Novel and Efficient Hyperdimensional Encoding of Surface

Electromyography Signals for Hand Gesture Recognition

Ancelin Salerno

and Sylvain Barraud

CEA, LETI, Minatec campus, University Grenoble Alpes, 38000 Grenoble, France

Keywords: High-Dimensional Computing, Surface Electromyography Signal, Gesture Recognition, Encoding,

Hypervectors, Ninapro Database.

Abstract: Gesture recognition has become a crucial component of human-computer interaction, with applications

ranging from virtual reality to assistive technologies. This study explores Hyperdimensional Computing

(HDC) as a powerful alternative to traditional machine learning techniques for real-time gesture recognition.

HDC is known for its robustness and efficiency, enabling fast and accurate classification though the use of

high-dimensional binary vectors. In this study, we introduce two key variants aimed at significantly improving

the performance of gesture recognition: (1) an enhancement of item memory representation enabling a better

gestures recognition, and (2) an advanced temporal encoding mechanism that captures the dynamic nature of

gestures more efficiently. These modifications are evaluated using a benchmark dataset of surface

electromyography (sEMG) signals, demonstrating significant improvements in both accuracy and

computational efficiency.

1 INTRODUCTION

Recent advances in the recognition and classification

of surface electromyography (sEMG) signals are

opening up new opportunities in fields such as

human-machine interfaces, robotic control, and

augmented/virtual reality. These advances rely

heavily on the accurate measurement of multi-

channel surface EMG signals and the application of

machine learning (ML) algorithms for gesture

identification. However, deploying ML models on

wearable edge devices presents both challenges and

opportunities. While edge computing enables real-

time processing with reduced latency and improved

privacy through on-device data handling, machine

learning models encounter significant challenges in

addressing the variability of sEMG signals (Hudgins

et al., 1993). Indeed, factors such as muscle fatigue,

electrode displacement, changes in arm posture, and

inter-subject/session variability can severely impact

classification performance, limiting the robustness of

conventional ML approaches.

Current neural network-based solutions, despite

their potential, are resource-intensive, requiring large

https://orcid.org/0009-0008-6432-4270

https://orcid.org/0000-0002-4334-9638

volumes of high-quality training data and incurring

substantial computational and power demands. This

makes their integration into embedded systems,

particularly challenging for real-time gesture

recognition (Hudgins et al., 1993; Benatti et al.,

2014). To address these issues, we introduce

CompHD; a novel hyperdimensional computing

(HDC) framework designed for the efficient encoding

and classification of sEMG signals in hand gesture

recognition.

Unlike traditional HDC methods, which rely on

random or continuous item memories for sequences

encoding (Rahimi et al., 2016; Sgambato &

Castellano, 2022), CompHD incorporates optimized

hyperdimensional representations that allow for more

efficient and accurate processing of gesture data. The

HDC-based brain-inspired architecture offers several

key advantages: it supports one-pass learning,

reducing energy consumption and accelerating the

learning process, while also being highly robust to

noise and computational errors. Moreover, CompHD

requires only a small training dataset to achieve

competitive accuracy, making it well suited for low-

power, real-time applications and most-importantly,

Salerno, A. and Barraud, S.

Novel and Efﬁcient Hyperdimensional Encoding of Surface Electromyography Signals for Hand Gesture Recognition.

DOI: 10.5220/0013257500003911

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 1, pages 763-772

ISBN: 978-989-758-731-3; ISSN: 2184-4305

763

for on-subject training, as sEMG data are highly

individual-specific.

For the first time, we present a comprehensive

quantitative comparison between the CompHD

algorithm and conventional machine learning models

for sEMG-based gesture recognition, using publicly

available datasets, including Ninapro—the largest and

most comprehensive sEMG database. Our evaluation

demonstrates that CompHD not only outperforms

traditional methods in terms of robustness and

computational efficiency, but also achieves

competitive accuracy on this challenging benchmark.

2 DATABASE DESCRIPTION

This study utilizes several publicly available

electromyography (EMG) datasets, including Master

(Rahimi et al., 2016), Pattern (Lobov et al., 2018),

Ninapro DB1, DB4 and DB5 (Atzori et al., 2012,

2014; Pizzolato et al., 2017; Wan et al., 2018). While

these datasets are commonly used for gesture

recognition tasks with various classifiers, we are the

first study to systematically compare them using the

same HDC model. The datasets differ in terms of the

number of gestures, subjects, sampling frequencies

and recording channels, providing a diverse and

comprehensive foundation for developing and

analyzing our HDC model for sEMG signals

recognition. They offer an opportunity to explore

different signal characteristics and patterns, thus

facilitating the development and evaluation of

machine learning models for gesture classification.

The key characteristics of these datasets are

summarized in Table 1.

Table 1: Number of gestures, subjects and channels

considered for each dataset used in this work.

The feature used in this work for gestures

classification is the Mean Amplitude Value (MAV)

which has been shown to be highly effective,

providing both high accuracy and computationally

efficiency (Scheme & Englehart, 2014).

Figure 1: Example of raw (a) and preprocessed (b) sEMG

signals of “Pattern” database. For further clarity, the sEMG

signals measured on 8 electrodes are shown for only one

repetition and only five gestures (including unmarked data).

The MAV is a widely used time-domain feature in

sEMG signals analysis. It represents the average of

the absolute values of the signal over a given time

window. In this study, we use 100ms windows. The

MAV is particularly effective at capturing muscle

activation patterns and is less sensitive to noise

compared to other features, making it well-suited for

applications such as prosthetic control. In the context

of sEMG data, where the frequency band of interest

typically ranges from 10 to 500 Hz, the MAV is

extracted following preprocessing steps such as

filtering to remove artifacts and baseline drift. For the

Ninapro DB1, this preprocessing can be performed

analogically before digitization to ensure that only

relevant muscle activation signals are captured. For

the other datasets, MAV extraction is conducted

using digital filters. This process involves summing

the absolute values of the signal within the specified

window and dividing by the number of samples,

providing a robust measure of the signal’s amplitude.

The MAV is particularly effective in detecting

changes in muscle activity levels, making it crucial

for real-time EMG signal classification.

MASTER

PATTERN

NINAPRO DB1

NINAPRO DB4

Gestures

number

Subjects

number

Database

Channels

number

NINAPRO DB5

52 10 16

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

764

3 HIGH-DIMENSIONAL

COMPUTING CLASSIFIER

The Hyperdimensional Computing (HDC) is an

emerging computational paradigm inspired by the

human brain's approach to processing information

and provides a robust, efficient framework for

managing data represented as high-dimensional

vectors, called hypervectors (HVs) (Kanerva, 2009;

Rahimi et al., 2017; Cohen & Widdows, 2015). The

central concept is to encode information in high-

dimensional vectors that capture rich and complex

patterns. Some mathematical preserving operations

are applied to process or retrieve information. As

information is distributed across all dimensions of a

hypervector, it is less susceptible to interference. This

holistic characteristic inherently enhances the

system’s robustness to noise and partial information

loss, as individual bit errors are unlikely to

compromise the encoded meaning.

3.1 Conventional HDC Encoding

This study specifically focuses on binary HVs, as

binary HDC provides significant energy efficiency

benefits and is particularly well-suited for embedded

device (Chen et al., 2022) or hardware-based in-

memory implementations (Abhijith & Shekhar, 2019;

Benatti et al., 2017; Karunaratne, et al., 2020, 2021;

Li et al., 2016). Figure 2 illustrates the main

components of the HDC algorithm.

The newly highlighted areas correspond to the

specific blocks we have targeted to enhance the

model’s performance. At each stage of the

spatiotemporal encoding process, binary operations

are performed on binary hypervectors. These

operations are dimension-preserving, ensuring that

the dimensionality (DIM) remains consistent across

both the input and output vectors. The core operations

in hyperdimensional encoding — binding and

bundling (Rahimi et al., 2017, Cohen & Widdows,

2015) — enable the creation and manipulation of

complex symbolic structures.

• Binding: This operation combines two

hypervectors (HVs) to create a new one that is distinct

from both. Binding is performed using element-wise

multiplication (or XOR for binary vectors), ensuring

that the resulting hypervector is unique while still

preserving information from both of the original

hypervectors.

• Bundling: This operation combines multiple

hypervectors into a single hypervector that reflects

the common features of all the elements in the set.

Figure 2: Overview of HDC classifier for gesture

recognition accounting for the new HD mapping and

temporal encoder proposed in this work.

Bundling captures similarities among different

vectors representing related information and is

implemented using an element-wise majority

function across the set of hypervectors.

The bundling operation is essential for linking

temporal information across multiple hypervectors

from different timestamps. The hypervector resulting

from bundling n consecutive temporal hypervectors

is referred to as n-gram. In hyperdimensional

computing-based algorithms, n is a hyperparameter

commonly referred to as NGRAM. Both DIM and

NGRAM are critical hyperparameters that

significantly influence the model's performance.

NGRAM captures contextual relationships by

encoding local dependencies, with larger n-grams

offering richer contextual information at the expense of

increased complexity. DIM, on the other hand,

determines the dimensionality of the hypervectors used

to represent data in high-dimensional space. It typically

ranges from several thousand to as many as 10,000

elements. This high dimensionality ensures that

randomly generated vectors are almost orthogonal,

minimizing overlap and enhancing distinguishability.

A higher DIM typically offers greater capacity for

encoding and distinguishing patterns. However, it

requires more computational resources. Striking the

right balance between these parameters is key to

optimizing accuracy, efficiency, and generalization in

hyperdimensional computing systems.

• Class HV of Gesture #1

• Class HV of Gesture #2

…

• Class HV of

Gesture

Similarity check

HD mapping (NEW)

Spatial Encoder

Temporal Encoder (NEW)

Testing

dataset

Learning

dataset

• Query HV (Gesture)

Classified Gesture

Spatio-Temporal Encoder

Preprocessed Signals

Associative Memory

Novel and Efﬁcient Hyperdimensional Encoding of Surface Electromyography Signals for Hand Gesture Recognition

765

3.2 Our Spatio-Temporal Encoder

The first modification we propose involves the

spatiotemporal encoder. In surface electromyography

(sEMG) signal processing, the sequence order of

input data is generally less critical than in tasks like

text or speech recognition, where the position of each

element in a sequence significantly contributes to its

meaning. sEMG signals capture the electrical activity

of muscles during contractions, which typically

produce patterns based on muscle activation levels

rather than strict temporal order. As a result, models

for sEMG analysis can often focus more on feature

extraction and less on temporal dependencies

compared to language processing tasks. This

characteristic allows for the use of simpler

preprocessing techniques and makes certain machine

learning approaches, such as convolutional neural

networks (CNNs), particularly effective for

interpreting sEMG data. Instead of using the binding

operation to encode sequences—preserving both the

values and their order—we propose employing a

bundling operation to compute the mean vector over

a temporal window. This method enhances the

model's robustness to noisy sEMG signals, where

amplitude variability across time samples could

otherwise impair sequence encoding and compromise

signal interpretation.

3.3 Our Novel Composite Encoding

(CompHD)

In text recognition (Abhijith & Shekhar, 2019; Cohen

& Widdows, 2015; Rahimi et al., 2017), mapping

data to high-dimensional hypervectors that are

orthogonal ensures that letters or symbols are

equidistant from each other, preventing any bias in

the representation of specific characters. This

orthogonality is especially advantageous for discrete

classification tasks, where each character is treated as

a distinct entity. However, for continuous data such

as surface electromyographic (sEMG) signals, it is

crucial to capture subtle variations in signal amplitude

and frequency. In this case, projecting the data into

continuous hypervectors offers a more effective

representation of the continuous nature of sEMG

amplitudes (Cohen & Widdows, 2015; Rahimi et al.,

2016/2017; Salerno & Barraud, 2024).

In this case, the distance between hypervectors

should reflect the magnitude of the data, allowing for

smoother transitions and a more nuanced encoding of

continuous signal variations. This continuous

projection is better suited for tasks that require fine-

grained discrimination between data points, such as

analyzing muscle activation patterns in sEMG. Thus,

while orthogonal hypervectors are optimal for

discrete data like text, continuous hypervectors are

more appropriate for representing dynamic,

continuous signals like sEMG.

Mapping data into hypervectors using a

combination of random (i.e., orthogonal)

hypervectors and continuous hypervectors (Figure 3)

offers significant advantages for pattern recognition,

particularly in tasks where data points are highly

similar to previously encountered examples. This

approach leverages the high-dimensional properties

of hypervectors to encode information in a way that

enables fine distinctions between closely related data,

while still preserving the ability to generalize to

broader patterns.

By associating the distance between hypervectors

with differences in their index positions, this method

enhances the recognition of subtle variations in data,

making it highly effective for tasks that require both

precise identification of known patterns and

flexibility in adapting to new, unseen data. This dual

capability is especially valuable in applications like

biomedical signal processing, where small variations

can be critical, yet robust generalization is essential to

handle variability across subjects or conditions

Figure 3: HD mapping scheme used in this work. A new

HD mapping (COM) is proposed to encode the

preprocessed and quantized sEMG signals into binary

hypervectors. The hamming distance between two HVs is

calculated and compared to conventional Random and

Continuous mapping.

3.4 Leave-P-Groups-out

Cross-Validation

In this section, we outline the specific training and

testing procedures used for CompHD, including

dataset partitioning, cross-validation techniques, and

performance evaluation metrics to assess model

accuracy. To evaluate and compare the gesture

recognition rates for different training sizes while

maintaining temporal integrity of gestures repetitions,

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

766

the Leave-P-Groups-Out Cross-Validation (LPGO) is

an excellent choice. LPGO provides a robust

approach for evaluating machine-learning models in

scenarios where data is structured into distinct

groups, such as gesture repetitions in human activity

recognition. This technique systematically partitions

the dataset into smaller groups, such as repetitions,

ensuring that the temporal structure of the data is

preserved during validation. During the inference

phase, the model trained on the remaining groups is

tested on P small groups, known as the left-out

groups. The LPGO cross-validation process is

illustrated in Figure 4.

In the context of finite sets of gesture repetitions,

LPGO ensures that the model is evaluated exclusively

on unseen repetitions, as it is reinitialized at each

iteration. Thus, LPGO closely simulates real-world

variability in user performance. This technique is

particularly valuable for applications like prosthetic

control or gesture recognition, where consistent

performance across repetitions is crucial. By

assessing how well the model generalizes across

different instances of the same gesture, LPGO

provides insights into its robustness.

Figure 4: Protocol used for learning and testing phases.

Leave-P-groups-Out Cross-Validation method is applied to

use all the repetitions for both training and testing. In this

example P=1, which corresponds to a simple 5-fold Cross-

Validation.

4 EMG-BASED GESTURES

RECOGNITION

Using this training and testing protocol, we evaluated

the performance of three state-of-the-art hypervector

memory models: Random Item Memory (RIM),

Continuous Item Memory (CIM), and our composite

model, CompHD (COM), across the databases listed

in Table 1. This comparison allows us to assess the

strengths and limitations of each approach in the

context of gesture recognition, highlighting how

CompHD leverages the advantages of both RIM and

CIM to achieve enhanced performance. The results

are presented in order from the simplest (Master) to

the most complex database (Ninapro). This

organization allows us to clearly demonstrate the

model's effectiveness and robustness, highlighting its

adaptability across different gesture recognition

scenarios.

In Table 2, we present a summary of the database

gesture information and the optimal NGRAM value,

one of the key hyperparameters in our model. The

tuning of these hyperparameters involves identifying

a set of optimal parameter values for the HDC model,

aimed at maximizing performance (i.e., recognition

accuracy) while minimizing memory resource

allocation. Higher-dimensional hypervectors are

critical as they help prevent data loss during

encoding. Several hyperparameters in HDC require

fine-tuning, including the NGRAM value, which

determines the number of bundled hypervectors

during temporal encoding, and the dimensionality

(DIM) of the hypervectors.

To optimize these hyperparameters, we conducted

an extensive Grid search across different NGRAM

values and dimensionalities. The tuning process

involved computing the average performance score

for various combinations of NGRAM and DIM.

Specifically, we evaluated these parameters using a

2D heatmap (see Appendix), which allowed us to

visualize the relationship between the dimensionality

and the NGRAM value. From this analysis, we

identified the optimal parameter set that maximized

the performance score for a given database, ensuring

the best possible recognition rate. These selected

values were then applied consistently across all

subjects in the database.

Table 2: Average gesture duration, optimal NGRAM and

dimension of HVs used for each database.

MASTER

PATTERN

NINAPRO DB1

NINAPRO DB4

Average gesture

duration (a.u.)

Optimal

NGRAM (a.u.)

26.5

25.08

37.06

Database

Dimension

of HVs

8192

16384

NINAPRO DB5

25.75 43 16384

Novel and Efﬁcient Hyperdimensional Encoding of Surface Electromyography Signals for Hand Gesture Recognition

767

4.1 Master Database

The first database, known as the Master database,

contains ten repetitions of five gestures performed by

four subjects (with one subject excluded due to

inconsistent repetitions). In our study, the CompHD

model consistently outperformed both CIM and RIM

in classification accuracy across all training sizes.

Notably, CompHD achieved higher accuracy rates at

each incremental training size, demonstrating its

robustness on both small and large datasets. For some

subjects, CompHD’s classification accuracy reached

an impressive 99.9%, further highlighting its

precision in delivering accurate classifications across

subjects. This consistent advantage at every stage of

training establishes CompHD as a significantly more

effective model than CIM and RIM for achieving

high-accuracy classifications under various training

conditions. Remarkably, CompHD requires only 40%

of the training data to outperform both CIM and RIM,

even when larger training sizes (up to 90%) are used.

This underscores CompHD's superior efficiency and

effectiveness in gesture recognition tasks.

Figure 5: Master database. (a) Classification accuracy was

averaged across all subjects and gestures for different

training sizes. (b) Averaged accuracy achieved per subject

across all five gestures using the largest training size.

4.2 Pattern Database

The second database, known as the Pattern database,

includes four repetitions of seven gestures performed

by 36 subjects. This dataset was recorded using eight

electrodes, providing a rich data source for analyzing

gesture recognition performance across diverse

subjects.

As with the previous database, the CompHD

model consistently outperformed both CIM and RIM

in classification accuracy across all training sizes.

Notably, CompHD achieved higher accuracy rates at

each incremental training size, demonstrating its

robustness on both smaller and larger datasets. For

some subjects, CompHD’s classification accuracy

reached an impressive 100%, further emphasizing its

effectiveness in delivering precise classifications

across different subjects.

This consistent advantage at every stage of

training reinforces CompHD as a significantly more

effective model than CIM and RIM for achieving

high-accuracy classifications under varied training

conditions.

Figure 6: Pattern database. (a) The classification

accuracy, averaged across all subjects and gestures, was

evaluated for different training sizes. (b) Averaged

accuracy achieved per subject across all seven gestures

using the largest training size.

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

768

4.3 Ninapro Database 1

The third database tested, Ninapro DB1, includes ten

repetitions of 52 gestures performed by 27 subjects.

This dataset was recorded using 8 electrodes.

Figure 7: Ninapro DB1. (a) Classification accuracy was

averaged across all subjects and gestures for different

training sizes. (b) Averaged accuracy achieved by each

subject across all 52 gestures using the largest training size.

While RIM outperforms CIM with larger training

sizes, the reverse holds true for smaller sample sizes.

Despite this, the CompHD model consistently

demonstrated superior classification performance

across all training sizes compared to both CIM and

RIM. While the average score amongst subjects is

approximately 70%.

For some subjects, CompHD’s classification

accuracy reached an impressive 85%, further

highlighting its precision in delivering accurate

classifications across subjects, even with more

complex datasets. Notably, CompHD requires only

40% of the training data to outperform both CIM and

RIM, even when larger training sizes (up to 90%) are

used.

4.4 Ninapro Database 4

The fourth database tested, Ninapro DB4, includes six

repetitions of 52 gestures performed by 10 subjects,

recorded using 10 electrodes. As with the previous

databases, the CompHD model consistently

outperformed both CIM and RIM in classification

accuracy across all training sizes.

Figure 8: Ninapro DB4. (a) Classification accuracy was

averaged across all subjects and gestures for different

training sizes. (b) Averaged accuracy achieved by each

subject across all 52 gestures with the largest training size.

For some subjects, CompHD’s classification

accuracy reached nearly 80%, further emphasizing its

ability to deliver precise classifications across

different subjects. Notably, CompHD requires only

50% of the training data to achieve similar accuracies

as both CIM and RIM, even when larger training sizes

(up to 83.33%) are used.

The random Item memory (RIM) outperforms the

Continuous one (CIM) only on the Ninapro database

1 and 4. However, our new Composite Item Memory

(COM) is robust across all database and outperforms

both of the traditional item memories used.

Novel and Efﬁcient Hyperdimensional Encoding of Surface Electromyography Signals for Hand Gesture Recognition

769

4.5 Ninapro Database 5

The fifth database: Ninapro DB5 includes six

repetitions of 52 gestures performed by 10 subjects

and was recorded using 16 electrodes.

Figure 9: Ninapro DB5. (a) Classification accuracy was

averaged across all subjects and gestures for different

training sizes. (b) Averaged accuracy achieved by each

subject across all 52 gestures using the largest training size.

Once again, the CompHD model demonstrated

superior classification performance across all training

sizes compared to both CIM and RIM. For some

subjects, CompHD’s classification accuracy reached

nearly 80%, further highlighting its efficacy in

delivering precise classifications across subjects.

Notably, CompHD requires only 50% of the training

data to outperform both CIM and RIM, even when

larger training sizes (up to 83.33%) are used.

Moreover, CompHD not only achieves better

accuracy than CIM and RIM, but the latter two

models also exhibit significant inconsistencies across

multiple databases. In contrast, CompHD has proven

its robustness across various datasets and training

sizes, consistently outperforming both CIM and RIM

in every case.

Beyond its impressive accuracy and robustness,

HDC (Hyperdimensional Computing) also

demonstrates notable resilience to errors, making it a

compelling choice for applications requiring fault

tolerance. This error resilience enhances HDC’s

suitability for real-world conditions in in-memory

computing, where environmental and electrical

variability, along with data imperfections, are

common challenges.

5 CONCLUSIONS

In this study, CompHD consistently outperformed

both state-of-the-art CIM and RIM, demonstrating

superior classification accuracy across all training

sizes while requiring significantly fewer samples.

This efficiency opens the door to embedded, on-

device training, reducing reliance on large,

centralized datasets and enabling more adaptive,

resource-efficient deployments.

CompHD’s robustness goes beyond performance;

it also exhibits resilience to data variability and

encoding errors, which is a critical advantage for real-

world applications that deal with imperfect data or

noisy environments. These qualities, combined with

its compatibility with low-complexity and massively

parallel operations, position CompHD as a highly

effective choice for embedded systems. Furthermore,

its high accuracy, efficiency, and fault tolerance make

it a promising candidate for in-memory computing

applications.

ACKNOWLEDGEMENTS

This project is supported by the “HDC” exploratory

Carnot project.

REFERENCES

Hudgins, B., Parker, P., Scott, R. N. (1993). A new strategy

for multifunction myoelectric control. IEEE

Transactions on Biomedical Engineering, vol.40(1),

82-94.

Benatti, S., Farella, E., Gruppioni, E., Benini, L. (2014).

Analysis of robust implementation of an EMG pattern

recognition based control. International Conference on

Bio-inspired Systems and Signal Processing 2014

BIOSIGNAL.

Rahimi, A., Benatti, S., Kanerva, P., Benini, L., Rabaey, J.

M. (2016). Hyperdimensional biosignal processing: A

case study for EMG-based hand gesture recognition.

Proceedings of the 2016 IEEE International

Conference on Rebooting Computing (ICRC), San

Diego, CA, USA, pp. 1-8.

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

770

Sgambato, B., Castellano, G. (2022). Performance

comparison of different classifiers applied to gesture

recognition from sEMG signals. In Bastos-Filho, T. F.,

de Oliveira Caldeira, E. M., Frizera-Neto, A. (Eds.),

XXVII Brazilian Congress on Biomedical Engineering.

CBEB 2020. IFMBE Proceedings, Vol. 83. Springer,

Cham.

Salerno, A., Barraud, S. (2024). Evaluation and

implementation of High-Dimensionnal Computing for

gesture recognition using sEMG signals. Proceedings

of the 2024 International Conference on Control,

Automation and Diagnosis (ICCAD)

Lobov, S., Krilova, N., Kastalskiy, I., Kazantsev, V.,

Makarov, V. A. (2018). Latent factors limiting the

performance of sEMG-interfaces. Sensors, 18(4), 1122.

Atzori, M., Gijsberts, A., Castellini, C., et al. (2014).

Electromyography data for non-invasive naturally-

controlled robotic hand prostheses. Scientific Data, 1,

140053

Atzori, M., et al. (2012). Building the Ninapro database: A

resource for the biorobotics community. Proceedings of

the 2012 4th IEEE RAS & EMBS International

Conference on Biomedical Robotics and

Biomechatronics (BioRob), Rome, Italy, pp. 1258-

1265.

Pizzolato, S., Tagliapietra, L., Cognolato, M., Reggiani, M.,

Müller, H., et al. (2017). Comparison of six

electromyography acquisition setups on hand

movement classification tasks. PLOS ONE, 12(10),

e0186132.

Wan, Y., Han, Z., Zhong, J., Chen, G. (2018). Pattern

recognition and bionic manipulator driving by surface

electromyography signals using convolutional neural

network. International Journal of Advanced Robotic

Systems, 15(5).

Scheme, E., Englehart, K. (2014). On the robustness of

EMG features for pattern recognition based myoelectric

control: A multi-dataset comparison. Proceedings of

the 2014 36th Annual International Conference of the

IEEE Engineering in Medicine and Biology Society,

Chicago, IL, USA, pp. 650-653.

Kanerva, P. (2009). Hyperdimensional Computing: An

Introduction to Computing in Distributed

Representation with High-Dimensional Random

Vectors. Cogn Comput 1, 139–159.

Widdows, D., Cohen, T. (2015). Reasoning with vectors: A

continuous model for fast robust inference. Logic

Journal of the IGPL, 23(2), 141-173.

Rahimi, A., et al. (2017). High-dimensional computing as a

nanoscalable paradigm. IEEE Transactions on Circuits

and Systems I: Regular Papers, 64(9), 2508-2521.

Chen H., Najafi, M. H., Sadredini, E., Imani, M. (2022).

Full Stack Parallel Online Hyperdimensional

Regression on FPGA. Proceedings of the 2022 IEEE

40th International Conference on Computer Design

(ICCD), Olympic Valley, CA, USA, 2022, pp. 517-524

Benatti, S., et al. (2017). A sub-10mW real-time

implementation for EMG hand gesture recognition

based on a multi-core biomedical SoC. Proceedings of

the 2017 7th IEEE International Workshop on

Advances in Sensors and Interfaces (IWASI), Vieste,

Italy, pp. 139-144.

Karunaratne G., Rahimi, A., L. Gallo, M., Cherubini, G.,

Sebastian A. (2021). Real-time Language Recognition

using Hyperdimensional Computing on Phase-change

Memory Array," IEEE 3rd International Conference on

Artificial Intelligence Circuits and Systems (AICAS),

Washington DC, DC, USA, 2021, pp. 1-1

Li, H., et al. (2016). Hyperdimensional computing with 3D

VRRAM in-memory kernels: Device-architecture co-

design for energy-efficient, error-resilient language

recognition. Proceedings of the 2016 IEEE

International Electron Devices Meeting (IEDM), San

Francisco, CA, USA, pp. 16.1.1-16.1.4.

Karunaratne, G., Le Gallo, M., Cherubini, G., Benini, L.,

Rahimi, A., Sebastian, A. (2020). In-memory

hyperdimensional computing. arXiv preprint,

1906.01548.

Abhijith, M., Shekhar, G. (2019). Language classification

technique using in-memory high dimensional

computing (HDC). Proceedings of the 2019 2nd

International Conference on Intelligent Computing,

Instrumentation and Control Technologies (ICICICT),

Kannur, India, pp. 293-298.

APPENDIX

Master DB: Heatmap showing the accuracy of our HDC