Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Emotional EEG Classification using Upscaled Connectivity Matrices

Chae-Won Lee, Jong-Seok Lee

TL;DR

The paper addresses the challenge that CNNs applied to EEG connectivity matrices can fail to capture salient local patterns due to limited input resolution. It proposes an interpolation-based upscaling scheme (nearest-neighbor and bilinear) to expand 32×32 connectivity matrices to higher resolutions before CNN processing, exploring multiple scaling factors and kernel sizes. Using the DEAP dataset, the approach yields state-of-the-art-like improvements, with nearest-neighbor upscaling at ×3.0 and a 3×3 kernel achieving up to 82.69% accuracy, and it shows that higher resolution enhances interpretability via Grad-CAM and t-SNE analyses. The findings demonstrate that simple spatial upscaling improves both discriminative performance and feature interpretability in EEG connectivity-based emotion recognition, suggesting future work on adaptive-resolution and hybrid architectures.

Abstract

In recent studies of emotional EEG classification, connectivity matrices have been successfully employed as input to convolutional neural networks (CNNs), which can effectively consider inter-regional interaction patterns in EEG. However, we find that such an approach has a limitation that important patterns in connectivity matrices may be lost during the convolutional operations in CNNs. To resolve this issue, we propose and validate an idea to upscale the connectivity matrices to strengthen the local patterns. Experimental results demonstrate that this simple idea can significantly enhance the classification performance.

Emotional EEG Classification using Upscaled Connectivity Matrices

TL;DR

The paper addresses the challenge that CNNs applied to EEG connectivity matrices can fail to capture salient local patterns due to limited input resolution. It proposes an interpolation-based upscaling scheme (nearest-neighbor and bilinear) to expand 32×32 connectivity matrices to higher resolutions before CNN processing, exploring multiple scaling factors and kernel sizes. Using the DEAP dataset, the approach yields state-of-the-art-like improvements, with nearest-neighbor upscaling at ×3.0 and a 3×3 kernel achieving up to 82.69% accuracy, and it shows that higher resolution enhances interpretability via Grad-CAM and t-SNE analyses. The findings demonstrate that simple spatial upscaling improves both discriminative performance and feature interpretability in EEG connectivity-based emotion recognition, suggesting future work on adaptive-resolution and hybrid architectures.

Abstract

In recent studies of emotional EEG classification, connectivity matrices have been successfully employed as input to convolutional neural networks (CNNs), which can effectively consider inter-regional interaction patterns in EEG. However, we find that such an approach has a limitation that important patterns in connectivity matrices may be lost during the convolutional operations in CNNs. To resolve this issue, we propose and validate an idea to upscale the connectivity matrices to strengthen the local patterns. Experimental results demonstrate that this simple idea can significantly enhance the classification performance.

Paper Structure

This paper contains 14 sections, 1 equation, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. EEG-Based Emotion Recognition
  4. Brain Connectivity in Emotion Analysis
  5. Proposed method
  6. Connectivity Matrix Construction
  7. Limitations of Conventional Approaches
  8. Proposed Method
  9. Experimental Setup
  10. Results
  11. Overall Performance
  12. Visualization of t-SNE
  13. Visualization of CNN Activation
  14. Conclusion

Figures (5)

  • Figure 1: Activation (right) after the first convolutional layer for a connectivity matrix (left)
  • Figure 3: Overview of the proposed classification system
  • Figure 4: Classification accuracy depending on the interpolation method and kernel size across upscaling factors. The upscaling factor of 1.0 corresponds to the conventional approach.
  • Figure 5: t-SNE visualization of the feature representations extracted by models with different convolutional kernel sizes (3×3, 5×5, 7×7) under the nearest-neighbor interpolation and 3× upscaling setting. Different colors indicate different classes.
  • Figure 7: One-dimensional plots of the input and the activations of the first convolutional layer along a particular row when the kernel size is 3×3.