Representational similarity encoding for fMRI: Pattern-based synthesis to predict brain activity using...

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Neuroscience

Background:

• Understanding neural activity patterns elicited by stimuli is a key challenge in neuroscience.
• Existing methods like stimulus-model-based-encoding and Representational Similarity Analysis (RSA) have limitations.
• Stimulus-model-based-encoding synthesizes neural activity but requires model fitting, risking overfitting.
• RSA assesses models by comparing similarity structures but cannot synthesize neural activity.

Purpose of the Study:

• To introduce a novel approach, representational similarity-encoding, that combines the strengths of stimulus-model-based-encoding and RSA.
• To enable robust stimulus-model-based neural encoding without the need for model fitting.
• To integrate encoding analyses within the Representational Similarity Analysis framework.

Main Methods:

• Developed representational similarity-encoding, a new method integrating stimulus-model encoding and RSA.
• This approach bypasses model fitting, thus avoiding overfitting issues and reducing computational cost.
• Applied the method to synthesize and decode functional magnetic resonance imaging (fMRI) patterns representing word meanings.

Main Results:

• The new approach successfully synthesizes and decodes fMRI patterns related to word meanings.
• Representational similarity-encoding robustly enables stimulus-model-based neural encoding.
• The method integrates encoding models and RSA, capturing the strengths of both.

Conclusions:

• Representational similarity-encoding offers a powerful, computationally efficient tool for analyzing neural representations.
• This unified approach enhances the understanding of how the brain encodes and represents information, with potential relevance to semantic memory.
• The method facilitates the synthesis of predicted neural activity patterns based on representational similarities.