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

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.

• The paper introduces a unified two-stage framework that combines denoising autoencoders, Lyapunov-based nonlinear dynamics, and contrastive objectives to enhance EEG signal decoding.
• The methodology employs a Denoising Autoencoder followed by a Transformer-CNN to improve cross-subject generalization in low-label, noise-rich settings.
• Empirical results show improved motor imagery classification accuracy and robust chaos detection, with ablation studies highlighting the necessity of each module.

Multitask Deep Learning on Noisy EEG: Integrating Denoising, Nonlinear Dynamics, and Contrastive Objectives

Summary and Motivation

The paper "Contrastive and Multi-Task Learning on Noisy Brain Signals with Nonlinear Dynamical Signatures" (2601.08549) addresses the principal challenge of decoding electroencephalography (EEG) signals for motor imagery (MI) classification in noise-rich, real-world settings. The authors propose a unified two-stage multitask framework that explicitly integrates signal denoising, dynamical regime discrimination via Lyapunov exponents, and contrastive learning. This approach is motivated by the limitations of existing deep models in EEG analysis, which are hindered by noise susceptibility, limited interpretability with respect to nonlinear dynamics, and the lack of robust, invariant representations for downstream tasks. The framework targets both improved cross-subject generalization and enhanced dynamical characterization of EEG, with architectural modularity allowing future extensibility to diverse BCI use cases.

Methodology

The methodology is structured in two stages:

1. Denoising Autoencoder (DAE) Preprocessing:

A 1D convolutional denoising autoencoder is trained to map noisy raw EEG inputs (subject to muscle, ocular, and environmental interference) to filtered, artifact-suppressed versions. The loss combines SmoothL1 for temporal fidelity and a spectral loss to preserve EEG spectral features. This unsupervised pretraining protocol is critical for ensuring downstream models operate in a higher-SNR regime.

2. Multitask Transformer-CNN Architecture:

The preprocessed signals are propagated through a lightweight convolutional stem for local spatial-temporal encoding, followed by a Transformer encoder for long-range temporal context. The output is branched into three task-specific heads:

• Motor Imagery Classification: Supervised with cross-entropy loss, discriminates real versus imagined movement.
• Dynamics Classification: Uses Lyapunov-based labels (derived unsupervised from piecewise-linear RNN estimation and energy/entropy metrics) to classify each trial as chaotic or non-chaotic.
• Contrastive Representation Head: Implements NT-Xent loss over noise- and artifact-augmented signal pairs to promote invariance to subject/session-specific nuisance factors.

The joint loss is a weighted sum of the three objectives, and the architecture supports end-to-end semi-supervised optimization.

Nonlinear Dynamics and Lyapunov Feature Engineering

The framework distinguishes itself by incorporating explicit dynamical system characterization. Lyapunov exponents, computed via shallow piecewise-linear RNNs (shPLRNNs) trained to fit EEG time series, are used to label segments as chaotic (positive maximal exponent) or non-chaotic (negative/zero, indicating periodic/quasiperiodic behavior or absence of attractor). Complementary energy-based tagging (spectral and permutation entropy) forms an unsupervised alternate pipeline, with high agreement (Cohen's Kappa: 0.90) between approaches.

These dynamical labels serve as auxiliary supervision—guiding the feature extractor toward sensitivity to intrinsic neural complexity and nonlinear attractors, in contrast to standard discriminative pipelines that ignore the underlying system structure.

Contrastive Learning for Invariant EEG Representations

Contrastive self-supervised objectives, instantiated via NT-Xent loss, require the model to align augmented views of the same trial while separating others in latent space. Augmentations include stochastic jitter, amplitude scaling, time masking, and channel dropout, all designed to simulate realistic recording perturbations. This promotes robustness against session, subject, and sensor variability.

Empirical analysis demonstrates that this regularization is particularly beneficial in the low-label and high-noise regime, and ablation studies confirm performance deterioration if contrastive or denoising modules are omitted.

Empirical Results

Model validation spans BCI2000 and BNCI Horizon 2020 datasets, using the LOSO protocol. The proposed multitask framework shows:

• Highest F1 (0.84/0.83) across both datasets, outperforming strong CNN and Transformer baselines (including EEGNet, DeepConvNet, and TS-TCC), as well as other self-supervised and single-task variants.
• Chaos/non-chaos detection achieves high accuracy (90%+) with Lyapunov-based features alone; MI classification benefits from the multitask regime, especially with light (less aggressive) contrastive augmentation.
• Ablation studies: Removing the DAE or contrastive objective results in a substantial decrease in MI classification performance, supporting the necessity of both modules.
• The joint multitask setup avoids performance trade-offs seen when optimizing MI or dynamics in isolation.

Notably, the framework demonstrates that a minimal Lyapunov feature set (sum and maximum exponent) effectively classifies large-scale state transitions (e.g., eyes open/closed), while robust MI decoding in realistic settings demands the hybrid multitask protocol.

Theoretical and Practical Implications

Theoretical Contributions:

• The explicit integration of nonlinear time series analysis (Lyapunov exponents, attractor dimension) with deep neural architectures adds dynamical interpretability to network-based EEG decoding. This aligns signal modeling with the underlying neuroscience, bridging a gap between classical nonlinear analysis and modern ML.
• The staged denoising-then-learning architecture demonstrates empirical and theoretical gains in data regimes where non-neural noise is prevalent, supporting modular pipeline design in BCI/EEG research.

Practical Significance:

• Robust cross-subject decoding in high-noise, low-label settings has direct implications for clinical BCI systems, real-time neurofeedback, and adaptive cognitive monitoring, where calibration effort and generalizability are critical.
• The architecture is modular; additional heads or alternate backbone architectures (e.g., GNNs for spatial modeling) can be incorporated as new neuroscientific hypotheses or practical requirements emerge.

Limitations and Future Directions

The authors acknowledge several constraints: limited annotated dataset diversity, the persistent challenge of residual noise (even with DAE), and the interpretability ceiling inherent to deep models. Future work directions include:

• Extending beyond binary chaos/non-chaos to broader dynamical regimes (periodic, quasiperiodic, no attractor) for richer cognitive and clinical analysis.
• Investigating alternative or hybrid backbone architectures (e.g., incorporating GNNs to directly model EEG channel topology).
• Enhancing explainability via attention or dynamical attribution methods, critical for clinical translation.
• Modularity points toward adaptation for other EEG applications such as mental workload, seizure prediction, or longitudinal monitoring.

Conclusion

This work demonstrates that combining denoising, explicit nonlinear dynamical supervision, and contrastive representation learning within a unified multitask deep learning framework realizes significant performance and robustness gains for EEG motor imagery decoding. The approach sets a template for the principled integration of dynamical systems theory and modern ML architectures in neurological signal analysis, with implications for both theoretical neuroscience and real-world BCI deployment.