Concept2Brain: An AI model for predicting subject-level neurophysiological responses to text and pictures

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Neuroscience

Background:

• Advancements in artificial intelligence (AI) offer new methods for analyzing complex neurophysiological data.
• AI enables the prediction and reproduction of brain responses to various stimuli.
• Understanding how the brain processes semantic and emotional information is a key challenge.

Purpose of the Study:

• To introduce the Concept2Brain model, a deep network architecture for generating synthetic electrophysiological responses.
• To demonstrate the model's ability to translate semantic/emotional information from images or text into neural signals.
• To provide an open-source tool for creating reproducible electroencephalography (EEG) datasets.

Main Methods:

• Utilized AI solutions, including OpenAI's CLIP, to create representations of input data (images/text).
• Developed a deep network architecture to map these representations into an electrophysiological latent space.
• Trained and validated the model using existing neurophysiological datasets.

Main Results:

• The Concept2Brain model successfully generated synthetic neural responses that closely mimic those observed in naturalistic scene perception studies.
• The model demonstrated the capability to predict brain responses to diverse semantic concepts and visual stimuli.
• The generated synthetic EEG data showed high similarity to real-world neurophysiological recordings.

Conclusions:

• The Concept2Brain model serves as a valuable, openly available resource for neuroscience research.
• This AI-driven tool facilitates the creation of reproducible EEG datasets and enables prediction of brain responses.
• The model advances AI-driven approaches to brain activity modeling and the study of neural representations.