Disentangling signal and noise in neural responses through generative modeling

Area of Science:

• Neuroscience
• Computational Neuroscience
• Data Analysis

Background:

• Neural responses to repeated stimuli show significant variability, termed noise.
• Distinguishing neural signal from noise is crucial for understanding brain function and avoiding misinterpretation.
• Current methods may not adequately separate signal and noise components.

Purpose of the Study:

• To introduce a principled modeling approach, Generative Modeling of Signal and Noise (GSN), for disentangling neural signal and noise.
• To improve the estimation of the signal distribution in neural response data.
• To demonstrate the utility of GSN in enhancing data analysis techniques like principal components analysis.

Main Methods:

• Developed Generative Modeling of Signal and Noise (GSN), a method modeling response measurements as sums of signal and noise distributions.
• Estimated the signal distribution by subtracting the noise distribution from the data distribution.
• Validated GSN using simulations and applied it to functional magnetic resonance imaging (fMRI) data.

Main Results:

• GSN effectively improves estimates of the signal distribution, but not individual event responses.
• GSN favorably compares to related methods in simulations.
• Application to fMRI data showed GSN denoises principal components analysis and improves dimensionality estimates.

Conclusions:

• GSN provides a robust framework for characterizing and separating signal and noise in neural responses.
• The method has implications for improving computational models of neural activity and estimating noise ceilings.
• GSN offers a valuable tool for neuroscientists, with available code in MATLAB and Python.