Voxelwise encoding models with non-spherical multivariate normal priors.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Predictive modeling for neural and fMRI data commonly utilizes regression techniques with parameter priors.
• Ridge regression, a popular method, assumes a spherical multivariate normal prior, implying equal and independent variance for all parameters.
• This spherical prior assumption may not always be optimal or suitable for capturing complex parameter structures.

Purpose of the Study:

• To explore the theoretical underpinnings of Tikhonov regression for neuroscience applications.
• To present an efficient computational approach for implementing Tikhonov regression.
• To demonstrate the advantages of Tikhonov regression in improving predictive models for fMRI data.

Main Methods:

• Discussed the theoretical framework of Tikhonov regression, a generalized form of ridge regression.
• Developed and demonstrated a computationally efficient method for applying Tikhonov regression.
• Applied Tikhonov regression to predictive modeling of fMRI data, comparing its performance to standard methods.

Main Results:

• Showcased several examples where Tikhonov regression significantly enhanced predictive models for fMRI data.
• Demonstrated that prior studies implicitly used Tikhonov regression through regressor transformations in ridge regression.
• Highlighted the improved predictive performance achievable with non-spherical priors in Tikhonov regression.

Conclusions:

• Tikhonov regression provides a more flexible and powerful alternative to standard ridge regression for neuroimaging data analysis.
• The proposed computational methods facilitate the practical application of Tikhonov regression in neuroscience research.
• Incorporating structured priors via Tikhonov regression can lead to more accurate and robust predictive models for fMRI.