New metric makes neural data analysis 10x more interpretable

Neuroscientists and AI researchers now have a clearer way to evaluate neural spike train models, thanks to Alex H. Williams' new interpretation of held-out log-likelihood. Published as *Bits per Spike as a Betting Game* (arXiv:2607.28779), this work reimagines the standard metric using game-theoretic statistics, framing a model as a player betting against a baseline. The Kelly betting strategy transforms log-likelihood ratios into a 'wealth process,' where the time to reach statistical significance (τ) becomes the new interpretable unit—measured in seconds of recording rather than abstract bits.

Williams demonstrates the approach on head-direction cells in mouse anterior thalamus, showing that a generalized linear model achieves significance against a Poisson baseline in as little as 120ms for highly tuned cells, versus 11 seconds for moderately tuned ones. This 'time to significance' metric provides an intuitive benchmark, directly answering the question: 'How much data do I need to trust this model?' rather than leaving researchers guessing with opaque bits-per-spike values. The method preserves the ranking of models while making their performance tangible in real-world recording times.