Dandisets / 001616

Dandiset 001616

✓ Verified by an independent agent

SUMMER: Dataset of human single-neuron activity during an 83-minute movie from 2,286 neurons and 29 patients

electrophysiological Human 2.4 GB 29 files
  • 6 reproduced
  • 5 partially reproduced
  • 9 blocked

View on DANDI Paper JSON How to read outcomes

Claims

From the paper Decoding movie content from neuronal population activity in the human medial temporal lobe

Reproducing Gerken, Darcher et al. (2025) from dandiset 001616 16

Notebook reproduce_gerken_2025.ipynb

  • Partially reproducedResults, first paragraph; Figure 1d; Patient-wise decoding performance

    The data are 2286 single and multi-units from 29 patients (17 female), 30 to 137 units per patient, in the amygdala (580), hippocampus (794), entorhinal cortex (440), parahippocampal cortex (373) and piriform cortex (99).

    29 patients, 2286 units and 17 female patients, as stated, with 30 to 137 units per patient. With AH, MH and PH grouped as hippocampus (the dataset's code), the regions hold A 559 (paper 580), H 846 (794), EC 493 (440), PHC 293 (373) and PIC 68 (99), and 27 units lie in FF, LG and PRC. A and H are within 10%; EC, PHC and PIC are not. Every unit's region matches that of its electrode.

    How the test differed from the paper’s method

    • Regions are read from the dandiset's `brain_region` and grouped as in the dataset's code; the paper gives no mapping.
  • Partially reproducedResults, Parahippocampal cortex drives decoding; Amygdala drives decoding; Figure 5

    Decoding from single regions depends on the feature; the amygdala is best and the hippocampus worst for Summer, and the parahippocampal cortex is best and the amygdala worst for Indoor/Outdoor.

    Summer A 0.291, EC 0.149, PHC 0.096, H 0.068, so A highest and H lowest hold. Indoor/Outdoor PHC 0.357, H 0.302, A 0.122, EC 0.117, so PHC highest holds; A is not lowest, though A and EC differ by less than their SEMs (0.041, 0.052). 3 of 4 orderings hold. Six region folds have test-shuffle p above 0.05 (A Indoor/Outdoor, EC Summer, EC Indoor/Outdoor and PHC Summer once each, H Summer twice).

    How the test differed from the paper’s method

    • Logistic regression in place of the LSTM.
    • Regions as stored in the dandiset, which hold different units from the paper's (PHC 293 against 373).
  • Partially reproducedResults, Differences in character's visual appearances; Figure 5b and 5cverifier disagreed at first

    Face labels decode slightly better than character labels (Summer 0.33, Tom 0.27), and Summer Presence (0.27) is decoded best from the amygdala (0.26).

    One of three conditions holds. Summer Presence is decoded best from the amygdala (0.292; PHC 0.132, EC 0.110, H 0.103; pooled 0.234). Face labels decode worse than the character labels, Summer Faces 0.217 against Summer 0.301 and Tom Faces 0.167 against Tom 0.218 (paper LSTM 0.33 against 0.31 and 0.27 against 0.23). Every fold of the three labels is above chance (p = 0.001). The face labels keep as negatives the frames the Methods exclude, which may lower their kappa; the dandiset does not mark those frames, so this cannot be checked.

    How the test differed from the paper’s method

    • Logistic regression in place of the LSTM; the criterion was written after the first verification round, before these labels were decoded.
    • Face labels keep the frames the Methods exclude (a face neither shown nor clearly absent) as negatives.
  • ReproducedResults, Decoding from neuronal activity; Figure 3b; Methods, Single-neuron decoding performance

    Thresholding the activity of single neurons does not reliably predict whether Summer is on screen; most neurons perform near chance, below the population decoder.

    Median unit kappa for Summer 0.0076 over 2286 units; 89.8% below 0.05; best unit 0.276 (patient 7, amygdala); no unit reaches 0.31 or the pooled logistic regression (0.301).

    How the test differed from the paper’s method

    • The threshold and its direction are searched over 50 quantiles of each unit's window counts on the validation set; the Methods say only that the best threshold was chosen on the validation set.
  • ReproducedResults, Choice of architecture and Patient-wise decoding performance; Figure 3c; Figure 4a

    A logistic regression on spike counts decodes characters and location about as well as the LSTM; pooled over patients it reaches kappa 0.28 for Summer and 0.31 for Indoor/Outdoor.

    Pooled logistic regression, mean over 5 folds, Summer 0.301 (SEM 0.056), Indoor/Outdoor 0.359 (SEM 0.049), both within tolerance of 0.28 and 0.31; Tom 0.218, McKenzie 0.324 and Persons 0.410 reach the LSTM values less their SEM. Test-label shuffles give p = 0.001 in every fold for Summer, McKenzie, Persons and Indoor/Outdoor; one Tom fold has kappa 0.002 (p = 0.45). Over the C grid the test kappa ranges from 0.266 to 0.311 (Summer) and from 0.349 to 0.393 (Indoor/Outdoor). With Summer labels shifted by 300 s the kappa is 0.015.

    How the test differed from the paper’s method

    • Class weights balanced in place of oversampling the minority class.
    • C chosen on the validation kappa from 0.001, 0.003, 0.01 and 0.03; the Methods do not give the range.
    • The fold layout is that of the dataset's code with the 1.6 s window added to the 32 s gap, which gives the paper's sample counts (30822, 7194, 7194 against 30800, 7200, 7200).
    • The 0.28 for Summer is the pooled value of Figure 4; the paper does not state the Figure 3c logistic regression values in numbers.
  • ReproducedResults, Choice of architecture; Figure 3c; Table S2

    For scene and camera cuts the logistic regression performs below the LSTM, by 0.1 and 0.07, that is about kappa 0.10 and 0.11.

    Scene Cuts 0.0771 (SEM 0.027) and Camera Cuts 0.0961 (SEM 0.020), within tolerance of 0.10 and 0.11 (Camera Cuts by 0.001 above the lower bound of 0.095) and below the LSTM's 0.20 and 0.18. Camera Cuts are above chance in every fold (p = 0.001); one Scene Cuts fold is not (kappa -0.008, p = 0.89).

    How the test differed from the paper’s method

    • Scene cut labels are extended to 520 ms after the cut, as the Methods state for camera cuts.
    • The expected values are derived from the stated drops (0.1, 0.07) and the LSTM values of Table S2.
  • ReproducedResults, Avoiding spurious decoding performance; Figure 3e; Figure S9

    Without temporal gaps between training, validation and test samples, decoding of Summer is almost perfect (kappa 0.99), and it falls to 0.31 as the gap grows to 32 s.

    Summer with a random split 0.948 (5 draws, 0.942 to 0.953); with gaps of 4, 8, 16 and 32 s 0.555, 0.524, 0.451 and 0.301. The LSTM values held in the dataset's code are 0.987, 0.584, 0.539, 0.429 and 0.309.

    How the test differed from the paper’s method

    • Logistic regression in place of the LSTM.
    • Gaps other than 32 s use the dataset's code table of repeated parts per gap.
  • ReproducedResults, Patient-wise decoding performance; Figure 4

    Decoding from a single patient's neurons is lower than from the pooled population; no patient matches the pooled performance, for Summer, Scene Cuts or Indoor/Outdoor.

    No patient reaches the pooled kappa. Summer best patient 7 at 0.268 against 0.301 (median 0.060); Scene Cuts patient 20 at 0.043 against 0.077 (median 0.010); Indoor/Outdoor patient 2 at 0.256 against 0.359 (median 0.099). By the test-label shuffle 5, 5 and 12 patients are above chance (p = 0.001) in all five folds for the three labels.

  • ReproducedResults, Patient-wise decoding performance; Figure 4b and 4d

    For Summer, patients 7, 10, 15 and 22 exceed kappa 0.2; for Indoor/Outdoor, patients 2, 13, 20 and 25 exceed 0.2.

    7 of 8 named pairs exceed 0.2; Summer patients 7 (0.268), 10 (0.227), 22 (0.241), and 15 at 0.197, each above chance in all five folds; Indoor/Outdoor patients 2 (0.256), 13 (0.210), 20 (0.217), 25 (0.201). For Scene Cuts three of the paper's patients (20, 19, 2) are among the four best here, and patient 10 is fifth.

    How the test differed from the paper’s method

    • Assumes the dandiset's subject ids are the paper's patient numbers; neither gives a mapping, and the Figure 4 legend gives no unit counts per patient to compare. The agreement of the named patients supports this.
  • BlockedResults, Stimulus-aligned responsive neurons; Figure 2b; Table S1· Methods underspecified

    Over half of the parahippocampal neurons responded to camera cuts (193 of 373, 51.74%).

    The Methods' test counts 222 of 293 PHC units (75.77%) responsive at alpha 0.001 and 228 (77.82%) at 0.01 (paper 193 of 373, 51.74%). The same test counts 72.35% of PHC units and 68.4% of all units responsive at random event times (69.2% in the per-label check with 884 random times), and all synthetic Poisson units at 2 Hz with 884 random events, so its false-positive rate is far from its alpha and the procedure differs from the paper's. The test compares integer counts in 100 ms bins with the baseline mean per 100 ms across hundreds of trials, which a signed-rank test finds different even without a response. The dataset code's ZETA test (responsive_zeta_cuts.ipynb, all units) is not anticonservative (none of 400 units at random times at alpha 0.001, 1.0% at 0.01) and finds 35 of 293 PHC units (11.95%) at 0.001 and 67 (22.87%) at 0.01, below the paper's fraction; neither test as run here gives the paper's count.

    How the test differed from the paper’s method

    • The responsive-unit test follows the Methods as fixed in the claims file (baseline count over -1000 to 0 ms divided by 10, 19 overlapping 100 ms bins, two-sided Wilcoxon signed-rank, Simes); the Methods do not say more.
    • The ZETA variant uses 100 resamplings in place of the code's 1000, for time; for PHC units and scene cuts, 1000 resamplings change the counts by 1 and 2 units.
    • The fraction is compared, as the dandiset's PHC holds 293 units against the paper's 373.
  • BlockedTable S1; Results, Stimulus-aligned responsive neurons· Methods underspecified

    Camera cuts evoked responses in a large share of neurons in every region (amygdala 46.90%, entorhinal cortex 45.68%, hippocampus 38.92%).

    Depends on fig2b-phc-camera-cuts. The miscalibrated test counts A 73.70%, EC 73.63%, H 64.54% and PHC 75.77% at alpha 0.001 (paper 46.90, 45.68, 38.92, 51.74%), and 62 to 73% of each region at random times. The ZETA test of the dataset's code counts A 3.22%, EC 0.61%, H 1.30% and PHC 11.95% (responsive_zeta_cuts.ipynb).

    How the test differed from the paper’s method

    • The responsive-unit test follows the Methods as fixed in the claims file (baseline count over -1000 to 0 ms divided by 10, 19 overlapping 100 ms bins, two-sided Wilcoxon signed-rank, Simes); the Methods do not say more.
  • BlockedResults, Stimulus-aligned responsive neurons; Figure 2b; Table S1· Methods underspecified

    Few parahippocampal neurons responded to scene cuts (25 of 373, 6.70%), but the modulation was consistent across the responsive neurons (cluster permutation test, p = 0.018).

    The Methods' test counts 138 of 293 PHC units (47.10%) at alpha 0.001 and 180 (61.43%) at 0.01 (paper 6.70%); with as many random times as there are scene cut onsets it counts 48.25% of all units, so it is miscalibrated. With this split the largest cluster has p = 0.055 with the Methods' z-scoring and p = 0.020 with the code's per-trial z-scoring at alpha 0.001, and there is no cluster at 0.01. The ZETA test finds 12 of 293 (4.10%) at 0.001 and 29 (9.90%) at 0.01, either side of the tolerance of 4.16 to 9.24%; with 1000 resamplings 11 and 31 (responsive_zeta_cuts.ipynb).

    How the test differed from the paper’s method

    • The responsive-unit test follows the Methods as fixed in the claims file (baseline count over -1000 to 0 ms divided by 10, 19 overlapping 100 ms bins, two-sided Wilcoxon signed-rank, Simes); the Methods do not say more.
    • Cluster p is the fraction of 1000 label permutations whose largest cluster is at least the observed largest cluster; the Methods compare each cluster with the permuted histogram.
  • BlockedAbstract; Results, Stimulus-aligned responsive neurons; Table S1· Methods underspecified

    Only a few neurons responded to the onset of semantic features (Summer 46, Tom 216, Persons 6 and Outdoor 4 of 2286 neurons), in contrast to camera cuts (1025).

    The test counts 1013 units responsive to Summer and 1174 to Tom at alpha 0.001 (paper 46 and 216), and 9.93 to 51.36% of units for each semantic label. At random times, as many as each label's onsets, it counts about as many (Summer 44.05% against 44.31% at the real onsets, Tom 50.26% against 51.36%), so the counts are dominated by false positives. The code's onset rule (no minimum duration) changes them little (1019 and 1191). The ZETA test of the dataset's code, conservative at random times, counts 11 for Summer and 13 for Tom (responsive_zeta_labels.ipynb), and 79 for camera cuts (paper 1025).

    How the test differed from the paper’s method

    • The responsive-unit test follows the Methods as fixed in the claims file (baseline count over -1000 to 0 ms divided by 10, 19 overlapping 100 ms bins, two-sided Wilcoxon signed-rank, Simes); the Methods do not say more.
    • Onsets for labels with a duration also last at least 1 s (Results); the code does not require this, shown as a variant.
  • BlockedResults, Stimulus-aligned responsive neurons; Table S1· Methods underspecified

    No neuron responded to the onset of indoor scenes.

    The test counts 299 units responsive to Indoor onsets at alpha 0.001 and 810 at 0.01, and 439 for McKenzie and 227 for Summer Presence at 0.001, where Table S1 has 0; with 44 random times in place of the Indoor onsets it counts 11.46% of units. An absence cannot be judged with this test. The paper's test as written is the one run here; the alpha of 0.001 is that of the Results paragraph making the claim. The ZETA test of the dataset's code, which is not the paper's test, counts 2 Indoor units at 0.001 (responsive_zeta_labels.ipynb).

    How the test differed from the paper’s method

    • The responsive-unit test follows the Methods as fixed in the claims file (baseline count over -1000 to 0 ms divided by 10, 19 overlapping 100 ms bins, two-sided Wilcoxon signed-rank, Simes); the Methods do not say more.
  • BlockedResults, Stimulus-aligned responsive neurons· Methods underspecified

    By the cluster permutation test, parahippocampal neurons responded to the largest set of features (Summer, Persons, Camera Cuts, Scene Cuts, Outdoor), compared with the hippocampus (Tom, Camera Cuts, Outdoor), the amygdala (Camera Cuts) and the entorhinal cortex (none).

    The responsive and non-responsive split comes from the miscalibrated test. With it, 11 of 32 cells agree with the paper at alpha 0.001 and 12 at 0.01; camera cuts give no cluster in any region (p = 1.000), and at 0.001 PHC has 6 significant labels against 7 in H.

    How the test differed from the paper’s method

    • The responsive-unit test follows the Methods as fixed in the claims file (baseline count over -1000 to 0 ms divided by 10, 19 overlapping 100 ms bins, two-sided Wilcoxon signed-rank, Simes); the Methods do not say more.
  • BlockedResults, Decoding with non-responsive versus responsive neurons; Figure 6· Methods underspecified

    Responsive neurons drive decoding of camera cuts but not of Summer; without them, Summer is decoded about as well as from all neurons while camera cuts drop the most, and responsive neurons alone decode camera cuts about as well as all neurons but Summer less well.

    Depends on the responsive sets of the miscalibrated test (1013 units for Summer, 1616 for Camera Cuts at alpha 0.001). With them three of four conditions hold; Summer 0.294 without responsive units, 0.256 with them only (condition b needs below 0.251), 0.301 complete; Camera Cuts 0.115 with responsive units only, 0.013 without, 0.096 complete. Size-matched non-responsive sets give 0.262 to 0.300 for Summer; none can be drawn for Camera Cuts (670 non-responsive units). 548 of those 670 spike in fewer than a third of the pseudo-trials and get p = 1 untested; the 122 that spike give 0.034, a drop of 0.062, against 0.060 for the 730 active non-responsive units of Summer.

    How the test differed from the paper’s method

    • Logistic regression in place of the LSTM, over all regions together.

Top-ranked units for decoding Summer (Figure 7) 2

Notebook top_ranked_units.ipynb

  • Partially reproducedResults, Top-ranked neurons; Figure 7b; Methods, Intersection of top-performing neurons

    Across the five rankings of neurons by logistic regression weights for Summer, the top 500 neurons share 78 neurons, against about 1 expected by chance.

    95 units are common to the top 500 of the five rankings, against 78 (tolerance 70.2 to 85.8) and a chance level of 1.14 by the paper's formula (1.12 simulated), 83 times chance. No unit in a top 500 was tied at zero; ties by unit index give the same 95. Chance values for 100, 350, 750 and 1000 (0.00, 0.19, 8.69, 36.6) match the paper's (0, 0, 9, 37), so the chance computation agrees; observed overlaps are 20, 65, 156 and 252.

    How the test differed from the paper’s method

    • The four sub-splits of a test fold take each of the other four folds once as validation; the Methods say only that they alternate training and validation data.
    • The Methods average the absolute coefficients 'across the five splits' within a set of four models; the mean over the four models is used.
  • Partially reproducedResults, Training on pre-selections of top-ranked neurons; Figure 7averifier disagreed at first

    Decoding Summer from the 500 top-ranked neurons gives the peak performance (kappa 0.38), above that of the full population.

    With the logistic regression and nested averaging over the 20 sub-splits, the top 500 units give kappa 0.375 against 0.357 for all 2286 units, a difference of 0.018, less than the SEMs (0.026, 0.029). The peak is at 229 units (0.394), with 150 units at 0.389 and the top 23 at 0.341. Every fit is above chance (p = 0.001). The paper's LSTM peaks at 0.38 with 500 units.

    How the test differed from the paper’s method

    • Logistic regression in place of the LSTM; the criterion was written after the first verification round, before any fit on top-ranked units.
    • Sizes of 23, 114, 150, 229, 500, 750, 1000, 1500 and 2286 units; the paper does not list its sizes besides 500 and 150.

Not tested in a notebook 2

2 methods underspecified
  • BlockedResults, Decoding from neuronal activity; Figure 3c; Table S2· Methods underspecified

    An LSTM decodes every label above chance from the pooled population, with kappa 0.36 for Persons, 0.31 for Indoor/Outdoor, 0.31 for Summer, 0.32 for McKenzie, 0.23 for Tom, 0.20 for Scene Cuts and 0.18 for Camera Cuts.

    The LSTM's code is not in the paper's materials, the dataset's GitHub repository (commit 59fc242) or its Zenodo archive, which hold a two-layer linear baseline; the Methods choose layers, hidden sizes, batch size, learning rate, weight decay and dropout per label without stating them. Training 700 epochs per label and fold on a CPU would also exceed the time limit; this was not tried. The logistic regression results are under fig3c-logreg-characters-location and fig3c-logreg-cuts-lower.

  • BlockedResults, Choice of architecture; Figure 3d· Methods underspecified

    Shuffling the order of time bins in the test data lowers the LSTM's performance, most for visual transitions.

    Needs the trained LSTM; see fig3c-lstm-decoding. The logistic regression on window sums does not see the order of the 20 bins, so it cannot stand in.

Findings about the data

What attempts found about the dandiset itself, for its owner and the archive.

  • OtherThe units' `brain_region` gives A 559, hippocampus 846 (AH 418, MH 16, PH 412), EC 493, PHC 293 and PIC 68 units, and 27 in FF, LG and PRC, where the paper reports A 580, H 794, EC 440, PHC 373 and PIC 99 for the same 2286 units. Neither the dandiset nor the paper maps one assignment to the other, so the paper's region analyses cannot be repeated on the paper's regions.
  • File format errorSpike times in the units table are in milliseconds (its description says so), where NWB requires seconds; the watchlogs' recording times are also in milliseconds.
  • Metadata errorThe start and stop columns of `annotations_patient_aligned` are described as seconds but hold milliseconds of the recording clock, as the table's own description says.
  • Description missingThe face labels (summer-faces, tom-faces and others) hold only 0 and 1. The Methods define face labels with frames excluded when a face is neither shown nor clearly absent; those frames cannot be identified.
  • OtherPublished version 0.260519.1133 holds 57 files, the 29 of the later version and 28 older copies (`ses-1_behavior+ecephys`) with the same units and spike times, one per patient except patient 3. Version 0.260702.0824 holds each recording once.
  • Published code disagreesThe dataset's code identifies responsive units with the ZETA test (zetapy; alpha 0.01 in its notebook), where the Methods describe a bin-wise Wilcoxon signed-rank test with Simes correction. With ZETA, 79 units respond to camera cuts at alpha 0.001 (Table S1 1025), 11 to Summer (46) and 13 to Tom (216), and 11.95% of PHC units to camera cuts (51.74%).
  • Methods not availableThe responsive-unit test as the Methods describe it (spike counts in 19 bins of 100 ms against the baseline count per 100 ms, signed-rank across pseudo-trials, Simes) counts 68% of units responsive at random event times on this data with 884 events, and 10 to 55% with as many as each label's onsets, so the procedure behind Figure 2 and Table S1 cannot be recovered from the text. The ZETA test of the dataset's code does not give the paper's counts either. The paper's LSTM code and its per-label hyperparameters are not published.
  • Paper inconsistentThe threshold for responsive units is alpha 0.01 in the Methods and 0.001 in the Results, the Figure 2 legend and Table S1; the cluster test section splits units at p 0.01.
  • Paper inconsistentThe Results give the 29 patients' ages as 22 to 63; the Methods give 19 to 62 for all 46 patients recorded. The dandiset's 29 patients are 19 to 62.
  • OtherThe dandiset has 958 camera cuts in the 83-minute film, one every 5.25 s; the supplement says one every 3 s.
  • Published code disagreesThe training code of the dataset's repository (ML_framework/nwb_loading/nwb_loading.py, get_unit_ids_for_patient_nwb) compares `brain_region` with the region group name in place of its subregions; no unit is labelled `H`, so asking for the hippocampus selects no unit.
  • Identifier missingNeither the dandiset nor the paper states that subject ids are the paper's patient numbers. Taking them as such, the patients the paper names in Figure 4 are among the best here (7 of 8 above kappa 0.2).

Notebooks

Every registered entry that uses this dandiset. An entry can hold several notebooks, and can use more than one dandiset.

Reproduction Single units of the human medial temporal lobe during a full-length movie (Gerken, Darcher et al. 2025)

Tests the findings of Gerken, Darcher et al. (eLife reviewed preprint 106758, 2025) on dandiset 001616 (SUMMER), which holds 2286 units from 29 patients watching "500 Days of Summer" with 53 frame-wise annotations. It counts units responsive to the onsets of movie features, runs the cluster permutation test by region, and decodes characters, location and cuts with the paper's logistic regression, pooled, per patient, per region and on subsets of units. The paper's LSTM is not run. All files

Verification

  • Round 1
    2026-10-14    Agreed with 18 of 20 outcomes. Independent agent run by CatalystNeuro (same harness, separate context, did not write the notebook). Reviewed: The first executed notebook with the conclusion cells and the summary removed, the helper module and the claims file, with the paper's full text. The verifier listed the version's assets, and tested wilcoxon_simes, split, kappa, ranking, threshold_unit and the L1 setting of fit_logreg on synthetic inputs.
    • fig5-faces-presence: the verifier’s outcome was “Blocked”. LSTM claim. The summer-presence, summer-faces and tom-faces labels are in the files (0/1 only), so a logistic regression stand-in for the orderings could have been tested, as was done for fig5-region-decoding. Its list of problems marks this block as affecting a conclusion.
    • fig7a-500-neurons: the verifier’s outcome was “Blocked”. The Results compare the full-population value with the LSTM's 0.31, which suggests an LSTM, but the Methods describe the 20-split logistic regression. The 20 sub-split fits and rankings already exist, so a logistic regression stand-in on the top-k units was cheap to run. Its list of problems marks this block as affecting a conclusion, and it asked for kappa on the top-k units (k = 100 to 2286) with nested averaging over the 20 sub-splits.
  • Round 2
    2026-10-15    Agreed with 20 of 20 outcomes. Independent agent run by CatalystNeuro (same harness, separate context, did not write the notebook). Reviewed: The full executed notebooks of the revision (the main notebook, top_ranked_units.ipynb, responsive_zeta_cuts.ipynb and responsive_zeta_labels.ipynb) with their conclusions and summary, the helper module, the claims file, the README and this record.

Method differences

How the tests differed from the paper’s method or data, and the claims each applies to.

  • The responsive-unit test follows the Methods as fixed in the claims file (baseline count over -1000 to 0 ms divided by 10, 19 overlapping 100 ms bins, two-sided Wilcoxon signed-rank, Simes); the Methods do not say more.fig2b-phc-camera-cuts tables1-camera-cuts-regions fig2b-phc-scene-cuts tables1-few-semantic-responses tables1-no-indoor-response results-cluster-regions
  • Logistic regression in place of the LSTM.fig3e-temporal-gap fig5-region-decoding
  • Regions are read from the dandiset's `brain_region` and grouped as in the dataset's code; the paper gives no mapping.dataset-composition
  • The ZETA variant uses 100 resamplings in place of the code's 1000, for time; for PHC units and scene cuts, 1000 resamplings change the counts by 1 and 2 units.fig2b-phc-camera-cuts
  • The fraction is compared, as the dandiset's PHC holds 293 units against the paper's 373.fig2b-phc-camera-cuts
  • Cluster p is the fraction of 1000 label permutations whose largest cluster is at least the observed largest cluster; the Methods compare each cluster with the permuted histogram.fig2b-phc-scene-cuts
  • Onsets for labels with a duration also last at least 1 s (Results); the code does not require this, shown as a variant.tables1-few-semantic-responses
  • The threshold and its direction are searched over 50 quantiles of each unit's window counts on the validation set; the Methods say only that the best threshold was chosen on the validation set.fig3b-single-neuron-threshold
  • Class weights balanced in place of oversampling the minority class.fig3c-logreg-characters-location
  • C chosen on the validation kappa from 0.001, 0.003, 0.01 and 0.03; the Methods do not give the range.fig3c-logreg-characters-location
  • The fold layout is that of the dataset's code with the 1.6 s window added to the 32 s gap, which gives the paper's sample counts (30822, 7194, 7194 against 30800, 7200, 7200).fig3c-logreg-characters-location
  • The 0.28 for Summer is the pooled value of Figure 4; the paper does not state the Figure 3c logistic regression values in numbers.fig3c-logreg-characters-location
  • Scene cut labels are extended to 520 ms after the cut, as the Methods state for camera cuts.fig3c-logreg-cuts-lower
  • The expected values are derived from the stated drops (0.1, 0.07) and the LSTM values of Table S2.fig3c-logreg-cuts-lower
  • Gaps other than 32 s use the dataset's code table of repeated parts per gap.fig3e-temporal-gap
  • Assumes the dandiset's subject ids are the paper's patient numbers; neither gives a mapping, and the Figure 4 legend gives no unit counts per patient to compare. The agreement of the named patients supports this.fig4-named-patients
  • Regions as stored in the dandiset, which hold different units from the paper's (PHC 293 against 373).fig5-region-decoding
  • Logistic regression in place of the LSTM; the criterion was written after the first verification round, before these labels were decoded.fig5-faces-presence
  • Face labels keep the frames the Methods exclude (a face neither shown nor clearly absent) as negatives.fig5-faces-presence
  • Logistic regression in place of the LSTM, over all regions together.fig6-responsive-subsets
  • The four sub-splits of a test fold take each of the other four folds once as validation; the Methods say only that they alternate training and validation data.fig7b-top500-overlap
  • The Methods average the absolute coefficients 'across the five splits' within a set of four models; the mean over the four models is used.fig7b-top500-overlap
  • Logistic regression in place of the LSTM; the criterion was written after the first verification round, before any fit on top-ranked units.fig7a-500-neurons
  • Sizes of 23, 114, 150, 229, 500, 750, 1000, 1500 and 2286 units; the paper does not list its sizes besides 500 and 150.fig7a-500-neurons

Provenance

Attempt record
001616-gerken-2025-01
Paper
Decoding movie content from neuronal population activity in the human medial temporal lobe
Date
2026-10-14
Made by
An agent, run by CatalystNeuro
Dandiset version
0.260702.0824
Harness
Claude Code (GitHub Actions)
Model
claude-opus-5-5
Protocol
pilot-6

Transcripts of the agent runs

Each run’s prompt, messages and tool calls, with long output cut and the text of the paper left out, and its final report.

← Dandiset 001550Dandiset 001636 →