Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Contrastive and Multi-Task Learning on Noisy Brain Signals with Nonlinear Dynamical Signatures

Sucheta Ghosh, Zahra Monfared, Felix Dietrich

TL;DR

This work tackles robust EEG motor imagery decoding in noisy, non-stationary signals by integrating denoising, dynamical-system based chaos detection, and self-supervised representation learning. It introduces a two-stage pipeline: a denoising autoencoder preprocesses EEG, followed by a multitask CNN–Transformer with three heads for real-vs-imagery MI classification, Lyapunov-exponent-based chaos tagging, and NT-Xent contrastive learning. Key contributions include a DAE pretraining scheme, LE-based chaos labeling via shallow PLRNNs, a shared encoder with three task heads, and comprehensive ablations showing synergistic gains across two benchmark datasets. The approach yields improved robustness and generalization, surpassing several state-of-the-art EEG MI methods and illustrating the value of combining denoising, dynamical signatures, and self-supervised learning for noisy neural signals.

Abstract

We introduce a two-stage multitask learning framework for analyzing Electroencephalography (EEG) signals that integrates denoising, dynamical modeling, and representation learning. In the first stage, a denoising autoencoder is trained to suppress artifacts and stabilize temporal dynamics, providing robust signal representations. In the second stage, a multitask architecture processes these denoised signals to achieve three objectives: motor imagery classification, chaotic versus non-chaotic regime discrimination using Lyapunov exponent-based labels, and self-supervised contrastive representation learning with NT-Xent loss. A convolutional backbone combined with a Transformer encoder captures spatial-temporal structure, while the dynamical task encourages sensitivity to nonlinear brain dynamics. This staged design mitigates interference between reconstruction and discriminative goals, improves stability across datasets, and supports reproducible training by clearly separating noise reduction from higher-level feature learning. Empirical studies show that our framework not only enhances robustness and generalization but also surpasses strong baselines and recent state-of-the-art methods in EEG decoding, highlighting the effectiveness of combining denoising, dynamical features, and self-supervised learning.

Contrastive and Multi-Task Learning on Noisy Brain Signals with Nonlinear Dynamical Signatures

TL;DR

This work tackles robust EEG motor imagery decoding in noisy, non-stationary signals by integrating denoising, dynamical-system based chaos detection, and self-supervised representation learning. It introduces a two-stage pipeline: a denoising autoencoder preprocesses EEG, followed by a multitask CNN–Transformer with three heads for real-vs-imagery MI classification, Lyapunov-exponent-based chaos tagging, and NT-Xent contrastive learning. Key contributions include a DAE pretraining scheme, LE-based chaos labeling via shallow PLRNNs, a shared encoder with three task heads, and comprehensive ablations showing synergistic gains across two benchmark datasets. The approach yields improved robustness and generalization, surpassing several state-of-the-art EEG MI methods and illustrating the value of combining denoising, dynamical signatures, and self-supervised learning for noisy neural signals.

Abstract

We introduce a two-stage multitask learning framework for analyzing Electroencephalography (EEG) signals that integrates denoising, dynamical modeling, and representation learning. In the first stage, a denoising autoencoder is trained to suppress artifacts and stabilize temporal dynamics, providing robust signal representations. In the second stage, a multitask architecture processes these denoised signals to achieve three objectives: motor imagery classification, chaotic versus non-chaotic regime discrimination using Lyapunov exponent-based labels, and self-supervised contrastive representation learning with NT-Xent loss. A convolutional backbone combined with a Transformer encoder captures spatial-temporal structure, while the dynamical task encourages sensitivity to nonlinear brain dynamics. This staged design mitigates interference between reconstruction and discriminative goals, improves stability across datasets, and supports reproducible training by clearly separating noise reduction from higher-level feature learning. Empirical studies show that our framework not only enhances robustness and generalization but also surpasses strong baselines and recent state-of-the-art methods in EEG decoding, highlighting the effectiveness of combining denoising, dynamical features, and self-supervised learning.
Paper Structure (51 sections, 32 equations, 6 figures, 10 tables, 2 algorithms)

This paper contains 51 sections, 32 equations, 6 figures, 10 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related work
  3. Mathematical preliminaries
  4. Lyapunov exponents
  5. Methodology
  6. Denoising autoencoder for EEG signal preprocessing
  7. Multitask Transformer architecture
  8. Optimization and training procedure of semi-supervised framework
  9. Computational experiments
  10. Training protocol and results
  11. Results from the Ablation Study
  12. Comparison of State-of-the-Art (SOTA) Models: results with two datasets
  13. Conclusions
  14. Appendix
  15. Data
  16. ...and 36 more sections

Figures (6)

  • Figure 1: Schematic illustration of our multitask learning framework architecture for noisy EEG brain signals.
  • Figure 2: Raw and denoised EEG signals for channel 32 (subject S001R04) using our proposed DAE. The DAE effectively suppresses high-frequency and non-physiological noise while preserving task-relevant spectral features, as demonstrated by the PSD plot.
  • Figure 3: Multidomain visualization of EEG signal properties for the first subject (S001R04). (A) Autocorrelation plots across all 64 EEG channels, revealing temporal dependencies in the signal. (B) Signal-to-Noise Ratio (SNR) enhancement after filtering (1–40 Hz), highlighting boosted channel quality. (C) Power Spectral Density (PSD) highlighting the frequency content with annotated canonical EEG bands (delta, theta, alpha, beta, gamma). (D) Mean PSD across all channels. (E) Band-specific EEG power (extracted for each canonical band) across channels. (F) Continuous wavelet transform of the FC5 channel, illustrating time-frequency dynamics. (G) Summary of descriptive statistics (mean, median, variance, standard deviation) across EEG channels. (H) Mean EEG amplitude computed per channel to assess overall signal strength distribution.
  • Figure 4: Mean PSD Across 64 EEG Channels (in dB). The plot shows the mean PSD computed across all 64 EEG channels, comparing the original signal (gray line) and a boosted SNR version (blue line). The x-axis represents frequency (0–80 Hz), while the y-axis shows PSD in decibels (dB). Colored vertical bands denote canonical EEG frequency ranges: Delta (0.5–4 Hz), Theta (4–8 Hz), Alpha (8–13 Hz), Beta (13–30 Hz), Gamma (30–80 Hz).
  • Figure 5: Autocorrelation of EEG band power across frequency bands. Autocorrelation plots of normalized EEG band power are shown for five canonical frequency bands across multiple EEG channels or subjects:(A) Delta (0.5–4 Hz), (B) Theta (4–8 Hz), (C) Alpha (8–13 Hz), (D) Beta (13–30 Hz), and (E) Gamma (30–70 Hz). Each line represents the autocorrelation of band power time series for a single channel/subject, computed over a ±1 second lag window. The y-axis shows the normalized autocorrelation, and the x-axis represents time lag (in seconds).
  • ...and 1 more figures