Practices and pitfalls in inferring neural representations.

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience
• Neuroimaging Analysis

Background:

• Inferring neural representational schemes from stimulus features is crucial in cognitive neuroscience.
• Stimulus-feature-based encoding models are widely used for this purpose.

Purpose of the Study:

• To investigate the validity of inferring neural representational spaces using stimulus-feature-based encoding models.
• To demonstrate that high prediction accuracy does not equate to accurate representational inference.

Main Methods:

• Conducted three simulations to test encoding model predictions.
• Analyzed prediction accuracy and deviations from ground-truth representations.
• Examined scenarios with differing representational geometries and dimensions.

Main Results:

• Achieved high prediction accuracy even when neural and stimulus representations had different geometries and dimensions.
• Identified systematic deviations between model predictions and ground-truth representations.
• Demonstrated that successful prediction is insufficient for valid representational inference.

Conclusions:

• Caution is advised when inferring neural codes solely based on high prediction accuracy from encoding models.
• Alternative methods like model comparison and visualization are needed to overcome inferential limitations.
• Understanding the limitations of current encoding models is vital for advancing cognitive neuroscience research.