EEG-Based Decoding of Color and Visual Category Representations Is Reliable Within and Across Sessions

Technical Summary: EEG-Based Decoding of Color and Visual Category Representations

Problem Statement
Traditional methods for studying human visual representations, such as Event-Related Potentials (ERPs), have successfully identified distinctions in brain activity elicited by different stimulus categories. However, these univariate approaches fail to capture information embedded in the spatial distributions (patterns) of brain activity. Furthermore, traditional methods often lack the sensitivity required to analyze visual representations at the single-participant or single-trial level. While Time-resolved Multivariate Pattern Classification Analysis (MVPA), or "Decoding," has emerged as a popular solution for extracting visual representations from EEG topography without a-priori assumptions regarding spatiotemporal effects, significant uncertainties remain. Specifically, it is unclear whether decoding accuracy is consistent within individual participants across multiple sessions, tasks, and attentional conditions, or if it varies significantly across different visual features. Additionally, the stability of the underlying spatial patterns (classification weights) across these variables has not been fully established.

Methodology
To address these gaps, the study employed a rigorous experimental design involving human participants performing three distinct visual tasks across two separate sessions (spaced 1–7 days apart). The researchers utilized Time-resolved MVPA to decode EEG data, focusing on two primary visual features: color and visual category.

The analysis focused on several key dimensions of reliability and generalization:

1. Correlation Analysis: The study examined correlations in decoding accuracy across multiple axes: within the cross-validation set, between the two sessions, between the different visual features (color vs. category), and against standard ERP signal measurements and behavioral performance metrics.
2. Generalization Testing: The researchers tested how well the trained decoding models generalized to different tasks and varying attentional conditions.
3. Spatial Pattern Analysis: The study analyzed the classification weights (spatial patterns) to determine if the neural signatures underlying the decoding remained stable across sessions, tasks, and attention conditions.

Key Results
The study yielded several critical findings regarding the stability and nature of EEG-based decoding:

• Inter-Participant Variability: Decoding accuracies varied substantially across different participants, indicating that the ability to decode visual representations is not uniform across the population.
• Intra-Participant Reliability: Despite inter-participant variability, decoding accuracy was found to be highly reliable within individual participants. This reliability held true across different sessions, attentional conditions, and tasks.
• Trait-Like Decodability: The consistency of decoding performance within individuals suggests that "decodability" behaves as an individual trait.
• Generalization of Spatial Patterns: The spatial patterns underlying the decoding (classification weights) successfully generalized across different tasks, attentional conditions, and sessions. This indicates minimal representational drift at the spatial resolution provided by EEG.

Significance and Claims
The paper concludes that EEG-based decoding is a robust and reliable method for studying visual representations. The primary significance of the work lies in establishing that visual representations, as captured by EEG, are stable over time and across varying cognitive contexts (tasks and attention). By demonstrating that the spatial patterns of brain activity generalize across conditions and that decoding accuracy is a stable individual characteristic, the study validates the use of MVPA for single-participant analyses. The findings suggest that the visual system's representation of stimuli, specifically regarding color and category, remains consistent enough to be reliably decoded across multiple experimental sessions, reinforcing the stability of these neural representations.