Representational structure or task structure? Bias in neural representational similarity analysis and a Bayesian...

Area of Science:

• Cognitive Neuroscience
• Neuroimaging Analysis
• Machine Learning in Neuroscience

Background:

• Neural population activity patterns vary with tasks and stimuli, reflecting cognitive states.
• Current methods estimate neural patterns via linear regression before assessing similarity, introducing bias.
• This bias is pronounced in low signal-to-noise ratio (SNR) fMRI data and non-randomized tasks.

Purpose of the Study:

• To develop a novel method for unbiased estimation of neural representational similarity.
• To improve the precision of representational similarity analysis, particularly in challenging neuroimaging scenarios.
• To provide a unified framework for analyzing neural representations in fMRI data.

Main Methods:

• Introduced Bayesian Representational Similarity Analysis (BRSA), a generative modeling approach.
• BRSA treats the neural activity covariance structure as a hyper-parameter, directly estimating it from imaging data.
• Marginalizes over unknown activity patterns, enabling direct covariance estimation and joint group analysis.

Main Results:

• BRSA significantly reduces bias compared to traditional linear regression-based methods.
• Achieves higher precision in representational similarity estimation, especially at low SNR.
• Simultaneously estimates SNR maps and provides priors for improved fMRI decoding.

Conclusions:

• BRSA offers a principled and accurate approach to analyzing neural representations in fMRI.
• The method enhances the reliability of representational similarity analysis under various experimental conditions.
• BRSA provides a unified framework for neuroimaging data analysis, with publicly available tools (BrainIAK).