Model-based targeted dimensionality reduction for neuronal population data.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Data Analysis

Background:

• Analyzing high-dimensional neuronal population activity often requires summarizing data using fewer parameters.
• Targeted dimensionality reduction methods identify distinct low-dimensional subspaces related to experimental variables.
• Existing methods lack probabilistic rigor, limiting flexibility and interpretability.

Purpose of the Study:

• To develop a model-based method for targeted dimensionality reduction using probabilistic generative models.
• To improve the accuracy and interpretability of analyzing neuronal population activity.
• To offer a flexible and robust alternative to existing ad hoc methods.

Main Methods:

• Proposed a probabilistic generative model for population response data.
• Utilized a low-rank factorization of a linear regression model for low-dimensional structure.
• Employed expectation maximization and direct marginal likelihood maximization for efficient inference.
• Developed an efficient method for estimating subspace dimensionality.

Main Results:

• The proposed method demonstrated superior performance in parameter estimation (mean squared error) compared to alternatives.
• Successfully identified the correct dimensionality of neural encoding in simulated data.
• Provided more accurate inference of low-dimensional subspaces than demixed PCA.

Conclusions:

• The novel model-based approach offers a more interpretable and flexible framework for targeted dimensionality reduction.
• This method advances the analysis of neuronal population activity by integrating probabilistic modeling.
• The findings suggest improved accuracy and reliability in uncovering neural computation structures.