File size: 2,195 Bytes
f406c82 94711f6 4c179c0 94711f6 f406c82 94711f6 f406c82 94711f6 f406c82 4c179c0 f406c82 94711f6 f406c82 4c179c0 f406c82 4c179c0 f406c82 94711f6 f406c82 94711f6 f406c82 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 | {
"model_name": "hierarchical_gmm_30_types_300_subtypes",
"release_tag": "2026_W45",
"data_vintage": "2026_W41",
"architecture": {
"modalities": [
"waveform",
"acg",
"stpc"
],
"modality_latent_dim": 20,
"joint_latent_dim": 60,
"density": "hierarchical Gaussian mixture, full covariances",
"n_types": 30,
"n_subtypes": 300,
"max_subtypes_per_type": 10,
"min_units_per_subtype": 20,
"assignment": "top-down: P(k, j | z) = P(k | z) P(j | k, z), P(j | k, z) among the sub-types of k",
"context_conditioning": "sub-type weights gamma_k(x) gamma_j|k(x), from two context networks",
"context_definition": "50 MERFISH PCs + 50 AGEA PCs",
"phenotype_readout": "members: each sub-type's TRAIN units from anywhere in the brain, weighted by their top-down posterior of the sub-type"
},
"split_counts": {
"train_units": 70564,
"validation_units": 11052,
"test_units": 20206,
"train_pids": 724,
"validation_pids": 96,
"test_pids": 198
},
"heldout_basic_test": {
"model": "hierarchical_gmm_30_types_300_subtypes",
"split": "held-out test PIDs",
"reconstruction": {
"waveform_mse": 0.0008023193222470582,
"acg_mse": 2.9368129617068917e-05,
"stpc_mse": 0.010675105266273022,
"n_test_units": 4096
},
"latent_nll": {
"context_weights": -29.075083474935514,
"brain_wide_weights": -28.435720701046566,
"context_improvement": 0.6393627738889478,
"n_test_units": 20206,
"lower_is_better": true
},
"sanity_checks": {
"all_metrics_finite": true,
"context_beats_brain_wide": true,
"subtypes_inside_types": true
}
},
"invariants": [
"TEST PIDs are never used for AE/mixture/context training or as readout members.",
"The type and sub-type means and covariance matrices are global.",
"Molecular context changes only the sub-type weights gamma_k(x) gamma_j|k(x).",
"A unit's sub-type always lies inside its type (top-down assignment).",
"All spatial modeling uses the configured single-hemisphere mirroring convention."
]
}
|