TLDR: FDC-Net is a new deep learning framework that improves EEG-based emotion recognition by deeply integrating artifact removal and emotion classification. Unlike traditional methods that treat these as separate steps, FDC-Net uses a feedback mechanism to allow the tasks to learn collaboratively, ensuring that emotion-relevant information isn’t lost during denoising. This results in significantly more accurate and robust emotion recognition, even in very noisy real-world EEG recordings.
Electroencephalogram (EEG) signals are crucial for understanding human emotions, playing a vital role in fields like affective computing and brain-computer interfaces. However, real-world EEG recordings are often contaminated by various physiological artifacts, such as muscle movements (EMG) and eye blinks (EOG). These interferences significantly reduce the quality of the signals, making accurate emotion recognition a challenging task.
Traditionally, researchers have tackled this problem by treating artifact removal (denoising) and emotion recognition as two separate, sequential steps. First, the EEG data is cleaned, and then, the ‘denoised’ data is fed into an emotion recognition model. While seemingly logical, this cascaded approach has significant drawbacks. It can lead to an accumulation of errors from one stage to the next, and more importantly, it fails to leverage any potential synergies between the two tasks. Often, denoising methods might inadvertently remove valuable emotion-related information along with the noise, especially if they operate under the assumption of ‘perfectly clean data’.
Introducing FDC-Net: A Collaborative Solution
To overcome these limitations, a novel framework called FDC-Net, short for Feedback-Driven Collaborative Network for Denoising-Classification Nexus, has been proposed. This innovative approach deeply couples the denoising and emotion recognition tasks, allowing them to work together in an end-to-end manner, leading to more robust emotion recognition even in noisy environments. The core idea behind FDC-Net is to establish a dynamic collaborative mechanism between artifact removal and emotion recognition.
The primary innovations of FDC-Net include: (1) a bidirectional gradient propagation with joint optimization strategies, meaning that both tasks influence each other’s learning process; and (2) a gated attention mechanism integrated with a frequency-adaptive Transformer, which uses learnable band-position encoding to better process EEG signals.
How FDC-Net Works
FDC-Net comprises two main modules: a denoising module and a classification module, which interact through a shared layer and a feedback mechanism. The denoising module is responsible for cleaning the noisy EEG signals. It incorporates a specialized EEG-specific Transformer (EEGSPTransformer) that is designed to handle the unique characteristics of EEG data. This includes a ‘Band-Limited Positional Encoding’ that focuses on relevant EEG frequency bands (4-45Hz) and ‘Channel-Aware Dynamic Gating’ to understand spatial relationships between brain regions. It also uses ‘Time-Frequency Attention’ to capture both transient and rhythmic activities in the signal.
The classification module then takes the features, enhanced by channel attention, to predict emotion labels (e.g., valence and arousal). What makes FDC-Net unique is the ‘Dual-Path Feature Interaction with Feedback’. Unlike traditional methods where denoising and classification are independent, FDC-Net allows the emotion recognition task to provide feedback to the denoising module. This feedback guides the denoising process to retain features that are crucial for emotion recognition, preventing the accidental removal of important emotional information. This collaborative optimization ensures that both tasks mutually promote and improve each other’s performance.
The entire system is trained jointly, with a combined loss function that balances the goals of both denoising and classification. This joint training, along with a curriculum learning strategy that gradually introduces more noise, enhances the model’s robustness.
Impressive Performance in Noisy Environments
The FDC-Net framework was rigorously evaluated on two widely used EEG-based emotion datasets: DEAP and DREAMER. These datasets contain multi-dimensional emotional labels, allowing for comprehensive testing. The experiments simulated real-world noise by adding a mix of EMG and EOG signals at various signal-to-noise ratios (SNR).
The results demonstrated FDC-Net’s superior performance compared to nine state-of-the-art methods. In terms of denoising, FDC-Net achieved high correlation coefficients (CC) and low mean squared errors (MSE), indicating its effectiveness in cleaning signals. More importantly, under physiological artifact interference, FDC-Net achieved significantly higher emotion recognition accuracies (82.3% on DEAP and 88.1% on DREAMER) than other methods, especially in very noisy conditions. This highlights its robustness and ability to maintain performance where other models falter.
While integrating a Transformer architecture means FDC-Net has a relatively larger parameter size than some simpler models, its optimized design ensures fast running speeds, making it suitable for practical applications requiring real-time processing. Ablation studies further confirmed that the collaborative mechanism and the EEGSPTransformer module are critical components contributing to FDC-Net’s enhanced performance.
Also Read:
- IMAC: Enhancing EEG Signal Classification Through Spatial Imputation
- AnnoSense: A New Framework for Capturing Real-World Emotions for AI
A New Direction for EEG Analysis
FDC-Net offers a compelling new paradigm for EEG analysis, particularly in environments where noise is a significant factor. By deeply coupling denoising and emotion recognition, it addresses the long-standing challenge of error accumulation and the loss of emotion-relevant data. This collaborative framework not only enhances both artifact suppression and emotion recognition but also opens doors for its extension to other physiological signal processing fields and downstream tasks like clinical mood disorder diagnosis and cognitive load prediction. For more details, you can refer to the full research paper: FDC-Net: Rethinking the association between EEG artifact removal and multi-dimensional affective computing.
