Modeling and Interpreting Correlations, Null Distributions and Significance Levels in Neural Tracking of Natural Stimuli

What the math gives to ML

The paper provides a practical way to distinguish genuine predictive or tracking ability from correlation that is merely easy to obtain because of signal statistics. Its transferable asset is a semi-parametric null model: Fisher-transformed correlations are approximately zero-mean Gaussian, while their variance can be estimated from a modest number of misalignment permutations and rescaled across evaluation-window lengths. This can turn raw correlation into a calibrated score for temporal neural networks, representation models, and feature or architecture selection. The most useful adaptation is to use null-normalized correlation for validation, early stopping, model comparison, and optionally as a training regularizer rather than optimizing or reporting raw correlation alone.