Estimating nonlinear receptive fields from natural images.

Area of Science:

• Computational Neuroscience
• Visual System Modeling
• Image Processing

Background:

• Visual cell responses are nonlinear and optimized for natural images.
• Traditional Volterra models are complex and data-intensive.
• A low-dimensional subspace hypothesis aids in simplifying visual cell response modeling.

Purpose of the Study:

• To compare the Volterra model with histogram-based techniques for characterizing visual cell responses.
• To evaluate projection pursuit regression (PPR) for estimating the low-dimensional subspace.
• To demonstrate the efficacy of the Volterra relevant-space technique.

Main Methods:

• Utilized the Volterra relevant-space technique to estimate high-order Volterra models.
• Employed projection pursuit regression (PPR) to identify the relevant low-dimensional subspace.
• Compared PPR against spike-triggered average (STA) and spike-triggered covariance (STC).
• Tested models using simulated cortical simple and complex cell data, plus physiological complex cell data.

Main Results:

• Volterra models demonstrated equal or superior predictive power across all tested conditions.
• Projection pursuit regression (PPR) showed advantages over STA and STC for subspace estimation.
• The Volterra relevant-space technique effectively estimated high-order Volterra models.

Conclusions:

• The Volterra relevant-space technique is a powerful alternative to histogram-based methods for modeling visual cell responses.
• Projection pursuit regression (PPR) is a viable algorithm for recovering relevant subspaces from natural images.
• High-order Volterra models, when estimated efficiently, provide accurate characterizations of visual cell function.