Representation (FHRR) [26], relies on commutative based binding, which cannot capture order and directionality. Sequences and graphs must therefore be encoded through additional permutation logic. On hardware, this requirement results in costly, irregular memory access, which can undermine the efficiency of prior FPGA and ASIC accelerators [3, 9].
This work is licensed under a Creative Commons Attribution 4.0 International License. ISLPED ’26, Evanston, IL, USA
© 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2748-1/2026/08 https://doi.org/10.1145/3816440.3818598 Generalized Holographic Reduced Representation (GHRR) [34] overcomes this bottleneck by redefining binding as matrix multiplication in . Unlike commutative element-wise products, matrix multiplication is non-commutative such that embedding structure and order can be directly done in the representation space. Beyond its representational benefits, GHRR’s core operations align naturally with GEMM-like operations, a primitive already optimized across accelerator architectures. By collapsing compositional reasoning into GEMM-style datapaths and non-commutative binding, GHRR eliminates the need for permutation-specific hardware and enables symbolic reasoning to scale on existing high-throughput pipelines.
This work introduces a permutation-free accelerator design that natively supports GHRR. Rather than positional embeddings or permutations, we reframe the encoding symbolic reasoning as a series of matrix multiplications, allowing binding, unbinding, and similarity to execute directly on a unified complex-valued datapath. Our 28 nm ASIC prototype integrates runtime normalization, interleaved real–imaginary buffers, and dual DMA streaming to sustain high throughput, achieving 1.54 TFLOPS peak performance within 4.77 mm and 4.22 W. Compared to a PyTorch baseline on an NVIDIA RTX 4090 GPU, the design delivers 1.36–1.56× higher throughput and 16.2–18.6× energy efficiency.
By rethinking GHRR through a hardware lens, we demonstrate that the path to efficient symbolic reasoning can be achieved through additional parallelism or memory optimization, and eliminating permutation altogether. This permutation-free methodology simplifies accelerator design, scales predictably with dimensionality and parallelism, and provides a practical foundation for embedding brain-inspired reasoning into next-generation machine learning hardware.
In this work, we present a unified algorithm–architecture framework for scalable, permutation-free GHRR acceleration. First, we show that GHRR similarity can be computed exactly in a flattened complex vector space, enabling efficient training and inference without explicit matrix-valued computation. Second, we design a streaming ASIC accelerator that maps binding, unbinding, and similarity onto a unified complex-valued datapath. Finally, we demonstrate that the proposed design achieves FHRR-level inference efficiency while preserving GHRR’s non-commutative representational advantages.
Background & Related Works
Hyperdimensional Computing for Structured Data
Hyperdimensional Computing (HDC) represents symbolic entities as high-dimensional vectors and manipulates them through algebraic operations such as bundling (superposition) and binding (association) [8, 11, 28, 32, 37, 38, 39, 40]. These operations are massively parallel, noise-tolerant, and hardware-friendly, enabling applications in domains such as edge intelligence, robotics, and reinforcement learning [1, 10, 22, 23, 31].
Holographic Reduced Representations (HRR) [26] are one of the most established forms of HDC, defining binding as circular convolution, bundling as addition, and similarity as dot product between high-dimensional vectors. The Fourier variant (FHRR) [27] replaces convolution with element-wise complex multiplication, which allows exact unbinding through conjugation and further simplifies hardware realization. Both HRR and FHRR preserve distributed and robust encodings and form the foundation of many neurosymbolic HDC-based frameworks. However, because element-wise multiplication is commutative, these encodings cannot naturally capture order or directionality. To represent sequences or directed graphs, additional permutation operations are typically employed [5], which complicate interpretability, hardware mapping, and information retrieval [17, 18, 30].
More recently, graph-centric HDC methods have pursued graph-based symbolic reasoning. MoleHD [21] embeds molecular graphs into hypervectors at low energy cost, GraphHD [28] encodes entire graph structures as single hypervectors, and HDReason [2] and RelHD [12] incorporate relational and multi-hop information; recent work has also examined reasoning quality and bias in graph HDC [19, 20]. Despite their success, these methods remain reliant on commutative binding operators and thus fall back on permutations or handcrafted role vectors to encode structural information. This reliance introduces additional hardware overheads and limits compositional reasoning efficiency.
Generalized Holographic Reduced Representations
To address the problem of HDC encoding not naturally capturing order [6, 7, 34], various recent solutions have been proposed. GHRR address the commutativity limitation by extending FHRR into the matrix domain. Instead of unitary complex scalars , GHRR employs unitary matrices , yielding hypervectors represented as tensors , where is defined as the complexity. Binding is defined as matrix multiplication and bundling as addition, with similarity computed via normalized Frobenius inner products. When , GHRR reduces exactly to FHRR, making it a natural generalization. This formulation retains the robustness and distributed properties of HDC while enabling non-commutative binding and richer representational capacity. It also broadens the kernel approximation view of FHRR, which is closely related to Random Fourier Features (RFF) [29]. Through Bochner’s theorem, FHRR encodings approximate shift-invariant kernels such as the RBF kernel when Gaussian sampling is used [33]. GHRR inherits this kernel interpretation but extends it by supporting matrix-valued encodings that can realize a wider family of kernels.
Despite its theoretical advantages, empirical demonstrations of GHRR have so far been limited to small-scale tests of non-commutativity and decoding accuracy. No prior work has systematically benchmarked its performance on large-scale workloads, nor have its hardware implications been studied. Our work fills this gap by presenting the first hardware software acceleration co-design of GHRR by introducing a flattening technique alongside a custom accelerator design that maintains hardware efficiency while exploiting GHRR’s non-commutative algebra.
Hardware Acceleration for HDC
Research on hardware acceleration of HDC has spanned FPGA, PIM, and ASIC platforms [13, 14]. Early efforts targeted lightweight, low-power classification tasks by simplifying HDC algebra into binary or real-valued operations for embedded and edge devices [9, 15, 16, 24, 25]. These designs leveraged the inherent parallelism of bundling and similarity search but offered limited support for structured data. A complementary line of work instead targets the storage cost of HDC inference itself, compressing class representations under extreme memory budgets through decomposed or logarithmically reduced codebooks [35, 36], an axis of scalability orthogonal to the binding operator targeted in this work.
Figure 1. FHRR relies on permutations to encode structure while GHRR can encode direction through non-commutative binding. As a result, GHRR’s representations are better fit for encoding complex structures whereas FHRR requires hand-engineered permutation operations.
More recent work has focused on graph reasoning with HDC. HyperGRAF [3] introduced a domain-specific accelerator for graph-based HDC, demonstrating significant efficiency gains but relying on commutative binding and additional control for role encodings. RelHD [12] and HDReason [2] propose memory-centric designs that embed role binding directly into FeFET arrays or systolic fabrics, achieving scalability for undirected graphs. However, these architectures remain tied to commutative operators, which forces the use of handcrafted role vectors or positional permutations to represent directionality. As a result, this reliance inflates area, latency, and design complexity. Hardware accelerators for HDC frequently optimize only the bundling stages (e.g., Hadamard transforms), while overlooking the core binding operator that governs structure-preserving composition.
In contrast, our approach eliminates permutation entirely by aligning GHRR’s non-commutative binding with dense matrix operations and flattening its matrix space into a vector space better supported by HDC hardware. This matters because permutations, though trivial in software, are a major hardware bottleneck: a permutation moves all elements of a hypervector across the memory hierarchy with negligible computation, yielding low arithmetic intensity. On GPUs this disrupts access and tensor-core utilization; on ASICs and FPGAs it forces general-purpose shuffle or crossbar networks [3] that consume routing resources and add latency. This ties directly to the memory wall, where DRAM/HBM traffic dominates accelerator energy and execution time [4]. By replacing permutations with non-commutative binding, our design reduces off-chip traffic and yields a compute–memory balance better suited to modern accelerators, enabling hardware-friendly symbolic reasoning that scales to sequential and relational tasks while retaining the throughput expected of HDC accelerators.
GHRR Mathematics
A GHRR hypervector is composed of unitary matrix blocks of size :
where each is drawn from a distribution over unitary matrices. Here, is the dimensionality of the hypervector itself whereas is defined as the complexity and size of the matrix.
Bundling (superposition) in GHRR is defined as element-wise addition:
whereas binding (association) is defined as element-wise matrix multiplication:
which is non-commutative and therefore captures directionality between associations. Similarity (retrieval) uses the Frobenius inner product as the similarity metric:
where refers to taking the real component and is the trace of the resulting matrix averaged across .
Flattening Equivalence
Matrix-valued representations increase expressivity but introduce computational overhead during training and inference. The key observation enabling our hardware design is that GHRR similarity depends only on the Frobenius inner products of the blocks, which admit an exact vector form:
Therefore, for any ,
Applying (6) blockwise to and gives the following:
where
concatenates the vectorized blocks. Substituting into (4) gives:
Since the two inner products are complex conjugates,
Taking real part of (10) yields
Therefore, the normalized similarity in the matrix space equals a normalized real part of an inner product in the flattened vector space. This justifies training and inference directly on with standard high-throughput dot-product engines, without losing the GHRR semantics.
Utilizing GHRR Encoding
Now that we have established that GHRR can be flattened for inference, we will now explain GHRR encoding works. Given a data point , we construct its GHRR representation blockwise as
where each is a unitary matrix (either fixed or varying across ). The diagonal modulation is defined by
with vectors, , drawn from a distribution . For two inputs , the similarity yields an approximation to a kernel .
After the GHRR encoding, we use in any standard classifier or HDC routine. Bundling, updates, and similarity operate identically in the flattened domain because of (12). This results in an effective dimension of , preserving GHRR’s non-commutative binding association while enabling learning on an efficient vector space for hardware acceleration.
Hardware Acceleration
Now that GHRR can be mapped into a vector space for inference, we co-design a hardware accelerator tailored to its core operations: binding-based encoding and similarity-based inference, making GHRR viable for deployment in low-latency or resource-constrained environments. Unlike traditional HDC systems that rely on commutative binding and permutation logic, our ASIC design natively supports GHRR’s matrix-based, non-commutative algebra while maintaining end-to-end differentiability and throughput scalability. The overall system architecture is shown in Figure 2. The accelerator consists of three main subsystems: the Encoder Block, which transforms input features into hyperdimensional GHRR representations; the Inference Block, which evaluates normalized similarity against a learned codebook; and a control/data interface managed by a lightweight Host Interface. Communication with off-chip memory is handled by a dual-channel DMA Engine, which streams both raw inputs and codebook vectors from external memory into the on-chip buffers.
Figure 2. Top-level architecture of the GHRR accelerator. Host interface coordinates encoder and inference blocks via DMA streaming from external memory.
Encoder Block
The encoder module translates input data into GHRR hypervectors using a complex-valued transformation pipeline. As shown in Figure 3, the core of this pipeline is a Binding Unit that multiplies pairs of complex-valued matrices element-wise using the standard rule for complex multiplication:
Figure 3. Encoder acceleration block
This multiplication is implemented via parallel fused-multiply-add (FMA) units, with real and imaginary parts stored in dual buffers of shape , where is the hyperdimension. Each input vector is projected through a bank of fixed or randomized unitary transformations , followed by an element-wise complex exponential controlled by . These steps are pipelined for latency hiding and throughput maximization, with the resulting encoding streamed directly into downstream inference buffers.
Inference Block
Inference proceeds by computing the similarity between the current encoding and a fixed hyperdimensional codebook. In hardware, this is realized by a fully parallel Similarity Function Unit (Figure 4(a)) that compares candidate vectors in parallel. Each comparison computes a normalized dot product:
Figure 4. Inference acceleration block
Here, and are interleaved real-valued encodings of the original complex hypervectors. The system avoids complex conjugation for efficiency and instead directly evaluates the real-valued similarity approximation. This approximation maps cleanly to vector dot product units and eliminates the need for complex arithmetic.
The norm of each input vector is computed at runtime using accumulation (), and their multiplicative inverses and are generated in hardware. These values scale the final dot product through two multiplication stages. The result of each unit is forwarded to a lightweight top-1 engine, which identifies the index of the most similar codebook vector across all candidates in the current batch.
This runtime normalization is essential: unlike Hermitian inner products, the dropped conjugation means that is not fixed to the dimensionality , and may vary significantly based on the input distribution. Without normalization, similarity scores would be biased by magnitude differences, impairing robustness and accuracy.
To support high-throughput inference, the reciprocal square root is computed on-chip using a pipelined logic based on the Newton-Raphson method. This value is then refined across multiple stages of the pipeline to produce the final , which is used to scale the accumulated dot product.
The inference pipeline overlaps input loading, dot product evaluation, norm accumulation, reciprocal square root computation, and top-1 selection with minimal control overhead. This architecture maintains high utilization even as batch size and dimensionality increase.
Throughput and Parallelism
Both the encoder and inference blocks are designed to support data-level parallelism. The pipeline architecture overlaps data fetch, binding, and accumulation phases, while a streaming input controller ensures high utilization across all compute units. Our accelerator supports configurable parallelism degree , enabling adaptation to varying batch sizes and real-time throughput requirements without sacrificing accuracy.
Advantages Over Permutation-Based Designs
By eliminating the need for positional permutations or synthetic encodings, our hardware design simplifies memory access patterns and reduces logic area. All compositional reasoning steps, binding, unbinding, and comparison, collapse into GEMM-like operations with minimal control overhead. As a result, GHRR’s algebra maps cleanly onto both spatial (e.g., systolic) and temporal (e.g., pipelined) architectures. The matrix-based encoding allows us to recover order and hierarchy directly from the algebra, avoiding the combinatorial explosion of permutation states typical in traditional FHRR accelerators.
Results
Inference Experiments
To validate our flattening method, we benchmark FHRR, GHRR (Flatten), and GHRR (Matrix) on a synthetic sequence classification dataset with samples of sequence of length = 5 and 10 possible classes. The goal is to measure the inference cost of a given naive GHRR implementation when compared to FHRR and GHRR (Flatten). We run every model on a NVIDIA RTX 3060Ti GPU for 50 trials and report the average inference times.
We compare the inference speeds between all three implementations to gauge how efficient flattening after encoding is relative to FHRR and a normal GHRR implementation as well as the effect of hyperdimensional on the inference time in Figure 5. In our experiments we note that GHRR (Matrix) takes exponentially longer during inference whereas GHRR (Flatten) remains equivalent speeds as FHRR. Across all dimensions, GHRR (Flatten) remains within 0.95× - 1.05× the inference time as FHRR. On the other hand, GHRR’s (Matrix) inference times relative to FHRR is 38.2× at , 16.8× at , 15.5× at , and 12.3× at , showing an extreme cost of keeping GHRR representations as matrices. However, we do note that the gap between the inference times does decrease as increases. Figure 5 confirms our motivation to accelerate GHRR through vector space flattening for a scalable non-commutative-based HDC learning frameworks.
Figure 5. Inference cost of FHRR, GHRR Flatten, and GHRR Matrix on a synthetic sequence based dataset.
Hardware Acceleration
In this section, we implement the custom GHRR accelerator as a register-transfer level (RTL) design in Chisel, enabling modular and flexible architectural configurations. Functional verification is conducted using Verilator 5.009, and the design is synthesized using Synopsys Design Compiler targeting the 28 nm CMOS process, meeting the timing constraints at 1.0 GHz.
Performance and Efficiency
To assess the practicality of our architecture on reasonable HDC workloads, we synthesize the full GHRR pipeline in Figure 2 and evaluate area, power, and throughput. Experiments use a reasonable default hardware configuration of hyperdimension , complexity , and parallelism , aligned with the benchmark workloads in this section. These settings reflect the target deployment regime of the accelerator and enable direct comparison against state-of-the-art GPU implementation.
Table 1 summarizes the resulting specifications. The complete system occupies 4.77 mm of silicon area, operates at 1.00 GHz, and achieves a peak throughput of 1.54 TFLOPS, with a maximum power draw of 4.22 W.
| Parameter | Total Area | Max Power | Frequency | Peak Throughput |
| Value | 4.77 mm | 4.22 W | 1.00 GHz | 1.54 TFLOPS |
Table 1. Hardware Specifications
We benchmark our ASIC design against an optimized software baseline running on an NVIDIA RTX 4090 GPU. As shown in Figure 6, the accelerator achieves up to 1.56× higher throughput for encoding and 1.36× for inference. The energy efficiency gains are even more substantial, with 18.6× and 16.2× reductions in energy per operation, respectively.
Figure 6. Comparison of throughput and energy efficiency against an NVIDIA RTX 4090 baseline.
These improvements arise from architectural co-design that fuses complex-valued HDC primitives into a single streamlined datapath, eliminating permutation-heavy shuffles common in prior work. With on-the-fly normalization and a fully pipelined binding–similarity flow, the design sustains high utilization across and batch sizes, translating algorithmic regularity directly into throughput and energy gains.
Large Dimensionality Ablation
To thoroughly test the scalability of our proposed accelerator, we additionally benchmark our ASIC across hypervector dimensions as a definitive stress test across the scalability across large dimensionalities. Again, the accelerator consistently outperforms the GPU: encoder throughput improves by 1.16×–3.66× (peaking at =64K), while inference throughput improves by 1.36×–17.51× (peaking at =32K). Energy per operation drops by 13.83×–43.7× for encoding and 16.19×–121.87× for inference, with the largest gains occurring at mid-range .
Figure 7 summarizes these trends across various dimensionalities for a complete stress test on any possible complex tasks that need large representational capacity.
Figure 7. Ablation study comparison of throughput and energy efficiency against an NVIDIA RTX 4090 baseline across large hyperdimensions.
Conclusion
We present the first hardware software co-design accelerated pipeline for GHRR to unify binding, unbinding, and inference, removing the need for expensive permutation logic as well as reducing GHRR’s expensive matrix representations to simpler traditional vector representations for inference for exponential reduction during inference time. Additionally, our accelerator achieves over 1.36×–1.56× higher throughput and achieves 16.2×–18.6× better energy efficiency when compared to an NVIDIA RTX 4090 GPU, demonstrating compatibility with custom accelerators. Through this work, we demonstrate that GHRR can be practically deployed as a scalable alternative to traditional HDC architectures through flattening for inference and provides a promising foundation for future, energy-efficient brain-inspired computing architectures on custom or emerging hardware.
Acknowledgment
This work was supported in part by the DARPA Young Faculty Award, the National Science Foundation (NSF) under Grants #2127780, #2319198, #2321840, #2312517, and #2235472, the Semiconductor Research Corporation (SRC), the Office of Naval Research through the Young Investigator Program Award, and Grants #N00014-21-1-2225 and #N00014-24-1-2547, Army Research Office Grant #W911NF2410360. Additionally, support was provided by the Air Force Office of Scientific Research under Award #FA9550-22-1-0253.
References
References
- [1]Rajat Bhattacharya, Woohyeok Park, Arnab Sarkar, Hyunwoo Oh, Mohsen Imani, and Nikil Dutt. 2026. HYPERDOA: Robust and Efficient DoA Estimation Using Hyperdimensional Computing. In ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 20841–20845.
- [2]Hanning Chen, Yang Ni, Ali Zakeri, Zhuowen Zou, Sanggeon Yun, Fei Wen, Behnam Khaleghi, Narayan Srinivasa, Hugo Latapie, and Mohsen Imani. 2024. HDreason: Algorithm-hardware codesign for hyperdimensional knowledge graph reasoning. arXiv preprint arXiv:2403.05763 (2024).arxiv.org/abs/2403.05763
- [3]Hanning Chen, Ali Zakeri, Fei Wen, Hamza Errrahmouni Barkam, and Mohsen Imani. 2023. Hypergraph: Hyperdimensional graph-based reasoning acceleration on fpga. In 2023 33rd International Conference on Field-Programmable Logic and Applications (FPL). IEEE, 34–41.DOI
- [4]William Youngwoo Chung, Hamza Errrahmouni Barkam, Tamoghno Das, and Mohsen Imani. 2025. Robust Reasoning and Learning with Brain-Inspired Representations under Hardware-Induced Nonlinearities. In Proceedings of the Great Lakes Symposium on VLSI 2025. 968–975.DOI
- [5]E Paxon Frady, Denis Kleyko, and Friedrich T Sommer. 2018. A theory of sequence indexing and working memory in recurrent neural networks. Neural Computation 30, 6 (2018), 1449–1513.
- [6]Stephen I Gallant and T Wendy Okaywe. 2013. Representing objects, relations, and sequences. Neural computation 25, 8 (2013), 2038–2078.DOI
- [7]Jan Gonsmann and Chris Eliasmith. 2019. Vector-derived transformation binding: An improved binding operation for deep symbol-like processing in neural networks. Neural computation 31, 5 (2019), 849–869.
- [8]Mohsen Imani, Yeseong Kim, Behnam Khaleghi, Justin Morris, Haleh Alimohamadi, Farhad Imani, and Hugo Latapie. 2023. Hierarchical, distributed and brain-inspired learning for internet of things systems. In 2023 IEEE 43rd International Conference on Distributed Computing Systems (ICDCS). IEEE, 511–522.DOI
- [9]Mohsen Imani, Sahand Salamat, Saransh Gupta, Jiani Huang, and Tajana Rosing. 2019. Graph-based acceleration of hyperdimensional computing by reducing computational complexity. In Proceedings of the 24th Asia and South Pacific Design Automation Conference. 493–498.
- [10]Haris Imani, Sima Shahhosseini, Yang Ni, Tianyi Hu, Danny Abraham, Amir M Rahmani, Nikil Dutt, and Mohsen Imani. 2022. Hyperdimensional hybrid learning on end-edge-cloud networks. In 2022 IEEE 40th International Conference on Computer Design (ICCD). IEEE, 652–655.DOI
- [11]Pentti Kanerva. 2009. Hyperdimensional computing: An introduction to computing in distributed representation with high-dimensional random vectors. Cognitive Computation 1, 2 (2009), 139–159.
- [12]Jaeyoung Kang, Minxuan Zhou, Abhinav Bhansali, Weihong Xu, Anthony Thomas, and Tajana Rosing. 2022. Graph-based learning on fefet with hyperdimensional computing. In 2022 IEEE 40th International Conference on Computer Design (ICCD). IEEE, 553–560.DOI
- [13]Geethan Karunaratne et al. 2020. In-memory hyperdimensional computing. Nature Electronics (2020), 1–11.
- [14]Denis Kleyko, Mike Davies, Edward Paxon Frady, Pentti Kanerva, Spencer J Kent, Bruno A Olshausen, Evgeny Osipov, Jan M Rabaey, Dmitri A Rachkovskij, Abbas Rahimi, et al. 2022. Vector symbolic architectures as a computing framework for emerging hardware. Proc. IEEE 110, 10 (2022), 1538–1571.pubmed.ncbi.nlm.nih.gov/37868615
- [15]Denis Kleyko, Nikita Lyamin, Evgeny Osipov, and Laurynas Riliskis. 2012. Deepvibe MAC layer architecture based on holographic data representation using hyper-dimensional binary spatter codes. In International Workshop on Multiple Access Communications. Springer, 134–145.DOI
- [16]Denis Kleyko, Evgeny Osipov, Daswin De Silva, Urban Wiklund, and Damminda Alahakoon. 2019. Integer self-organizing maps for digital hardware. In 2019 International Joint Conference on Neural Networks (IJCNN). IEEE, 1–8.DOI
- [17]Denis Kleyko, Dmitri Rachkovskij, Evgeny Osipov, and Abbas Rahimi. 2023. A survey on hyperdimensional computing aka vector symbolic architectures, part ii: Applications, cognitive models, and challenges. Comput. Surveys 55, 9 (2023), 1–52.DOI
- [18]Denis Kleyko, Dmitri A Rachkovskij, Evgeny Osipov, and Abbas Rahimi. 2022. A survey on hyperdimensional computing aka vector symbolic architectures, part i: Models and data transformations. Comput. Surveys 55, 6 (2022), 1–40.DOI
- [19]Yezi Liu, William Youngwoo Chung, Hanning Chen, Calvin Yeung, and Mohsen Imani. 2025. Are Hypervectors Enough? Single-Call LLM Reasoning over Knowledge Graphs. arXiv preprint arXiv:2512.09369 (2025).
- [20]Yezi Liu, William Youngwoo Chung, Yang Ni, Hanning Chen, and Mohsen Imani. 2025. Mitigating Bias in Graph Hyperdimensional Computing. arXiv preprint arXiv:2512.07433 (2025).arxiv.org/abs/2512.07433
- [21]Dongning Ma, Rahul Thapa, and Xun Jiao. 2021. MoleHD: Ultra-low-cost drug discovery using hyperdimensional computing. arXiv preprint arXiv:2106.02894 (2021).
- [22]Peer Neubert, Stefan Schubert, and Peter Protzel. 2019. An introduction to hyperdimensional computing for robotics. KI-Künstliche Intelligenz 33, 4 (2019), 319–330.DOI
- [23]Yang Ni, William Y Chung, Samuel Cho, Zhuowen Zou, and Mohsen Imani. 2024. Efficient exploration in edge-friendly hyperdimensional reinforcement learning. In Proceedings of the Great Lakes Symposium on VLSI 2024. 111–118.DOI
- [24]Yang Ni, Yeseong Kim, Tajana Rosing, and Mohsen Imani. 2022. Algorithm-hardware co-design for efficient brain-inspired hyperdimensional learning on edge. In 2022 Design, Automation & Test in Europe Conference & Exhibition (DATE). IEEE, 292–297.DOI
- [25]Hyunwoo Oh, SungHeon Jeong, Suyeon Jang, Hanning Chen, Sanggeon Yun, Tamoghno Das, and Mohsen Imani. 2026. TorR: Towards Brain-Inspired Task-Oriented Reasoning via Cache-Oriented Algorithm-Architecture Co-design. arXiv preprint arXiv:2603.22855 (2026).arxiv.org/abs/2603.22855
- [26]Tony A Plate. 1995. Holographic reduced representations. IEEE Transactions on Neural Networks 6, 3 (1995), 623–641.
- [27]Tony A. Plate. 2003. Holographic Reduced Representation: Distributed Representation for Cognitive Structures. CSLI Publications, Stanford, CA.
- [28]Prathyush Poduval, Haleh Alimohamadi, Ali Zakeri, Farhad Imani, M Hassan Najafi, Tony Givargis, and Mohsen Imani. 2022. Graphhd: Graph-based hyperdimensional memorization for brain-like cognitive learning. Frontiers in Neuroscience 16 (2022), 757125.DOI
- [29]Ali Rahimi and Benjamin Recht. 2008. Random features for large-scale kernel machines. In Advances in neural information processing systems. 1177–1184.
- [30]Kenny Schlegel, Peer Neubert, and Peter Protzel. 2022. A comparison of vector symbolic architectures. Artificial Intelligence Review 55, 6 (2022), 4523–4555.DOI
- [31]Shay Snyder, Ryan Shea, Andrew Capodici, David Gorsich, and Maryam Parsa. 2025. Generalizable Reinforcement Learning with Biologically Inspired Hyperdimensional Occupancy Grid Maps for Exploration and Goal-Directed Path Planning. arXiv preprint arXiv:2502.09393 (2025).arxiv.org/abs/2502.09393
- [32]Michiel Stock, Wim Van Criekinge, Dimitri Boeckaerts, Steff Taelman, Maxime Van Haeverbeke, Pieter Derwulf, and Bernard De Baets. 2024. Hyperdimensional computing: A fast, robust, and interpretable paradigm for biological data. PLOS Computational Biology 20, 9 (2024), e1012426.arxiv.org/abs/2402.17572
- [33]Anthony Thomas, Sanjoy Dasgupta, and Tajana Rosing. 2021. A theoretical perspective on hyperdimensional computing. Journal of Artificial Intelligence Research 72 (2021), 215–249.DOI
- [34]Calvin Yeung, Zhuowen Zou, and Mohsen Imani. 2024. Generalized Holographic Reduced Representations. arXiv preprint arXiv:2405.09689 (2024).DOI
- [35]Sanggeon Yun, Hyunwoo Oh, Ryozo Masukawa, and Mohsen Imani. 2025. DecoHD: Decomposed Hyperdimensional Classification under Extreme Memory Budgets. arXiv preprint arXiv:2511.03911 (2025).arxiv.org/abs/2511.03911
- [36]Sanggeon Yun, Hyunwoo Oh, Ryozo Masukawa, Pietro Mercati, Nathaniel D Bastian, and Mohsen Imani. 2025. LogHD: Robust Compression of Hyperdimensional Classifiers via Logarithmic Class-Reduction. arXiv preprint arXiv:2511.03938 (2025).arxiv.org/abs/2511.03938
- [37]Zhuowen Zou et al. 2021. Scalable edge-based hyperdimensional learning system with brain-like neural adaptation. In SC. 1–15.DOI
- [38]Zhuowen Zou, Haleh Alimohamadi, Farhad Imani, Yeseong Kim, and Mohsen Imani. 2021. Spiking hyperdimensional network: Neuromorphic models integrated with memory-inspired framework. arXiv preprint arXiv:2110.00214 (2021).arxiv.org/abs/2110.00214
- [39]Zhuowen Zou, Haleh Alimohamadi, Yeseong Kim, M Hassan Najafi, Narayan Srinivasa, and Mohsen Imani. 2022. Eventhd: Robust and efficient hyperdimensional learning with neuromorphic sensor. Frontiers in Neuroscience 16 (2022), 858329.DOI
- [40]Zhuowen Zou, Haleh Alimohamadi, Ali Zakeri, Farhad Imani, Yeseong Kim, M Hassan Najafi, and Mohsen Imani. 2022. Memory-inspired spiking hyperdimensional network for robust online learning. Scientific reports 12, 1 (2022), 7641.DOI






