Deep Learning Emotion Classification via Attention Masking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing emotion recognition systems face limitations in generalization due to dependence on hand-crafted features, class imbalanced data, and inefficiencies in training with laboratory-generated data that does not accurately represent real-world emotional stimuli, leading to poor classification performance.
Innovation Solution
An apparatus for training a deep learning model using an attention layer that masks input matrices based on channel-wise and feature-wise statistics, combined with inter-class and intra-class loss functions to improve emotion classification performance, particularly for biosignals like EEG and ECG, by focusing on important features and reducing variance within classes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If hand-crafted features are used for emotion recognition, then the system can operate with simple feature extraction, but the generalization performance deteriorates due to dependence on specific feature sets
Solution Approach 1:
The patent replaces hand-crafted feature extraction (mechanical system) with a deep learning model that automatically learns features from raw biosignal data. The convolutional neural network performs automatic feature learning, eliminating the need for manual feature engineering while improving generalization performance across different datasets and subjects.
2Ease of manufacture
If standard stimulus databases are used for training, then the training process is simplified, but the classification performance deteriorates due to limited sample size and artificial stimulation conditions
Solution Approach 1:
The patent applies data augmentation techniques as a preliminary action to expand the training dataset before model training. By generating synthetic variations of existing biosignal data through noise addition, time warping, and other transformations, the system increases sample diversity and size, improving classification accuracy while maintaining training efficiency.
3Adaptability or versatility
If class imbalanced data is used for training, then the model learns from real-world distribution, but the classification performance deteriorates due to bias toward majority classes
Solution Approach 1:
The patent applies focal loss function as a partial action to address class imbalance. The focal loss modifies the standard cross-entropy loss by down-weighting easy examples (majority class) and up-weighting hard examples (minority class), allowing the model to focus training effort on underrepresented classes while maintaining overall performance.
4Loss of information
If biosignals from multiple channels are processed, then more emotion-related information is captured, but the computational complexity increases
Solution Approach 1:
The patent applies channel-wise attention mechanism that segments the multi-channel biosignal processing into independent attention computations for each channel. The attention layer computes importance weights for each channel separately, allowing the model to selectively focus on informative channels while suppressing redundant ones, thereby reducing computational complexity while preserving emotion-related information.
Data Source
AI summary
An apparatus for training a deep learning model for classifying emotions from biosignals includes: a memory configured to store a program for training the deep learning model; and a processor configured to train the deep learning model by executing the program, wherein, when the processor executes the program, the processor inputs an input matrix to an attention layer constituting the deep learning model, the input matrix being composed of a plurality of features each mapped to a plurality of channels and a plurality of feature groups as the biosignals are acquired from a plurality of channels and the biosignals acquired from each channel are divided into the plurality of feature groups, and the attention layer operates to mask the input matrix using an attention matrix in which an importance of features in each channel is reflected.