Analyzing multicomponent receptive fields from neural responses to natural stimuli

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Machine Learning for Neuroscience

Background:

• Accurately modeling neural responses to natural stimuli requires estimating multiple, jointly acting stimulus features.
• Selectivity for feature combinations is key for properties like contrast invariance in neural processing.
• The curse of dimensionality complicates the direct estimation of these joint features.

Purpose of the Study:

• To investigate the biases of sequential feature search strategies for neural models.
• To compare the predictive power of joint versus sequential optimization of stimulus dimensions.
• To assess the feasibility of joint optimization for modeling neural responses.

Main Methods:

• Analytic arguments and simulations of model cells were used to evaluate search strategies.
• Projection pursuit regression concepts were adapted for sequential dimension searching.
• Joint information maximization was employed for estimating multiple stimulus dimensions.

Main Results:

• Sequential search strategies introduce systematic biases when applied to natural stimuli.
• Joint optimization is computationally feasible for up to three dimensions with current algorithms.
• Models with three jointly optimized dimensions outperformed sequentially optimized models in predicting V1 neuronal responses.

Conclusions:

• Joint optimization of stimulus features offers improved predictive power for neural models.
• While the curse of dimensionality persists, several relevant dimensions can be estimated via joint information maximization.
• This approach advances the development of more accurate neural response models.