flâneur

Bits per Spike as a Betting Game · neurostatsblog

neurostatsblog.github.io · 5,473 words · saved by 1 readers

Whenever we fit a model to neural or behavioral data, we need to benchmark it against simpler or well-known baselines. Typically this is done by reporting the difference in log-likelihoods on heldout data. For example, the popular “bits per spike” performance metric (see e.g. Pillow et al. 2008) is simply the log (base 2) likelihood of the model minus the log (base 2) likelihood of a homogeneous Poisson process (or another appropriate baseline model), divided by the total number of spikes in the dataset.

Revision history v1 (2026-05-27): Initial version. v2 (2026-05-30): Added time to significance. Cite this post Copy @misc{williams2026_model, author = {Alex H Williams}, title = {Bits per Spike as a Betting Game}, year = {2026}, howpublished = {\url{https://doi.org/10.5281/zenodo.20418560}}, doi = {10.5281/zenodo.20418560}, note = {neurostatsblog}, } View as PDF Contents Basic Setup A Demo of the Betting Game Introducing the Game Choosing the optimal contract function Some interpretations Connection to Hypothesis Testing Take Home Message Further Reading Supplementary Note 1 Supplementary Note

saved by

related reading