Skip to content
DraconiticResearch
Search Sources API
220,782 studies31 sourcesUpdated 50 min ago
PreprintCC BY-NC-SAarXiv (Cornell University)

Aperiodic and Low-Frequency Spectral Bias in Reconstruction based EEG Foundation Models

  • Aditya Kommineni1
  • Emily Zhou2
  • Kleanthis Avramidis2
  • Simon Bock Segaard3
  • Jeppe Roden Münster3
  • Andreas Peter Juhl Hansen
  • Takfarinas Medani1
  • Tiantian Feng1
  • Richard M. Leahy1
  • Shrikanth Narayanan1,2
  1. 1Department of Electrical and Computer Engineering University of Southern California Thomas Lord Department of Computer Science
  2. 2University of Southern California Department of Mathematical Sciences
  3. 3Aalborg University
Open at sourcePDFDOICite

Abstract

EEG foundation models, pre-trained on large-scale unlabelled EEG data, have emerged as a promising direction towards learning generalizable EEG representations. Despite showing positive results in data-rich regimes, they often fail to outperform significantly smaller supervised models in low-resource settings compared to fully supervised models. We provide a mechanistic account of this shortcoming, attributing it to a fundamental mismatch between reconstruction-based pretext tasks and the idiosyncratic spectral structure of EEG signals, which decompose into distinct high-power aperiodic and low-power oscillatory components. Using controlled, synthetically-generated EEG inputs, we demonstrate that EEG foundation model embeddings are biased to capture the aperiodic components of the EEG signal while under-representing oscillatory components, particularly at higher frequencies. Additionally, linear probe evaluations on real-world BCI datasets further reveal that embeddings encode subject identity more strongly than task-relevant information, thereby reinforcing the low-frequency and aperiodic component bias in foundation model embeddings trained primarily on reconstruction based objectives. Together, these findings elucidate a failure mode in reconstruction based EEG foundation models and motivate future work to incorporate auxiliary losses explicitly targeting high-frequency oscillatory structure as a path toward more capable and generalizable EEG representations.

Related

  1. Resolution-aware spectral parameterization reveals robust context-dependent aperiodic EEG dynamics during real-world table tennis
    A. Guchait, C.-Y. Chen, N. G. Muggleton
    bioRxivPreprintCC BYFull textNeurosciencebioRxiv
  2. Acute scalp EEG signatures of anterior thalamic DBS in epilepsy:reproducible gamma power enhancement within an aperiodic spectral-state shift
    E. Sabbaghi, H. Mattila, E. Siitama +3
    bioRxivPreprintCC BYFull textNeurosciencebioRxiv
  3. Transcranial Random Aperiodic Stimulation Improves Working Memory Precision
    Q. van Engen, J. Riddle, B. Voytek
    bioRxivPreprintAll rights reservedFull textNeurosciencebioRxiv
  4. Aperiodic dynamics track cortical state shifts during sleep K-complexes
    E. S. Kaplan, J. Schindler, B. Voytek
    bioRxivPreprintAll rights reservedFull textNeurosciencebioRxiv
  5. Overlapping perceptual benefits of temporal attention in aperiodic and periodic streams are mediated by delta phase alignment, but not alpha suppression, in sensory circuits
    C. Bruckmann, A. Breska
    bioRxivPreprintAll rights reservedFull textNeurosciencebioRxiv

Cite

BibTeXRIS
BibTeX
@misc{kommineni2026aperiodic,
  title = {{Aperiodic and Low-Frequency Spectral Bias in Reconstruction based EEG Foundation Models}},
  author = {Aditya Kommineni and Emily Zhou and Kleanthis Avramidis and Simon Bock Segaard and Jeppe Roden Münster and Andreas Peter Juhl Hansen and Takfarinas Medani and Tiantian Feng and Richard M. Leahy and Shrikanth Narayanan},
  howpublished = {arXiv (Cornell University)},
  year = {2026},
  month = may,
  doi = {10.48550/arxiv.2605.26434},
  url = {https://arxiv.org/abs/2605.26434},
  note = {CC BY-NC-SA},
}
RIS
TY  - UNPB
TI  - Aperiodic and Low-Frequency Spectral Bias in Reconstruction based EEG Foundation Models
AU  - Aditya Kommineni
AU  - Emily Zhou
AU  - Kleanthis Avramidis
AU  - Simon Bock Segaard
AU  - Jeppe Roden Münster
AU  - Andreas Peter Juhl Hansen
AU  - Takfarinas Medani
AU  - Tiantian Feng
AU  - Richard M. Leahy
AU  - Shrikanth Narayanan
PY  - 2026
DA  - 2026/05/26
PB  - arXiv (Cornell University)
DO  - 10.48550/arxiv.2605.26434
UR  - https://arxiv.org/abs/2605.26434
AB  - EEG foundation models, pre-trained on large-scale unlabelled EEG data, have emerged as a promising direction towards learning generalizable EEG representations. Despite showing positive results in data-rich regimes, they often fail to outperform significantly smaller supervised models in low-resource settings compared to fully supervised models. We provide a mechanistic account of this shortcoming, attributing it to a fundamental mismatch between reconstruction-based pretext tasks and the idiosyncratic spectral structure of EEG signals, which decompose into distinct high-power aperiodic and low-power oscillatory components. Using controlled, synthetically-generated EEG inputs, we demonstrate that EEG foundation model embeddings are biased to capture the aperiodic components of the EEG signal while under-representing oscillatory components, particularly at higher frequencies. Additionally, linear probe evaluations on real-world BCI datasets further reveal that embeddings encode subject identity more strongly than task-relevant information, thereby reinforcing the low-frequency and aperiodic component bias in foundation model embeddings trained primarily on reconstruction based objectives. Together, these findings elucidate a failure mode in reconstruction based EEG foundation models and motivate future work to incorporate auxiliary losses explicitly targeting high-frequency oscillatory structure as a path toward more capable and generalizable EEG representations.
SN  - 2331-8422
ER  -