Abstract
The last years have seen a rise of interest in using electroencephalography-based brain computer interfacing methodology for investigating non-medical questions, beyond the purpose of communication and control. One of these novel applications is to examine how signal quality is being processed neurally, which is of particular interest for industry, besides providing neuroscientific insights. As for most behavioral experiments in the neurosciences, the assessment of a given stimulus by a subject is required. Based on an EEG study on speech quality of phonemes, we will first discuss the information contained in the neural correlate of this judgement. Typically, this is done by analyzing the data along behavioral responses/labels. However, participants in such complex experiments often guess at the threshold of perception. This leads to labels that are only partly correct, and oftentimes random, which is a problematic scenario for using supervised learning. Therefore, we propose a novel supervised-unsupervised learning scheme, which aims to differentiate true labels from random ones in a data-driven way. We show that this approach provides a more crisp view of the brain states that experimenters are looking for, besides discovering additional brain states to which the classical analysis is blind.
Introduction
In traditional Brain-Computer Interface (BCI) and neuroscience experiments, researchers rely heavily on "ground truth" labels provided by participants—for instance, a button press indicating whether a sound was heard. However, when stimuli are presented at the threshold of perception, these behavioral labels become unreliable. Participants often guess, creating systematic label noise where the reported behavior is fundamentally decoupled from the actual neural state.
Supervised learning models trained on such noisy data struggle to find clear decision boundaries. Furthermore, standard analysis is often "blind" to intermediate brain states, such as when the sensory cortex processes a signal but higher-order cognitive centers do not trigger a conscious response. This paper introduces a data-driven "supervised-unsupervised" learning framework designed to infer true neural labels without blindly trusting participant feedback.
Methodology: Sparsity-Inducing One-Class Learning
The proposed approach utilizes a two-step pipeline to refine EEG data and extract meaningful brain states:
- Artifact Removal: Unsupervised learning identifies and removes trials confounded by measurement noise, such as high-amplitude drifts caused by eye blinks or electrode movement.
- Sparsity-Inducing One-Class Learning: The core innovation is a reformulation of the One-Class Support Vector Machine (SVM). Instead of separating two labeled classes, this method identifies a "density level set" encompassing the "core" data points representing a clear neural response. By introducing an $l_1$-norm regularizer, the model becomes highly interpretable, selecting only the most relevant EEG features across specific time intervals and electrode locations.
This classification divides the data into three distinct categories: Core (clear neural signals), Plateau (borderline or subtle signals), and Outliers (noise or subliminal processing).
Experimental Setup
The framework was validated using a 64-channel EEG dataset from 11 subjects. Participants performed an "oddball" task listening to the phoneme /a/. Occasionally, the phoneme was degraded with signal-correlated noise at four levels (T1–T4), designed to achieve perception rates ranging from 100% (T1) down to 0% (T4). The raw data was reduced to 427 features by calculating the mean Event-Related Potential (ERP) signal across seven neurophysiologically relevant time intervals.
Decoding Conscious Perception
A key insight from the model is that the brain often "knows" more than the subject reports. The algorithm identified "neural hits" in trials where the subject behaviorally reported a "miss." In these instances, early sensory processing was evident in the EEG, but the late-stage cognitive component—specifically the P3 wave—was either missing or too weak to reach the threshold of conscious confidence required for a button press.
As shown in the topographic maps above, the "Core" class identified by the algorithm consistently aligns with a strong P3 component (occipital positivity occurring 400–1000ms post-stimulus). This component serves as a neural gatekeeper for conscious perception. By moving beyond binary behavioral labels, the researchers discovered a "Plateau" state where the auditory cortex is active, but the signal fails to propagate to the cognitive centers necessary for a behavioral response.
Key Results and Interpretability
The data-driven neural labels provided a significantly "crisper" view of brain states than behavioral labels alone. At the perception threshold (T2), where behavioral labels were most confounded, the algorithm successfully separated trials into those with and without conscious processing.
Furthermore, the model allowed for the quantification of label mixtures. For the most difficult stimuli (T4), nearly 100% of behavioral "hits" were assigned to the outlier or plateau classes, confirming that subjects were essentially guessing. Because the model was sparse, the weights could be mapped back to the scalp, showing that "Core" class weights perfectly matched expected bi-temporal (auditory) and occipital (cognitive) neural activity patterns.
Conclusion and Future Directions
This study serves as a proof-of-concept for using unsupervised methods to supervise the interpretation of noisy neural data. While currently limited by a small sample size and a focus on linear interpretability, the framework offers a path toward more robust BCI systems. Future work may apply this methodology to the calibration phase of real-time BCI systems, where cleaning initial label noise could significantly enhance final performance and reliability.
Summary updated