Biologically meaningful neural null models

Characterize and construct biologically meaningful null models for neural systems by specifying realistic structural properties these models should preserve and how randomness should be introduced, so that null-model comparisons of information-theoretic measures such as Partial Information Decomposition atoms are valid for brain activity analyses.

Background

The paper introduces a null-model based normalization (NuMIT) for comparing Partial Information Decomposition (PID) across systems with different mutual information, demonstrating its utility in synthetic and MEG datasets. However, the authors emphasize that the suitability of null models is crucial: null systems should meaningfully resemble the original data to provide informative baselines.

In the context of neuroscience, selecting or designing appropriate null models is particularly challenging given the complexity of brain dynamics. The authors explicitly note that biologically realistic structures for neural null models are not yet known, and that the choice of null models must be optimized to the specific system and hypothesis. Establishing such models would strengthen inference from information-theoretic analyses of neural data and improve cross-dataset comparisons.