A generalized linear model for estimating spectrotemporal receptive fields from responses to natural sounds

Area of Science:

• Neuroscience
• Auditory System Research
• Computational Neuroscience

Background:

• Single auditory neuron stimulus-response properties are often described by spectrotemporal receptive fields (STRFs).
• Estimating STRFs from natural stimuli involves various algorithms with differing models, cost functions, and regularization methods.

Purpose of the Study:

• To characterize auditory neuron stimulus-response functions using a generalized linear model (GLM).
• To compare the predictive power and STRF properties of GLM against normalized reverse correlation (NRC).

Main Methods:

• Utilized a GLM incorporating stimulus (STRF) and post-spike filters to model neural responses.
• Fitted the GLM using maximum penalized likelihood to zebra finch auditory midbrain neuron spiking activity.
• Compared GLM performance with NRC for responses to conspecific vocalizations and modulation-limited noise.

Main Results:

• A GLM with a sparse prior significantly outperformed NRC in predicting novel neural responses to both stimulus types.
• STRFs estimated by GLM and NRC differed substantially.
• GLM-derived STRFs exhibited greater consistency across different stimulus classes compared to NRC-derived STRFs.

Conclusions:

• A GLM with a sparse prior offers a more accurate characterization of spectrotemporal tuning in auditory neurons, especially for complex sounds.
• The GLM approach improves the prediction of spike train responses to novel stimuli.