Data-driven modelling of visual receptive fields: comparison between the generalized quadratic model and the...

Area of Science:

• Neuroscience
• Computational Neuroscience
• Visual Neuroscience

Background:

• Neurons in the primary visual cortex (V1) exhibit varied sensitivity to visual feature translations.
• Modeling V1 neural responses often involves linear spatial filters followed by nonlinear functions.
• Generalized Quadratic Models (GQM) and Nonlinear Input Models (NIM) are popular, parsimonious modeling approaches.

Purpose of the Study:

• To compare the performance of GQM and NIM in modeling V1 neural responses.
• To determine which model better captures visual feature selectivity and invariance.
• To guide the choice of data-driven receptive field modeling techniques.

Main Methods:

• Applied GQM and NIM to multielectrode recordings from cat V1.
• Used spatially white Gaussian noise as visual stimuli.
• Fitted models to 342 single units (SUs) and analyzed prediction accuracy.

Main Results:

• NIM predicted neural response rates as well as or better than GQM for 95% of SUs.
• NIM's superior performance was mainly linked to its ability to model complex nonlinearities.
• This advantage was more pronounced for units with higher average spike rates.

Conclusions:

• NIM offers a more flexible and accurate approach for modeling V1 receptive fields compared to GQM.
• The choice of model significantly impacts the ability to capture neural response characteristics.
• Findings aid researchers in selecting appropriate models for receptive field analysis.