MCCal: A pre-registered multiscale count-calibration audit of data-limited neural latent models

Count-calibration rankings reverse across model families under limited neural data. We introduce MCCal, a pre-registered multiscale count-calibration audit on the Neural Latents Benchmark (NLB), comparing smoothing, GPFA, svGPFA, LFADS and NDT with within-session trial-budget subsampling, fixed held-out neurons and paired condition-clustered inference. The executed budget variant uses three deep seeds instead of the five written in H2. Neither the predicted deep-model accuracy lead at every budget (H1) nor the predicted closing of the calibration gap by recalibration (H3) was supported. Neither deep model exceeds GPFA at any evaluated budget of at most 250 available trials. H2's predicted svGPFA calibration advantage is supported in the low-budget mean against LFADS but not NDT; small svGPFA advantages remain against both at the largest budget. Validation-only negative-binomial dispersion widens the svGPFA–LFADS gap by 16.3%. Exploratory synthesis reveals metric disagreement: NDT has the best low-budget coverage error but the worst randomized PIT distance and negative mean co-bps. All primary paired effects are below one coverage percentage point, against approximately 96–97% coverage for nominal 90% intervals. One monkey, four sessions, incomplete convergence evidence and an amended reproduction gate limit generalization. MCCal provides an audited count-prediction protocol and shows that Bayesian-versus-deep comparisons and single-metric calibration rankings should not be read as universal.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24369327
Primary Topic
Neural dynamics and brain function
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

MCCal: A pre-registered multiscale count-calibration audit of data-limited neural latent models

S. M. Abdulla Al Mamun
Zenodo (CERN European Organization for Nuclear Research)
Neural dynamics and brain function
preprint

MCCal: A pre-registered multiscale count-calibration audit of data-limited neural latent models

S. M. Abdulla Al Mamun
preprint en

Abstract

Count-calibration rankings reverse across model families under limited neural data. We introduce MCCal, a pre-registered multiscale count-calibration audit on the Neural Latents Benchmark (NLB), comparing smoothing, GPFA, svGPFA, LFADS and NDT with within-session trial-budget subsampling, fixed held-out neurons and paired condition-clustered inference. The executed budget variant uses three deep seeds instead of the five written in H2. Neither the predicted deep-model accuracy lead at every budget (H1) nor the predicted closing of the calibration gap by recalibration (H3) was supported. Neither deep model exceeds GPFA at any evaluated budget of at most 250 available trials. H2's predicted svGPFA calibration advantage is supported in the low-budget mean against LFADS but not NDT; small svGPFA advantages remain against both at the largest budget. Validation-only negative-binomial dispersion widens the svGPFA–LFADS gap by 16.3%. Exploratory synthesis reveals metric disagreement: NDT has the best low-budget coverage error but the worst randomized PIT distance and negative mean co-bps. All primary paired effects are below one coverage percentage point, against approximately 96–97% coverage for nominal 90% intervals. One monkey, four sessions, incomplete convergence evidence and an amended reproduction gate limit generalization. MCCal provides an audited count-prediction protocol and shows that Bayesian-versus-deep comparisons and single-metric calibration rankings should not be read as universal.

Zenodo (CERN European Organization for Nuclear Research)
University of Manchester (GB)
Neural dynamics and brain function
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon