Hyperdimensional computing in biomedical sciences: a brief review

Key principles and advantages of hyperdimensional computing

At its core, HDC operates on the following key principles (Kanerva, 2009):

• (1)

  High-dimensional representations: information is encoded using random binary or bipolar high-dimensional vectors, usually with 10-thousand dimensions. Note that this number is not arbitrary. In high-dimensional spaces, randomly generated vectors tend to be nearly orthogonal. This property guarantees that vectors representing different information remain distinct and easily distinguishable in the same space;

• (2)

  Randomized encoding: data is mapped to hypervectors using randomized encoding techniques, ensuring that similar data points are mapped to nearby points in the hyperspace, regardless of the nature of data (e.g., images, text, biomedical signals);

• (3)

  Holographic representation: information is distributed across the entire hypervector, providing robustness to noise and data corruption. Instead of having localized or sparse encoding where a small subset of dimensions carries most of the information, every component of the hypervector contributes to the representation. This brings a series of advantages:

  • (a)

    Robustness: since the information is spread out, the representation is resilient to noise or corruption in any subset of the dimensions. Even if some dimensions are altered or lost, the overall information can be recovered;

  • (b)

    Graceful degradation: partial damage leads to a proportional decrease in the fidelity of the representation rather than a catastrophic loss;

• (4)

  Simple operations: computation is performed using simple arithmetic operations like the element-wise addition, multiplication, and permutation of vectors. These operations are at the base of the so called Multiply-Add-Permute (MAP) model, and each of them carries different and unique mathematical properties:

• a.

  Binding ⊗: it is implemented as element-wise multiplication (or, for binary hypervectors, XOR). It brings the following mathematical properties:

  • (i)

    Invertibility: the binding operation is approximately invertible. It means that, given $c=a\otimes b$, both $a$ and $b$ can be recoverable;

  • (ii)

    Associativity and commutativity: the order of binding operations does not affect the result.

• b.

  Aggregation/bundling ⊕: it aggregates multiple hypervectors to form a single hypervector representation of a set or group of items $c=a\oplus b$. It is implemented as the element-wise sum followed by a normalization to make sure that the resulting hypervector remains in the same representational space as the original hypervectors. It is characterized by the following properties:

  • (i)

    Superposition principle: The bundling operation distributes information from all contributing hypervectors across the resulting hypervector in a way that retains key features of the original set. While individual hypervectors are not perfectly preserved, their influence remains in a distributed manner that allows to recognize and retrieve patterns from the aggregated representation;

  • (ii)

    Noise tolerance: because the operation involves summing up many high-dimensional vectors, the signal of interest is maintained while random noise tends to cancel out due to averaging effects.

• c.

  Permutation $p$: it is used to introduce order or structure into the hypervector representations. A permutation operation $p$ acts as a fixed, bijective reordering of the dimensions of a hypervector, shifting elements by a fixed number of positions. It also brings the following properties:

  • (i)

    Invertibility: since a permutation is a bijective mapping, it is invertible. This means that the original order of the elements can be recovered by applying the inverse permutation, or shifting elements back to the same number of positions initially used to build the permuted representation of the hypervector;

  • (ii)

    Preservation of similarity: although permutation changes the positions of the elements, it preserves the overall distribution and statistical properties of the hypervectors.

The set of hypervectors, together with the encoding logic and its arithmetic operations on vectors, is called vector-symbolic architecture.

These principles give rise to several advantages that make HDC particularly well-suited for a wide range of applications:

• Robustness to noise and errors: the distributed nature of information in hypervectors makes models encoded through the HDC paradigm tolerant to noise and errors;

• Computational efficiency: simple vector operations can be implemented efficiently in hardware, leading to fast and energy-efficient computation. An example is the development of an FPGA-based accelerator for HDC, which achieved significant speedups and energy savings compared to traditional CPU implementations. This makes HDC suitable for real-time applications in resource-constrained environments (Chen, Barkam & Imani, 2023);

• Data agnosticism: HDC can be applied to any kind of data, regardless of its nature, making it extremely versatile and suitable for a wide range of applications, ranging from robotics (Hassan et al., 2024), to natural language processing (Berster, Caleb Goodwin & Cohen, 2012), image recognition (Neubert & Schubert, 2021), and many others, demonstrating the HDC’s ability to handle diverse data sources in very different and specific contexts efficiently;

• Scalability: the inherent parallelism of HDC allows for efficient scaling to handle large datasets and complex computations. It is crucial for applications involving big data and real-time processing (Heddes et al., 2024).

The Dollar of Mexico and the promise of hyperdimensional computing

The foundation of HDC is rooted in the idea that information can be represented and manipulated using high-dimensional vectors, typically consisting of thousands of dimensions. One of the earliest illustrations of this concept is the “What’s the Dollar of Mexico?” problem (Kanerva, 2009), introduced by Pentti Kanerva to demonstrate how distributed representations can encode semantic relationships. The problem states that if a system knows the relationships Dollar → USA and Peso → Mexico, it should infer that Dollar of Mexico likely refers to Peso, despite never having encountered this specific phrase before.

Mathematically, this is achieved using vector-symbolic architectures (VSAs), where concepts are represented as high-dimensional random vectors, and relationships are encoded through arithmetic operations such as vector binding and bundling as discussed above. Given the vector representations of the concepts Currency and Country, and Dollar and USA, their relationship can be represented as:

$$V_{USA}=\left(Currency\otimes Dollar\right)\oplus \left(Country\otimes USA\right),$$where ⊗ and ⊕ denote the binding and bundling operations, respectively. Analogously, the relationship between Peso and Mexico can be defined as V_Mexico = $\left(Currency\otimes Peso\right)\oplus (Country\otimes Mexico)$. Finally, the binding between $V_{USA}$ and $V_{Mexico}$ produces a new vector that represents the reality that we are trying to encode. In order to answer the initial question, a simple binding between this final vector with the vector representation of Dollar produces a new vector which is very similar to the vector representation of Peso. This can be verified by computing the cosine similarity between this final vector and all the other vectors computed so far.

This principle—where information is stored and retrieved based on patterns of similarity rather than explicit lookup tables—forms the backbone of HDC. Over the years, researchers have leveraged these properties to build robust models capable of handling noisy, uncertain, and large-scale data, making HDC particularly appealing for different kinds of problems, including biomedical applications.

It represents a paradigm shift in computing, offering a powerful and efficient alternative to traditional approaches. Its unique ability to handle high-dimensional data, robustness to noise, and computational efficiency make it a promising candidate for addressing the challenges posed by the ever-growing complexity and scale of data-driven applications. As research in HDC continues to advance, we can expect to see its adoption in an even wider range of fields, paving the way for a new era of intelligent and efficient computing systems.

Hyperdimensional computing and biomedical sciences

While a broader potential of HDC is clear, its application to biomedical sciences presents a particularly compelling case. This review focuses specifically on the intersection of HDC and biomedical sciences, encompassing medical informatics, bioinformatics, and cheminformatics. As illustrated in Fig. 1, the number of publications exploring HDC within these domains has steadily increased over the past decade (with the exception of 2019 and 2023), highlighting the growing interest and potential of this technology in these specific scientific domains.

This review is particularly crucial due to the rapid evolution and relative novelty of HDC in the biomedical domain. While promising, this field is still in its early stage, characterized by a diversity of approaches and a lack of standardized methodologies.

The motivations for this review systems from several key factors that distinguish it from previous surveys, particularly those with a general or solely technical focus:

• (1)

  The biomedical applications of HDC are evolving at an unprecedented rate, with novel approaches and implementations emerging frequently. While previous surveys have outlined the theoretical and algorithmic foundations of HDC (Stock et al., 2024), this review captures the latest trends and breakthroughs specific to biomedical sciences. It offers an updated perspective that reflects the current research landscape, making it indispensable for readers who wish to stay abreast of state-of-the-art developments;

• (2)

  As mentioned above, unlike prior surveys that address HDC from a broad, interdisciplinary viewpoint, this paper zooms in on biomedical applications. By bridging the gap between general HDC methodologies and domain-specific needs, this review serves as a specialized resource for biomedical researchers and practitioners;

• (3)

  The field of HDC in biomedicine is still in its nascent stages, characterized by a diversity of approaches and a lack of standardized methodologies. This review synthesizes the existing literature to help readers identifying promising avenues for future innovations in this field;

Our target audience encompasses both researchers actively working in HDC and those in related biomedical fields seeking to explore the potential of this novel computational paradigm. By providing them with a critical evaluation of the current landscape, as well as identifying gaps and future directions, this review aims to equip readers with the knowledge and insights needed to advance HDC applications in biomedical sciences.

In summary, this article offers a unique contribution by focusing specifically on the intersection of HDC and biomedicine, delivering a synthesis of recent progress, a critical evaluation of existing challenges, and practical guidance for future research. For readers seeking to understand not only the theoretical foundations but also the practical potential and future trajectory of HDC in biomedical sciences, this review serves as an essential resource that complements and extends previous surveys.

To the best of our knowledge, no surveys on biomedical applications of HDC exist in the scientific literature. The already-mentioned study by Stock et al. (2024) recently presented a technical guide or introduction aimed at explaining the fundamentals of HDC and demonstrating its potential through specific applications in bioinformatics. In particular, their article is structured to provide an introduction to the HDC concepts, outline the computational advantages of HDC from a general perspective, and highlight selected case studies and applications across the bioinformatics field. Although useful and interesting, this article does not outline the applications of HDC in biomedical sciences; we fill this gap by presenting our short survey here.

Search strategy and selection criteria

Our bibliographic research employed a systematic search strategy across Google Scholar and the collection of scientific articles in the hd-computing repository at https://www.hd-computing.com. The latter represents a valuable resource to help the HDC/VSA scientific community by collecting software, video courses and webinars, in addition to scientific articles about the application of the HDC paradigm to specific technological problems.

To ensure a comprehensive and unbiased review, we conducted a systematic search of the literature using a multi-pronged approach. In particular, we utilized a combination of keywords and their variations to ensure a broad capture of relevant articles among those operating in the context of biomedical sciences: (“Hyperdimensional Computing” OR “Vector-Symbolic Architectures”) AND (“Bioinformatics” OR “Cheminformatics” OR “Medical Informatics” OR “Biomedical Sciences”)

Our search yielded a pool of 41 articles. To focus our review on the most relevant studies, we applied the following inclusion criteria:

• Focus on HDC: articles must primarily focus on the application of HDC techniques;

• Relevance to biomedical sciences: articles must address problems within bioinformatics, cheminformatics, or medical informatics;

• Peer-reviewed publications: we prioritized peer-reviewed journal articles and conference proceedings articles. However, we also considered preprints of significant interest and potential impact.

Data extraction and categorization

We extracted a comprehensive set of features from each selected article in support of our analysis of HDC applications in biomedical sciences. These features were carefully chosen to provide insights into various aspects of the research, including:

• (1)

  Bibliographic information:

  • Theme: categorized the primary application domain of the article into bioinformatics, cheminformatics, or medical informatics;

  • Venue: recorded the journal or conference name where the article was published;

  • Article type: classified the publication type as a journal article, conference proceedings article, or preprint;

  • Open access: noted whether the article was freely available in an open-access format.

• (2)

  HDC methodology:

  • Vector construction: documented the specific methods used to construct the hypervectors, such as random projection, learned embeddings, or domain-specific encodings;

  • Data encoding: described the techniques employed to encode different data types (e.g., DNA sequences, molecular structures, patient records) into hypervectors;

  • Data combination: analyzed the operations used to combine encoded data, including vector addition, multiplication, and permutation operations.

• (3)

  Application and evaluation:

  • Scientific problem: summarized the specific biomedical problem addressed in the article;

  • Number of patients and samples: recorded the dataset size, including the number of patients or samples used in the study;

  • Hardware: extracted information about any specialized hardware utilized for HDC implementation, such as classical CPUs, GPUs, ASICs, or FPGAs;

  • Robustness: analyzed any reported metrics or discussions related to the model’s robustness to noise;

  • Programming language(s): identified the programming language(s) used for implementing HDC models (for example, Python, MATLAB, R, C/C++, and Rust);

  • Software availability: determined whether the software or code developed for the study was publicly available.

Analysis framework

The extracted data allowed us to analyze the trends and characteristics of HDC applications in biomedical sciences. We will present our findings by focusing on each of the extracted features, highlighting the following aspects:

• Themes: we will discuss the prevalence of HDC in each of the three themes (bioinformatics, cheminformatics, and medical informatics), identifying areas where HDC has been particularly successful or is gaining traction;

• Methodological Trends: we will analyze the commonalities and differences in HDC methodologies employed across different studies, examining the suitability of specific techniques for particular biomedical problems;

• Open Science Practices: we will investigate the adoption of open science practices, such as open access publication and software sharing, within the HDC community.

By systematically analyzing these features, this literature review aims to provide a comprehensive and insightful overview of the current landscape of HDC in biomedical sciences, identifying research gaps, and highlighting future directions for this promising field.

We identified a pool of 62 research articles over three scientific themes, i.e., medical informatics (Buteau et al., 2024; Chen et al., 2024a, 2024b; Du et al., 2024; Gaddi et al., 2024; Ponzina et al., 2024; Salerno & Barraud, 2024; Cohen et al., 2012; Moon et al., 2013; Kleyko et al., 2016, 2017; Rahimi et al., 2017a, 2017b, 2018; Burrello et al., 2018, 2019a, 2019b, 2021; Lagunes & Lee, 2018; Burkhardt et al., 2019; Asgarinejad, Thomas & Rosing, 2020; Billmeyer & Parhi, 2021; Ge & Parhi, 2021, 2022; Menon et al., 2021, 2022; Pale, Teijeiro & Atienza, 2021, 2022a, 2022b, 2022c, 2023; Watkinson et al., 2021a, 2021b; Schindler & Rahimi, 2021; Ni et al., 2022; Shahroodi et al., 2022; Ung, Ge & Parhi, 2022; Wang, Ma & Jiao, 2022; Topic et al., 2022; Segura et al., 2024; Ge et al., 2024; Xu & Parhi, 2024; Katoozian, Hosseini-Nejad & Dehaqani, 2024; Jeong et al., 2024; Colonnese et al., 2025), bioinformatics (Barmpas et al., 2024; Fan et al., 2024; Mohammadi et al., 2024; Pinge et al., 2024; Imani et al., 2018; Kim et al., 2020; Cumbo, Cappelli & Weitschek, 2020; Poduval et al., 2021; Chen & Imani, 2022; Zou et al., 2022; Barkam et al., 2023; Verges et al., 2024; Xu et al., 2024; Cumbo et al., 2024), and cheminformatics (Ma, Thapa & Jiao, 2022; Jones et al., 2023, 2024), with 32 of them falling under the medical informatics category (78.05%), 7 belonging to the bioinformatics domain (17.07%), and the remaining 2 about cheminformatics (4.88%) (Fig. 2B).

In the context of three macro thematic categories, we manually classified the whole set of 62 articles based on the specific scientific problem that the authors aimed at addressing with their research. In particular, we identified 28 scientific problems, with the most discussed being seizure detection (14/39 articles under the medical informatics theme), followed by biosignal classification and sequence matching and alignment with roughly the same number of manuscripts (7/39 articles under the medical informatics theme, and 6/20 articles under the bioinformatics theme, respectively) (Fig. 2A).

Examining the methodological choices employed across the surveyed literature reveals a diverse landscape with a few dominant trends and some concerning ambiguities (Fig. 3).

In particular, Fig. 3A highlights the prevalence of random vector generation methods, with ~50% of the studies utilizing either binary (23 articles) or bipolar (12 articles), and the other ~50% failing to explicitly specify their approach (27 articles), making it difficult to assess the potential impact of vector generation choices on the reported results.

Analogously, Fig. 3B shows the different data encoding techniques employed in all the 62 studies, revealing a perfect balance between record-based encoding (20 articles) and n-gram representations (20 articles). Interestingly, a small subset (5 articles) employs a combination of both. However, a concerning number of studies (17 articles) again lack specificity in describing their encoding strategy.

Finally, Fig. 3C examines data modeling techniques, revealing a striking dominance of the MAP-model (45 articles). This finding suggests a potential convergence towards this approach within the HDC community, at least in the context of biomedical applications. However, a substantial number of studies (17 articles) also fails in clearly specifying the employed modeling technique.

It is worth noting that one critical aspect that still remains largely unexplored in the literature considered in this review is the role of randomness in hypervector generation and its potential impact on both model accuracy and hardware efficiency. This lack of investigation is likely due to the intrinsic nature of HDC, where information is encoded into randomly generated hypervectors by design. Since randomness is a fundamental principle of HDC, it is very likely that authors implicitly accept it without examining how different randomness strategies might affect accuracy or computational efficiency. However, vector generation and encoding are among the most computationally expensive steps in an HDC system, and understanding potential trade-offs between randomness, accuracy, and hardware performance could be particularly valuable specifically for biomedical applications, where precision and resource constraints are critical. Future research should explore whether alternative encoding methods—such as semi-random or structured approaches—could improve performance while maintaining the robustness of HDC models.

It is also worth to note that ~40% of the research articles considered in this study that propose a HDC-based software solution also do not mention the hardware architecture for which their software has been designed to work on (Fig. 4A). Furthermore, the limited exploration of specialized hardware like ASICs and neuromorphic systems indicates a potential gap in leveraging the full potential of HDC for computationally demanding biomedical tasks.

As a direct consequence, when the authors do not mention the specific hardware for which their software has been designed and tested, they also usually omit the programming language used to develop their software (Fig. 4B). For all the other cases, Python is the most used programming language in this context, followed by MATLAB, and finally R, C/C++, and Rust with no more than two references each.

A lack of transparency and reproducibility is evident in the alarming trend of withholding the source code used to design and build the HDC models in these specific scientific fields of biomedical sciences (Fig. 4C). This lack of openness represents a fundamental breach of scientific principles, severely hindering the progress and credibility of the field and raising concerns about the thoroughness and scientific rigor of these studies, clearly standing in the opposite direction with respect to the core values of FAIR principles (Wilkinson et al., 2016).

More importantly, the refusal to share source code is even more concerning considering that, in most of the cases, authors have taken the time to assign specific names to their software, highlighting a deliberate choice towards obfuscation rather than openness. Without access to the code, the scientific community is left to accept results on blind faith. Independent verification becomes impossible, hindering the identification of potential errors, biases, or overstated claims. This practice stifles collaborative progress, as researchers are unable to build upon existing work, forcing them to reinvent the wheel or resort to potentially less efficient solutions.

Analyzing the publication venues of HDC research in biomedical sciences reveals a significant skew towards conference proceedings over peer-reviewed journals (Table 1 for a list of journals and number of published articles, and Table 2 for a list of conferences alongside the number of articles published in their proceedings and their publisher). While it is generally common in rapidly evolving fields, this trend raises concerns one more time about the long-term impact and acceptance of HDC within the broader scientific community.

As it is evident from this data, a striking majority of publications appear in conference proceedings, with IEEE conferences dominating the landscape. While conferences offer a valuable platform for a rapid dissemination of research findings, they often lack the rigorous peer-review processes typical of academic journals.

The disparity in publication venues raises concerns about the rigor and transparency of HDC research in biomedical sciences. The fact that the number of articles published in conference proceedings (34) significantly outweigh those published in peer-reviewed journals (21) suggests a potential lack of peer-review scrutiny (Fig. 5A).

In fact, conference proceedings publications allow for a faster spread of new ideas, but they have several drawbacks compared to journal publications. Typically, the peer review process for conference proceedings is more superficial, as conference organizers need to accept a certain number of articles to cover the costs of organizing the conference. Additionally, conference proceedings are usually short (4–8 pages), which limits the space available for demonstrations, specific experiments, and in-depth scientific discussions (Ernst, 2006).

Moreover, journal articles have a stronger impact on scientific literature than conference proceedings (Lisée, Larivière & Archambault, 2008). As Cynthia Lisée and coauthors explain: “The evidence thus shows that conference proceedings have a relatively limited scientific impact, on average representing only about 2% of total citations, that their relative importance is shrinking, and that they become obsolete faster than the scientific literature in general” (Lisée, Larivière & Archambault, 2008).

Additionally, Fig. 5B highlights the overwhelming influence of IEEE (40 articles) as the primary publisher in the field. This dominance, while indicative of the strong presence of HDC research in IEEE conferences, also raises concerns about potential publication biases and the need for greater diversity in dissemination channels.

Most concerning, however, is the stark disparity in open access publication revealed in Fig. 5C. A mere 27 articles are classified as open access, compared to a staggering 35 articles that are not freely available.

In our literature review, we also collected data regarding the most present single authors (Table 3) and the most present groups of authors (Table 4). As one can notice, Tajana Rosing (currently at University of California San Diego), Abbas Rahimi (currently at IBM Research Zurich), and Mohsen Imani (currently at University of California Irvine) resulted being the researchers with the highest number of authored studies in our survey, with nine and seven publications respectively (Table 3).

It is also worth noting, as we highlighted in Table 3, that the most present single authors are more leaning towards the study of three scientific problems, that are biosignal classification, seizure detection, and sequence matching and alignment.

Regarding author groups, the team consisting of David Atienza, Una Pale, and Tomas Teijeiro has resulted being quite productive, by publishing five scientific articles on HDC applied to biomedical data (Table 4). The following most present groups of authors are (i) Luca Benini, Alessio Burrello, Abbas Rahimi, and Kaspar Schindler, (ii) Lulu Ge and Keshab Parhi, (iii) and Tony Givargis, Victor Jow, Alexandru Nicolau, Alexander Vaidenbaum, and Neftali Watkinson with two published scientific articles for each group. Other groups of authors have less publications. Interestingly, Abbas Rahimi and Mohsen Imani are the most productive authors in our survey, but they do not belong to the groups of most prolific authors.