EEG-Based Music Emotion Prediction Using Supervised Feature Extraction for MIDI Generation.

Area of Science:

• Artificial Intelligence
• Neuroscience
• Music Information Retrieval

Background:

• Algorithmic composition is advancing with AI-driven music emotion prediction.
• Bridging neural and auditory domains for AI music generation is challenging due to semantic gaps.
• Aligning low-level brain features with high-level musical concepts requires computationally intensive methods.

Purpose of the Study:

• To propose a deep learning framework for generating MIDI sequences aligned with emotion predictions.
• To address the semantic gap between neural and auditory data in AI music composition.
• To enhance the quality and emotional relevance of AI-generated music.

Main Methods:

• Utilized EEGNet for neural (Electroencephalography) data processing and an autoencoder for auditory data.
• Incorporated Centered Kernel Alignment to manage modality heterogeneity and improve emotion separation.
• Applied domain regression to reduce subject variability and clustered latent auditory representations for better MIDI reconstruction.

Main Results:

• Improved emotion classification accuracy, specifically for arousal and valence.
• Generated MIDI sequences demonstrated enhanced temporal alignment, tonal consistency, and structural integrity.
• Subject-specific analysis indicated that stronger imagery paradigms correlate with higher-quality MIDI outputs.

Conclusions:

• The proposed deep learning framework effectively bridges neural and auditory domains for emotion-aligned music generation.
• The method enhances both the accuracy of emotion prediction and the musical quality of AI-generated MIDI.
• Individual differences in neural patterns and imagery ability significantly impact the performance of AI music composition systems.