Hierarchical models in the brain.

Area of Science:

• Machine Learning
• Computational Neuroscience
• Statistical Modeling

Background:

• Parametric models for continuous data are diverse and often complex.
• Existing models include state-space models, dynamic causal models, and general linear models.
• Inverting these models typically requires specialized schemes for each type.

Purpose of the Study:

• To present a general hierarchical model that unifies many parametric models for continuous data.
• To demonstrate a single inversion scheme applicable to this broad class of models.
• To explore the relationship between diverse generative models and their inversion.

Main Methods:

• Development of a hierarchical model with interconnected hidden layers (state-space or dynamic causal models).
• Application of dynamic expectation maximization as a universal inversion scheme.
• Formulation of the inversion process as a neural network.

Main Results:

• The proposed model encompasses a wide range of special cases, from general linear models to nonlinear time-series analysis.
• Dynamic expectation maximization successfully inverts all models within this hierarchy.
• The inversion process can be represented as a neural network.

Conclusions:

• A single, unified model and inversion scheme (dynamic expectation maximization) can handle diverse parametric models for continuous data.
• This unified approach simplifies the analysis of complex data.
• The neural network formulation of inversion offers a potential metaphor for brain function in inference and learning.