ePPR: a new strategy for the characterization of sensory cells from input/output data

Area of Science:

• Systems neuroscience
• Computational neuroscience
• Computational vision

Background:

• Characterizing sensory input to neural output is a core challenge in systems neuroscience.
• Previous linear-nonlinear (LN) models often require simplified stimuli and estimate only partial parameters.
• The high dimensionality of sensory inputs complicates accurate neural modeling.

Purpose of the Study:

• To develop a novel algorithm, extended Projection Pursuit Regression (ePPR), for estimating generic linear-nonlinear (LN) models.
• To enable the estimation of all spatio-temporal parameters of LN models using arbitrary stimuli.
• To demonstrate the efficacy of ePPR in modeling neural responses from simulated and physiological data.

Main Methods:

• Developed the extended Projection Pursuit Regression (ePPR) algorithm.
• Proved the theoretical capability of ePPR models to approximate continuous functions.
• Applied ePPR to recover parameters of cortical cell models and analyze physiological data from the primary visual cortex.

Main Results:

• ePPR can uniformly approximate any continuous function to arbitrary precision.
• ePPR successfully recovered parameters for cortical cell models, including those not perfectly representable by ePPR.
• ePPR effectively characterized both simple and complex cells using natural and random stimuli, outperforming spike-triggered and information-theoretic methods.

Conclusions:

• ePPR provides a powerful and flexible method for estimating comprehensive LN models of neurons.
• This is the first method demonstrated to estimate multi-filter LN models of visual cells from natural stimuli.
• ePPR advances systems neuroscience by enabling more accurate characterization of neural computations from complex sensory inputs.