Eitanm/ddm-st-lans

DDM non-decision-time variability LANs (ddm_uniform_st, ddm_normal_st)

Two likelihood approximation networks for the drift diffusion model with trial-to-trial variability in non-decision time, with executed tutorials to follow once they are re-run under these names (they live on the labmate/st-networks branch of EItanm1999/HSSM meanwhile).

model	trial non-decision time	st means
ddm_uniform_st	Uniform(t - st, t + st)	half-width, SD = st / sqrt(3)
ddm_normal_st	Normal(t, st), unbounded	standard deviation

Equal st across the two is not an equal amount of variability.

Contents

ddm_uniform_st.onnx, ddm_normal_st.onnx   the networks, 85 KB each
lan/<model>/                              training bundle: jax state, configs, history
manifest.json                             file inventory

NOTE: the tutorials are re-added here once they are re-executed under the new model names.

Getting a network

from huggingface_hub import hf_hub_download
onnx = hf_hub_download(repo_id="Eitanm/ddm-st-lans", filename="ddm_uniform_st.onnx")

Fitting

These models are not in a released HSSM. Install the branch that registers them and carries the st-aware admissibility guard, then pass the network file explicitly:

pip install "hssm @ git+https://github.com/EItanm1999/HSSM.git@labmate/st-networks"

import hssm
m = hssm.HSSM(data=df, model="ddm_uniform_st", loglik=onnx,
              loglik_kind="approx_differentiable",
              model_config={"bounds": {"v": (-2.0, 2.0), "a": (0.5, 2.0), "z": (0.35, 0.65),
                                       "t": (0.25, 1.5), "st": (0.01, 0.25)}},
              link_settings="log_logit", initval_jitter=0)

Two things that bite, both covered at length in the tutorials on that branch. Against a stock HSSM release the admissibility guard floors the band where these kernels put density, which biases t and never raises an error. And st should be bounded below at 0.01 rather than the trained edge of 0.001, because below 0.01 the network fabricates a maximum on the bound.

Each model card under lan/ carries its own training box, architecture and caveats.

Status

Interim location. These networks belong in franklab/HSSM alongside the other HSSM LANs, and move there once the model registrations are merged upstream. Until then they live here so the tutorials run without a lab filesystem.