Cross-Subject EEG Feedback for Implicit Image Generation.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computer Vision

Background:

• Generative models create novel content but typically require explicit user input to align with human goals.
• Integrating implicit, diverse feedback from multiple users into generative models remains a significant challenge.
• Current methods lack intuitive mechanisms for users to steer content generation based on subconscious preferences.

Purpose of the Study:

• To develop a novel system for generating novel face images by inferring human goals directly from brain signals.
• To explore the integration of implicit user feedback, derived from neural data, into generative models.
• To demonstrate the feasibility of using electroencephalography (EEG) to guide generative adversarial networks (GANs) for personalized image synthesis.

Main Methods:

• Recorded electroencephalography (EEG) brain responses from 30 subjects viewing images of faces with specific salient visual features (VFs).
• Decoded subject preferences for VFs from their brain responses.
• Utilized decoded neural preferences as implicit feedback to guide a generative adversarial network (GAN) in generating new face images.

Main Results:

• The GAN successfully generated new face images guided by the decoded brain feedback.
• A follow-up user study confirmed that the generated images reflected the subjects' intended goals regarding salient VFs.
• The quality and goal-alignment of brain-feedback-generated images were comparable to those produced with manual feedback.

Conclusions:

• The study presents a first-of-its-kind system for inferring human goals from brain signals to guide generative models.
• This methodology represents a significant step towards developing 'humans-in-the-loop' systems for image generation.
• The findings open avenues for more intuitive and personalized human-AI interaction in creative content generation.