• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Friday, November 20, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Technology

Spiking Neural Networks Protect Brain Computer Interfaces from Severe Signal Noise

by
November 9, 2026
in Technology
Reading Time: 5 mins read
0
Spiking Neural Networks Protect Brain Computer Interfaces from Severe Signal Noise

Spiking Neural Networks Protect Brain Computer Interfaces from Severe Signal Noise

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

A new comparative benchmark has shown that spiking neural networks can substantially reduce the sensitivity of non-invasive brain-computer interface systems to severe signal corruption, sensor failure, and physiological artifacts. In an experimental evaluation assessing four-class motor imagery decoding from electroencephalography signals, researchers found that replacing continuous-activation artificial neurons with event-driven spiking counterparts allowed decoders to maintain high classification accuracy even under signal disruptions that caused a matched conventional convolutional neural network to collapse to near-chance levels. The investigation, conducted by researchers at Sapienza University of Rome and Tohoku University, demonstrates that threshold-based binary spike computation provides inherent noise resistance for neural decoding pipelines driving robotic actuation.

Brain-computer interfaces translate patterns of neural activity into control commands for external devices, serving critical functions across neurorehabilitation, assistive communication, and human-robot collaboration. Among available non-invasive recording techniques, electroencephalography remains the primary approach because it is portable, relatively inexpensive, and safe. Within these systems, motor imagery paradigms represent a widely studied methodology, requiring users to mentally imagine specific movements, such as moving a limb or grasping an object, without executing them physically. This rehearsal activates coordinated sensorimotor networks in the brain. However, translating these systems outside strictly controlled laboratory environments has remained a fundamental engineering challenge, because real-world electroencephalography recordings are notoriously weak and vulnerable to ambient electromagnetic interference, electrode displacement, and involuntary user movements.

To rigorously evaluate how different neural architectures endure signal degradation, the research team established an experimental framework benchmarking a compact one-dimensional convolutional baseline, designated ConvShallowNet, against two spiking neural network counterparts designated as ConvShallowNetSpiking. The spiking models incorporated either standard Leaky Integrate-and-Fire neurons or Adaptive Leaky Integrate-and-Fire neurons. Crucially, the architectures were constructed with matching capacity to ensure that differences in resilience could not be attributed to model parameter size or network depth. The conventional network possessed 43,624 trainable parameters, while the spiking variants comprised 43,264 parameters, with the minimal difference arising strictly from the batch normalization statistics utilized exclusively by the non-spiking model.

The benchmark evaluated decoding performance using the established BNCI 2014-001 motor imagery dataset, which includes 22-channel electroencephalography recordings from nine healthy subjects performing four distinct imagined motor tasks: left hand, right hand, both feet, and tongue movement. Each participant took part in two sessions conducted on separate days, performing 288 trials per session. Signals were sampled at 250 Hz, band-pass filtered between 0.5 and 100 Hz, and processed with a 50 Hz notch filter to eliminate power line noise. To eliminate confounding variables associated with inter-subject and inter-session distribution shifts, the researchers pooled data across all subjects and sessions into a 70% training, 15% validation, and 15% testing split, maintaining identical input tensors and scale across all tested network models.

On pristine, uncorrupted electroencephalography data, the conventional convolutional network achieved a mean classification accuracy of 61.52%, holding a modest advantage over the standard Leaky Integrate-and-Fire model at 58.34% and the adaptive variant at 59.12%, where balanced four-class chance accuracy is 25.00%. However, this slight baseline superiority reversed dramatically when signal perturbations were introduced. Under global white Gaussian noise injected across all recording channels at a standard deviation of one, the conventional convolutional model suffered a devastating 58.56% relative accuracy drop, plunging from 61.52% to 25.49% and effectively reducing its performance to random guessing. In sharp contrast, the standard Leaky Integrate-and-Fire spiking network experienced an accuracy drop of only 0.48%, and the Adaptive Leaky Integrate-and-Fire model dropped by only 0.80%, maintaining functional decoding accuracies of 58.06% and 58.65%.

The researchers also investigated sensor dropouts by testing single-channel deletion, which represents a severe failure mode when physical electrodes lose scalp contact or malfunction. When models trained on clean data faced the sudden complete removal of an individual channel, the conventional convolutional baseline lost an average of 41.84% of its decoding accuracy across the montage. The spiking variants proved substantially more tolerant to missing inputs, with the standard leaky model experiencing an average drop of 28.78% and the adaptive model losing 33.07%. When single-channel Gaussian noise was applied, the baseline dropped by an average of 4.61%, with losses reaching up to 19.08% on critical sensor locations, whereas the spiking architectures demonstrated average drops of just 0.01% and 1.32% across the electrode array.

To explore scenarios where calibration data is gathered under compromised conditions, the researchers tested a paradigm in which networks were trained on corrupted data and subsequently evaluated on clean signals. Under general noise, the conventional model exhibited an 18.33% accuracy deficit, whereas the standard spiking model showed a modest drop of 3.01% and the adaptive model dropped by 7.38%. The most pronounced separation of the entire study emerged under the channel deletion condition within this training paradigm: when trained on data lacking an electrode and tested on the complete 22-channel array, the baseline accuracy collapsed by an average of 36.00%. In contrast, the standard spiking network dropped by just 0.72% and the adaptive model dropped by a negligible 0.34%, proving that spiking decoders can generalize cleanly when missing sensory channels are restored.

Beyond stylized Gaussian noise and channel loss, the investigation analyzed model performance under semi-simulated physiological contamination, incorporating real ocular waveforms from electrooculograms, real muscular signals from electromyograms, and synthetic cardiac traces from electrocardiograms. While physiological artifacts possess temporal and spectral structures that partially overlap with motor imagery bands, spiking decoders consistently degraded less than the conventional network across all artifact categories. Under full-montage contamination at a 0 dB signal-to-noise ratio, the adaptive spiking model experienced relative drops of 19.6% for ocular, 47.0% for muscular, 4.9% for cardiac, and 38.8% for mixed artifacts, outperforming the baseline which suffered corresponding drops of 27.8%, 55.1%, 18.5%, and 49.8%.

The underlying mechanism enabling this resilience stems from the event-driven non-linearity of spiking neurons. In the spiking architecture, convolutional feature maps are converted into sparse binary spike trains by Leaky Integrate-and-Fire units operating as thresholding elements. Synaptic inputs must cross a strict threshold potential to trigger a spike event, meaning continuous background fluctuations and sub-threshold noise are intrinsically filtered out before reaching downstream layers. In addition to mathematical robustness, this sparse binary operation offers significant computational benefits. In the spiking network, the second convolutional layer processes discrete binary events rather than floating-point values, converting 18.00 million multiply-accumulate operations per trial into energy-efficient accumulate-only additions. Based on a standard 45-nanometer hardware proxy, this conversion reduces estimated arithmetic energy demand from 113.2 microjoules per trial to between 30.6 and 47.2 microjoules.

To validate real-time operational feasibility, the authors embedded the spiking model into an end-to-end decoding-to-actuation pipeline controlling a Franka Emika Research 3 robotic arm through the Robot Operating System 2 environment. Replaying pre-recorded electroencephalography streams, the spiking decoder reliably mapped motor imagery classes to high-level kinematic actions, including circular Cartesian motions, pick-and-place routines, forward pushes, and pouring trajectories generated with cubic spline interpolation. The robotic arm successfully carried out all commanded tasks without communications faults or controller exceptions, confirming that the spiking decoder satisfies the temporal constraints required for robotic interfaces.

The study authors noted important constraints surrounding their findings. The primary evaluation relied on synthetically added Gaussian noise and stylized channel deletions to strictly isolate architectural effects from complex biological variables, which trades off real-world artifact complexity for experimental control. Furthermore, the experiments focused exclusively on a single four-class motor imagery benchmark using a subject-pooled data split, meaning the findings do not directly measure performance in subject-independent calibration regimes or across evoked-potential paradigms such as P300 or steady-state visual evoked potentials. Closed-loop human adaptation was also not evaluated. Nonetheless, the results indicate that event-driven spiking architectures provide an effective, energy-efficient path toward robust neural decoding in noisy real-world environments.

Subject of Research: Noise resilience of spiking neural networks in EEG-based brain-computer interfaces for robotic arm control

Article Title: Noise resilience in brain decoding driven by spiking neural networks for responsive robot control

Article References: Doku, M., Hayashibe, M., & De Luca, A. (2026). Noise resilience in brain decoding driven by spiking neural networks for responsive robot control. Frontiers in Neurorobotics, 20, Article 1931325. https://doi.org/10.3389/fnbot.2026.1931325

Image Credits: AI Generated

DOI: 10.3389/fnbot.2026.1931325

Keywords: brain-computer interfaces, spiking neural networks, electroencephalography, motor imagery, noise resilience, robot control, computational neuroscience, deep learning, Noise, resilience, brain, decoding

News Source: Cassandra Pierce. (November 9, 2026). Spiking Neural Networks Protect Brain Computer Interfaces from Severe Signal Noise. Scienmag.

Tags: Brainbrain-computer interfacesComputational neuroscienceDecodingdeep learningelectroencephalographymotor imageryNoisenoise resilienceresiliencerobot controlspiking neural networks
Share12Tweet7Share2ShareShareShare1

Related Posts

Cross-Attention Multimodal Framework Halves Climate Prediction Error Across Ukrainian Data

Cross-Attention Multimodal Framework Halves Climate Prediction Error Across Ukrainian Data

November 20, 2026
Nanostructured Oxide Fillers Stabilize Ionic Conductivity in Eutectic Salt Composites

Nanostructured Oxide Fillers Stabilize Ionic Conductivity in Eutectic Salt Composites

November 20, 2026

Buffalo Bone Mineral Alters PLLA and PCL Composite Printing Filament Properties

November 20, 2026

Biased Bubble Chain Navigation Helps Indoor Mobile Robots Accommodate Approaching Pedestrians

November 20, 2026

POPULAR NEWS

  • Biased Bubble Chain Navigation Helps Indoor Mobile Robots Accommodate Approaching Pedestrians

    Biased Bubble Chain Navigation Helps Indoor Mobile Robots Accommodate Approaching Pedestrians

    29 shares
    Share 12 Tweet 7
  • Thermophilic Fungus Engineered for High-Yield Bacterial Dipeptidyl Peptidase 4 Secretion

    29 shares
    Share 12 Tweet 7
  • Anterior Talar Osteotomy Corrects Equinus Deformity While Preserving Joint Congruence

    29 shares
    Share 12 Tweet 7
  • Biosensor Mouse Shows Mammary Defects Reversed by Estrogen and Genetic Crossing

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Microelectronic Telemetry Chips Monitor Chemical Reactions Wirelessly Inside Test Tubes

Study Links METTL7B and SPP1 Signaling to Lung Squamous Cell Progression

Elastic Nanovesicle Gel Delivers Vardenafil to Relieve Peripheral Tissue Ischemia

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 89 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.