PrincetonUniversity / PsyNeuLink / 28814065630

84%

master: 85%

push

github

Prototype neural likelihood estimation against the KDE simulation likelihood

nle/: MNLE-style mixture density network over (parameters, trial
features) -> p(choice, rt), trained on simulator draws, plus an exact
Navarro-Fuss reference density. On the DDM benchmark, EM with the exact
or neural likelihood matches full-Bayes HSSM per-subject accuracy (RMSE
0.084 vs 0.086), attributing the KDE-EM gap to KDE noise rather than the
Laplace EM, at ~60x less wall time. stabflex_nle extends the test to the
stability-flexibility composition: training data generated by sweeping
theta through a worker-cached PEC (rebuilding a composition per config
leaks PNL/LLVM state until dask pauses the workers) with the sequential
gen -> train -> em pipeline in the SLURM script.

11060 of 14440 branches covered (76.59%)

Branch coverage included in aggregate %.

37807 of 43800 relevant lines covered (86.32%)

0.86 hits per line