Cross-Subject EEG Emotion Recognition With Contrastive Latent Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing emotion recognition models based on EEG signals face challenges in cross-subject scenarios due to individual differences in skull shapes and sensitivity to stimuli, leading to a decline in performance when applied to new subjects without personalized training, and existing deep learning methods like transfer learning and domain adaptation require test set data for retraining.
Innovation Solution
A method involving constructing positive and negative samples using a generator, encoding these samples in a latent space with an autoregressive model, and training an encoder with a supervised contrastive loss function to learn domain-invariant features, followed by fine-tuning a classifier without gradient propagation, utilizing contrastive predictive coding and supervised contrastive learning to enhance cross-subject emotion recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional machine learning algorithms are used for cross-subject emotion recognition, then the model can be trained with existing data, but the performance sharply declines when applied to new subjects due to individual differences
Solution Approach 1:
The patent introduces a domain adapter module as an intermediary component that bridges the training domain and test domain. This adapter learns to map features from the source domain (training subjects) to the target domain (test subjects), effectively mediating the domain gap and enabling accurate cross-subject recognition without requiring subject-specific training models
Solution Approach 2:
The patent employs domain adaptation techniques that modify model parameters to adapt to different subject characteristics. By changing the parameter space representation and using domain-invariant features, the model can adjust its decision boundaries to work effectively across multiple subjects while maintaining high recognition accuracy
2Reliability
If domain adaptation methods are used to improve cross-subject accuracy, then recognition accuracy improves by 20%, but the model requires test set data for retraining which increases complexity
Solution Approach 1:
The patent performs preliminary domain adaptation training using only source domain data before actual inference. The domain adapter is pre-trained to learn domain-invariant representations, allowing the model to handle new subjects without requiring test set data for retraining, thus reducing overall system complexity while maintaining high accuracy
3Reliability
If contrastive learning is applied to maximize similarity of positive sample pairs, then recognition accuracy reaches 86%, but the use of large numbers of pseudo-negative samples affects final recognition accuracy
Solution Approach 1:
The patent applies local quality by treating positive and negative samples differently in the loss function. Instead of uniformly weighting all samples, the method uses label information to assign different weights to positive sample pairs (same emotion) and negative sample pairs (different emotions), allowing the model to focus on meaningful comparisons while reducing the harmful impact of pseudo-negative samples
Data Source
AI summary
A method of emotion recognition in cross-subject EEG signals, belonging to technical field of deep learning, includes the following steps: S1, constructing the extracted DE features into positive and negative samples by using a positive and negative sample generator; S2, sending the DE features of an anchor and the positive and negative samples into the encoder for coding, mapping the DE features to a latent space, performing regression prediction on the encoded anchor samples in the latent space by using an autoregressive model, training the encoder by using a probability supervision contrastive loss function; and S3, connecting the trained encoder to the classifier for fine tuning, and training the classifier through the cross entropy loss function; in this process, the encoder does not perform gradient propagation to complete cross-subject emotion recognition.