Introduction
The "cocktail party problem" describes the human ability to isolate a single auditory stimulus in a noisy environment. Quantifying this auditory attention is vital for cognitive research and the development of intelligent hearing aids. However, traditional analysis of electroencephalography (EEG) data is hindered by the noisy, non-stationary nature of brain signals. This study proposes a novel deep learning pipeline that treats attention estimation as a computer vision regression task by transforming 1D EEG signals into 2D images using Gramian Angular Difference Fields (GADF).
Methodology
The primary challenge in EEG analysis is capturing temporal dependencies effectively. The researchers addressed this by converting 1D time-series data into a spatial representation.
Data Representation via GADF
The core of the approach involves the Gramian Angular Difference Field (GADF). This technique represents time-series data in a polar coordinate system, preserving temporal correlations within a matrix. For this study, 14-channel EEG signals were individually converted into GADF images and then stacked to form a single 14-channel input tensor. This transformation allows the model to leverage advanced 2D and 3D convolutional neural network (CNN) architectures originally designed for image recognition.
Model Architectures
The study evaluated several deep learning configurations:
- 2D CNN: Designed to extract spatial features across the stacked channels.
- 3D CNN: Aimed at capturing deeper inter-channel relationships, though limited by computational complexity.
- Convolutional Autoencoder: Used for feature compression, with latent representations fed into Random Forest and XGBoost regressors.
A key engineering feature was the inclusion of Adaptive Max Pooling layers. Since audio stimuli varied in length, the resulting EEG images had inconsistent dimensions. Adaptive pooling allowed the network to process variable-sized inputs and compress them into a fixed-size representation for the final regression layers without losing critical neural information through padding or cropping.
Experimental Setup
The models were trained and validated using the Phyaat dataset, which includes EEG recordings from 25 subjects across 144 trials. Subjects performed tasks involving listening to audio with varying Signal-to-Noise Ratios (SNRs). The ground truth "Attention Score" was derived from the subjects' ability to correctly identify words from the audio clips.
Results and Discussion
Performance was evaluated using Mean Absolute Error (MAE), where a lower score indicates higher accuracy. The 2D CNN emerged as the superior model, achieving a validation MAE of 29.43, outperforming the previous baseline Support Vector Machine (SVM) score of 29.65.
Interestingly, the 2D CNN outperformed the 3D CNN (MAE 29.72). This suggests that for this specific dataset size and hardware environment, the ability to train a deeper 2D model was more beneficial than the theoretical advantage of a shallower 3D model in capturing channel relationships. The Autoencoder combined with XGBoost also showed competitive results with an MAE of 29.99.
Conclusion and Future Work
By reframing EEG signal processing as a computer vision problem, this research demonstrates that spatial modeling of temporal brain activity can improve auditory attention prediction. Future work will focus on larger datasets to mitigate overfitting and exploring hybrid architectures, such as CNN-LSTMs, to further refine the capture of inter-channel temporal dynamics.
Summary updated