Scalable Variational Inference for Low-Rank Spatiotemporal Receptive Fields.

Area of Science:

• Systems neuroscience
• Computational neuroscience

Background:

• Characterizing neural integration of sensory inputs across space and time is crucial in systems neuroscience.
• Linear receptive fields are common tools for quantifying neural responses but are challenging to estimate with limited data or high-dimensional, correlated stimuli.

Purpose of the Study:

• To propose a novel hierarchical model for flexible parameterization of low-rank receptive fields.
• To develop a scalable algorithm for variational Bayesian inference of receptive field components and hyperparameters.

Main Methods:

• The model incorporates Gaussian Process priors for spatial and temporal receptive field components, promoting smoothness.
• Introduced 'temporal relevance determination,' a prior imposing variable smoothness over time lags.
• Derived a scalable variational Bayesian inference algorithm suitable for high-dimensional data.

Main Results:

• The proposed estimator effectively handles high-dimensional settings where traditional methods are intractable.
• Demonstrated superior performance compared to existing estimators on neural data from rat retina and primate cortex.
• The method significantly outperforms current receptive field estimation techniques.

Conclusions:

• The developed hierarchical model and inference algorithm offer a scalable solution for estimating low-rank receptive fields.
• This approach overcomes limitations of existing methods in challenging data regimes.
• The framework has potential for extension to other high-dimensional inference problems with smooth or low-rank structures.