{ "dataset": { "archive": "dandi", "id": "000289", "label": "DANDI:000289" }, "page": "datasets/dandi/000289/", "generated": "2026-11-05T07:41:23+00:00", "caution": "A failure to reproduce is not a finding that the paper is wrong. It may be a failure of the attempt.", "outcome_counts": { "reproduced": 0, "partially_reproduced": 0, "not_reproduced": 0, "blocked": 1 }, "outcome_counts_with_outside_data": { "reproduced": 0, "partially_reproduced": 0, "not_reproduced": 0, "blocked": 0 }, "outcome_counts_elsewhere": { "reproduced": 7, "partially_reproduced": 4, "not_reproduced": 1, "blocked": 0 }, "outcome_counts_links": { "reproduced": 0, "partially_reproduced": 0, "not_reproduced": 0, "blocked": 0 }, "attempts": [ { "id": "000289-kutschireiter-2023-01", "dataset": { "archive": "dandi", "id": "000289", "version": "0.240801.1040", "doi": "10.48324/dandi.000289/0.240801.1040" }, "key": "dandi:000289", "label": "DANDI:000289", "dataset_version": "0.240801.1040", "paper": "10.1073/pnas.2210622120", "paper_title": null, "date": "2026-10-29", "kind": "agent", "accountable": "CatalystNeuro", "agent": { "harness": "Claude Code", "model": "claude-opus-5-5", "protocol": "pilot-12" }, "runner": { "platform": "GitHub Actions", "url": "https://github.com/catalystneuro/encore/actions/runs/40969113892" }, "tier": "agent_verified", "deprecated": "", "deprecated_on": null, "entry": "000289-kutschireiter-2023", "entry_title": "Bayesian inference in ring attractor networks (Kutschireiter et al. 2023)", "description": "Tests the findings of Kutschireiter, Basnak, Wilson and Drugowitsch (2023), a theory paper. Its simulation findings are tested on the paper's released code and processed outputs (Zenodo), its connectome finding on the hemibrain connectome, and its prediction that bump amplitude is higher with a reliable cue on dandiset 000289, which holds two-photon imaging of EPG neurons in flies shown a visual cue at three contrasts, from a separate study by the same lab.", "notebooks": [ { "path": "reproduce_kutschireiter_2023.ipynb", "title": "Testing Kutschireiter et al. (2023) with dandiset 000289", "executed": "notebooks/000289-kutschireiter-2023/reproduce_kutschireiter_2023.html" } ], "outcomes": [ { "claim": "fig1-conflict-decreases-certainty", "statement": "In the circular Kalman filter, an HD observation consistent with the current estimate increases certainty, while one that strongly conflicts with it (opposite direction) decreases certainty.", "source": "Results, Circular Kalman Filtering, HD Observations, paragraph 1; Fig. 1 D and E", "quote": "HD observations increase the observer\u2019s certainty if they are confirmatory ... if HD observations strongly conflict with the current estimate (e.g., if they point in the opposite direction), they actually decrease certainty", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_kutschireiter_2023.ipynb", "notes": "With the code's vM_Projection (kappa_phi 1, kappa_v 1, dt 0.01), for kappa_t 1, 2 and 5 and per-step reliabilities 0.1416 (gamma_z 1/s) and 0.4529 (10/s), a confirmatory observation raises and an opposite one lowers the certainty in all 6 cases, both for the observation update alone (for example kappa_t 1, 10/s, 1.453 and 0.547) and for the full step against the step without observation (1.447 and 0.546 against 0.997). The first comparison holds for any weight alpha between 0 and 2 kappa_t, so it is a weak test; the second is the informative one.", "deviations": [], "assumed": false, "disagreed": false, "data": "elsewhere", "where": "Zenodo doi:10.5281/zenodo.7615975 (BayesianRingAttractor v1.0, commit a9db6f9), circularFiltering.py vM_Projection_Run", "proposed": null }, { "claim": "fig2h-bump-tracks-circkf", "statement": "In a simulated Bayesian ring attractor of 80 neurons, the bump position tracks HD about as accurately as the circular Kalman filter, the bump amplitude tracks the filter's certainty, and when the visual cue is removed the amplitude decays toward its baseline.", "source": "Results, Neural Network Implementation, paragraph 6; Fig. 2H", "quote": "This network exhibits a spatially localized bump whose position tracks HD, with an accuracy similar to that of the circular Kalman filter itself ( Fig. 2 H ). Meanwhile, the amplitude of the bump accurately tracks the fluctuating HD posterior certainty in the circular Kalman filter ... When the visual cue is removed, the bump amplitude decays toward baseline", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_kutschireiter_2023.ipynb", "notes": "The simulation cells of Figure2h.ipynb were run from the deposit. With rate 15 (notebook) and 10/s (Methods), the network's position differs from the circKF's by a mean of 0.016 and 0.021 rad, against a circKF error to the true HD of 0.483 and 0.491 rad. Bump amplitude and circKF certainty correlate at r 0.994 and 0.993 over the cue periods. In darkness (5 to 10 s) the amplitude falls from 7.21 to 1.22 (rate 15) and from 5.89 to 1.21 (rate 10), closer to the baseline of 1 than at its start.", "deviations": [ "The notebook calls vM_Projection_Run with the keyword kappa_y, which the function names kappa_v; the keyword was renamed (claims file header). Nothing else in the cells was changed except the rate for the second reading." ], "assumed": false, "disagreed": false, "data": "elsewhere", "where": "Zenodo doi:10.5281/zenodo.7615975 (BayesianRingAttractor v1.0, commit a9db6f9), Figure2h.ipynb with circularFiltering.py", "proposed": null }, { "claim": "fig2i-accuracy-and-certainty-vs-reliability", "statement": "The Bayesian ring attractor tracks HD with the same accuracy as the circular Kalman filter when HD observations are reliable and with slightly lower accuracy when they are unreliable; in the unreliable regime its certainty stays near its baseline and overestimates the filter's certainty.", "source": "Fig. 2I legend; Results, Neural Network Implementation, paragraph 7", "quote": "The Bayesian ring attractor network tracks the true HD with the same accuracy ... if HD observations are reliable and, therefore, more informative but with slightly lower accuracy once they become less reliable, and therefore less informative. This drop co-occurs with an overestimate in the belief\u2019s certainty", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_kutschireiter_2023.ipynb", "notes": "From figure2i.npz, the accuracies of the circKF and the Bayesian ring attractor differ by at most 0.0010 at rates of 1/s and above and 0.0003 at 3/s and above, so all four combinations hold. At 0.01, 0.03 and 0.1/s the attractor is less accurate (by 0.073, 0.084 and 0.057) and its mean certainty is higher (1.00 to 1.05 against 0.18 to 0.56).", "deviations": [], "assumed": false, "disagreed": false, "data": "elsewhere", "where": "Zenodo doi:10.5281/zenodo.7615975 (BayesianRingAttractor v1.0, commit a9db6f9), data_processed/figure2i.npz with Figure2i.ipynb", "proposed": null }, { "claim": "fig3b-bayesian-beats-conventional", "statement": "A conventional ring attractor (fast decay to a fixed bump amplitude, with that amplitude optimized) is less accurate than the Bayesian ring attractor at every HD observation reliability, and reaches its accuracy only when observations are highly unreliable.", "source": "Results, Bayesian vs. Conventional Ring Attractors, paragraphs 2 and 3; Fig. 3B", "quote": "Even after this optimization of the conventional ring attractor, it does not rival the accuracy of the Bayesian ring attractor. The Bayesian attractor performs consistently better, regardless of the amount of information available to the network", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_kutschireiter_2023.ipynb", "notes": "The optimal kappa* of the fixed-certainty filter is 1.7 (weighted accuracy 0.626 against 0.687 for the Bayesian attractor). The Bayesian attractor is more accurate at all 23 rates (smallest difference 0.0115 at 0.01/s, largest 0.131 at 100/s) and strictly so from 0.1 to 10/s (smallest 0.034). The reported part, not judged, holds only in part. The difference is smallest at 0.01/s, but at 0.03/s (0.036) it is larger than at 2 and 3/s (0.034 and 0.035).", "deviations": [], "assumed": false, "disagreed": false, "data": "elsewhere", "where": "Zenodo doi:10.5281/zenodo.7615975 (BayesianRingAttractor v1.0, commit a9db6f9), data_processed/figure2i.npz with Figure3b,c,e,d_FigureS3.ipynb", "proposed": null }, { "claim": "fig3c-update-weights", "statement": "The weight given to a new HD observation by the Bayesian ring attractor depends on both the observation's reliability and the current certainty and is close to that of the circular Kalman filter, deviating only for very uninformative observations; a conventional attractor weights observations by their reliability alone.", "source": "Fig. 3C legend; Results, Bayesian vs. Conventional Ring Attractors, paragraph 3; SI Appendix, Fig. S3 legend", "quote": "The update weight of the Bayesian attractor is close to optimal, visually indistinguishable from the circKF", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_kutschireiter_2023.ipynb", "notes": "At rates of 1/s and above, the weights computed with the circKF's and the attractor's mean certainty differ by at most 0.0010 with the Methods' weight (dt 0.01) and 0.0062 with the notebook's weight (dt 0.1), inside both bounds (0.02, 0.05). The largest difference, 0.021 and 0.130, falls at the lowest rate, 0.01/s.", "deviations": [], "assumed": false, "disagreed": false, "data": "elsewhere", "where": "Zenodo doi:10.5281/zenodo.7615975 (BayesianRingAttractor v1.0, commit a9db6f9), data_processed/figure2i.npz with Figure3b,c,e,d_FigureS3.ipynb", "proposed": null }, { "claim": "fig3d-best-network-slow-decay", "statement": "Across ring attractors with different bump decay speeds and baseline amplitudes, the best network has a slow decay (low beta) and performs about as well as the Bayesian ring attractor; performance drops as decay speed increases, which a larger baseline amplitude partly offsets.", "source": "Results, Bayesian vs. Conventional Ring Attractors, paragraph 4; Fig. 3D", "quote": "The network with the best performance overall had a slow bump decay speed (low \u03b2 ), as expected ( Fig. 3 D and E ). While it featured similar performance to the Bayesian ring attractor, it had slightly different \u03b2 and \u03ba * parameters ... As the bump decay speed \u03b2 increased further, performance dropped.", "statistics": [], "kind": "finding", "inferred": "", "outcome": "partially_reproduced", "blocked_by": null, "notebook": "reproduce_kutschireiter_2023.ipynb", "notes": "Normalized by the particle filter reference, the best network is kappa* 0.7, beta 0.3 (1.0012). Conditions 1, 2, 3 and 5 hold. The best beta is at most 1, the best network differs from (1, 0.5), the Bayesian attractor is 0.57% below the best, and above the best beta the maximizing kappa* is 0.8 to 1.6, never below 0.7. Condition 4 fails as written. The maximum over kappa* falls from 1.0012 (beta 0.3) to 0.9322 (beta 8), Spearman rho -0.98, but rises twice, by 0.0008 (beta 0.7 to 0.8) and 0.0014 (0.9 to 1). The two rises are smaller than the falls around them. With 10,000 runs per rate in each grid cell, the standard error of a cell's normalized performance is 0.0018 and that of a difference of two cells 0.0026, so the rises are 0.32 and 0.52 standard errors, within sampling error. The cells are independent samples, since performance_scan_figure3d.py runs each (kappa*, beta) as its own process and sets no seed (its lines are printed). The criterion gives no tolerance, so condition 4 fails as written.", "deviations": [], "assumed": false, "disagreed": false, "data": "elsewhere", "where": "Zenodo doi:10.5281/zenodo.7615975 (BayesianRingAttractor v1.0, commit a9db6f9), data_processed/figure3d.npz with Figure3b,c,e,d_FigureS3.ipynb", "proposed": null }, { "claim": "fig3f-noise-drop-below-70", "statement": "In a network of 64 neurons with additive neural noise, inference accuracy drops below 70% of its noise-free value once the noise is large enough.", "source": "Results, The Impact of Neural Noise on Inference Accuracy, paragraph 1; Fig. 3F", "quote": "Indeed, simulating a network of 64 noisy neurons shows that, once the noise becomes sufficiently large, inference accuracy drops to below 70% of its noise-free value", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_kutschireiter_2023.ipynb", "notes": "The noise-free optimum is kappa* 0.5, beta 0.2. Kept at those values, the network's relative performance is 63.1% at noise SD 10 and 38.1% at SD 20, under the notebook's normalization (particle filter) and the legend's (same network at SD 0, 63.07% and 38.06%), below 70% under all four combinations. It first falls below 70% at SD 9 (67.4%).", "deviations": [], "assumed": false, "disagreed": false, "data": "elsewhere", "where": "Zenodo doi:10.5281/zenodo.7615975 (BayesianRingAttractor v1.0, commit a9db6f9), data_processed/figure3f/sigma=*.npz with Figure3f_FigureS4.ipynb", "proposed": null }, { "claim": "fig3f-retuning-reduces-drop", "statement": "Retuning the bump decay speed and baseline amplitude for each noise level reduces the drop in inference accuracy by about 35% compared with the network tuned for noise-free performance, and raises the signal-to-noise ratio.", "source": "Results, The Impact of Neural Noise on Inference Accuracy, paragraph 3; Fig. 3F; SI Appendix, Fig. S4", "quote": "Indeed, doing so reduced the drop in inference accuracy by about 35% when compared to the network tuned to optimize noise-free performance ( Fig. 3 D and F ) and also boosted the network\u2019s signal-to-noise ratio", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_kutschireiter_2023.ipynb", "notes": "The reduction of the drop by retuning is 34.64% at SD 10 with the particle filter normalization (the first reading, given as measured) and 34.56% with the legend's, and 31.67% for the mean over SD 1 to 10 with the particle filter normalization, inside 31.5 to 38.5%. The other readings are outside it, 31.19% (mean, legend normalization) and 29.80 and 29.76% (SD 20). The retuned signal-to-noise ratio is above the untuned one at every noise SD above 0 (ratio 1.05 to 3.00). Fig. S4, reported only, the best kappa* grows from 0.3 to 4.8 between SD 1 and 10, and the best beta from 0.1 to 2.0 but not steadily, and both fall again at SD 15 and 20.", "deviations": [ "The paper does not say at which noise level the 35% is taken; the claims file names six readings, three of them match, and all are shown." ], "assumed": false, "disagreed": false, "data": "elsewhere", "where": "Zenodo doi:10.5281/zenodo.7615975 (BayesianRingAttractor v1.0, commit a9db6f9), data_processed/figure3f/sigma=*.npz with Figure3f_FigureS4.ipynb", "proposed": null }, { "claim": "noise-effects-on-mu-and-kappa", "statement": "Neural noise makes the bump position diffuse without bias and makes the bump amplitude drift upward and diffuse, which leads to overconfidence; for additive noise, a four-fold larger network halves the noise's effective SD.", "source": "Results, The Impact of Neural Noise on Inference Accuracy, paragraphs 2 and 3; SI Appendix, section 5", "quote": "First, it causes an unbiased random diffusion of \u03bc t and, thus, an increasingly imprecise memory of the HD estimate. Second, it causes a positive drift and random diffusion of \u03ba t", "statistics": [], "kind": "finding", "inferred": "", "outcome": "partially_reproduced", "blocked_by": null, "notebook": "reproduce_kutschireiter_2023.ipynb", "notes": "2,000 runs of T 5 per setting, noise SD 1 and 2, N 64 and 256. (1) holds. The mean change of position is within 2 standard errors of 0 in all four settings (z -0.74 to 0.82) and its variance grows at each second (for N 64, SD 1, 0.034 at 1 s to 0.232 at 5 s). (3) holds. The SD of the one-step amplitude increments at N 256 is 0.500 and 0.501 times that at N 64. (2) fails. The mean amplitude falls throughout the run and ends at 0.83, 0.65, 0.82 and 0.42, all below the noise-free 1.00, so there is no upward drift and no overconfidence in this network. In RNN_filter_Run the noise also raises the rectified sum that drives the global inhibition (from 0.5 without noise to 0.57 to 0.77 at 1 s), and the zero mode stays near 0. With the inhibition held at its noise-free constant the mean amplitude rises to 1.08, 1.38, 1.01 and 1.08 after 5 s, within 1.9 standard errors of the mean that SI Eq. S124 implies (1.08, 1.35, 1.02, 1.08). That control removes both the noise's effect on the inhibition and the restoring force toward kappa* 1, so it confirms the SI's noise term but does not isolate the cause. With the inhibition set to the value a noise-free bump of the decoded amplitude gives, which keeps the restoring force, the mean amplitude ends at 1.002, 1.045, 1.003 and 1.005, above the noise-free 1.000 to 1.001 in all four settings (z 0.35, 6.2, 1.3, 1.1) and within 1.2 standard errors of the reduced two-dimensional SDE of the first mode. So the noise's direct effect gives a small upward drift, and in the released network the fall comes from the noise's own contribution to the rectified sum that drives the inhibition. The criterion tests the full network; whether the paper claims the drift for the full network is asked of a reader (questions/000289.yaml).", "deviations": [ "The runs use a batched copy of RNN_filter_Run (bra.rnn_batch), the same update term by term; on the same noise draws it matches the released function to 4e-16.", "The released RNN_filter_Run has no z parameter and raises NameError; z=None was added to its signature, as the claims file says." ], "assumed": false, "disagreed": false, "data": "elsewhere", "where": "Zenodo doi:10.5281/zenodo.7615975 (BayesianRingAttractor v1.0, commit a9db6f9), circularFiltering.py RNN_filter_Run", "proposed": null }, { "claim": "fig4b-connectome-motifs", "statement": "The connectivity among EPG, Delta7 and PEN1 neurons in the hemibrain connectome resembles the connectivity of a network that implements the approximate circular Kalman filter, with shifted PEN1 to EPG projections and broad EPG to Delta7 to EPG inhibition.", "source": "Fig. 4 B and C legends; SI Appendix, section 4A", "quote": "The connectivity profile of a recurrent neural network (RNN) ( Fig. 2 D \u2013 F ) that implements an approximate circKF is strikingly similar to the connectivity of neurons in the Drosophila HD system.", "statistics": [], "kind": "finding", "inferred": "", "outcome": "partially_reproduced", "blocked_by": null, "notebook": "reproduce_kutschireiter_2023.ipynb", "notes": "Sorted by the PB glomerulus in the name (reading (a)), left PEN1 neurons send their EPG output a mean of +0.56 glomeruli from their EPG input (10 of 10 positive, Wilcoxon p 0.002) and right PEN1 neurons -0.48 (9 of 10 negative, p 0.004), among same-side EPG neurons; with partners on both sides, +0.37 and -0.35. The two edge neurons (L9, R9) have offsets near 0. Every Delta7 receives EPG input from 13 to 16 of the 16 EPG glomeruli and sends output to EPG neurons in 9 to 15, more than half. Counting only glomeruli with at least 3 (or 5) synapses, inputs still reach more than half for all 42 Delta7, but outputs do for 7 (or 0) of 42, so the breadth of the Delta7 to EPG side depends on the threshold. Reading (b), by ellipsoid body or gall position, could not be run, and the criterion requires both, so the claim holds in part (changed from reproduced after verification round 1). Whether reading (b) belongs to the claim is asked of a reader (questions/000289.yaml).", "deviations": [ "The criterion asks for two sorts. Reading (b), by the ellipsoid body or gall position in the neuron's name, cannot be formed: hemibrain instance names give only the PB glomerulus (EPG(PB08)_L1, PEN_a(PB06a)_R2). The outcome rests on reading (a) alone.", "Assumed, as the Fig. 4B legend counts \"the total number of synaptic connections between each cell pair\" with no threshold, that a connection is any synapse; the counts at 3 and 5 synapses are shown.", "Assumed a measure of the offset, since the paper gives none. For each PEN1 neuron, it is the synapse-weighted circular mean position (period 8 glomeruli) of its EPG targets minus that of its EPG inputs, among same-side EPG neurons, with both sides shown.", "The connectome is the public hemibrain v1.2 export, not v1.2.1 as in the legend. neuPrint, which serves v1.2.1, now requires a personal API token for all data." ], "assumed": true, "disagreed": true, "data": "elsewhere", "where": "neuPrint hemibrain:v1.2.1 (Scheffer et al. 2020, eLife 9, e57443), synapse counts between EPG, Delta7 and PEN1 neurons", "proposed": null }, { "claim": "fig4gh-drosophila-network-matches", "statement": "A Drosophila-like multipopulation network, constrained by the connectome, tracks HD with an accuracy indistinguishable from the idealized Bayesian ring attractor across observation reliabilities.", "source": "Results, Tuning a Biological Ring Attractor, paragraph 3; Fig. 4 G and H", "quote": "We found that this network achieves a HD encoding accuracy which is indistinguishable from that of our idealized Bayesian ring attractor network ( Fig. 4 G and H ).", "statistics": [], "kind": "finding", "inferred": "", "outcome": "partially_reproduced", "blocked_by": null, "notebook": "reproduce_kutschireiter_2023.ipynb", "notes": "From figure4g.npz, the Drosophila-like network's accuracy is within 0.02 of the Bayesian ring attractor's at all 23 rates (largest 0.019 at 0.1/s), so reading (b) holds, but it exceeds 0.01 at 0.1, 0.2 and 0.3/s (0.019, 0.014, 0.010), so reading (a) fails; the criterion needs both. At 0.4/s and above the difference is at most 0.004. The three differences above 0.01 are 1.95, 1.51 and 1.16 times the standard error of a difference of two independent scans (the two networks share their runs, so the actual error is smaller), and the same Bayesian attractor differs by 0.022 between figure4g.npz and figure2i.npz at 0.1/s, so the 0.01 bound is near the sampling error there. Mean certainties, reported only, are lower in the Drosophila-like network at every rate (0.88 against 1.00 at 0.01/s, 13.8 against 14.5 at 100/s).", "deviations": [], "assumed": false, "disagreed": false, "data": "elsewhere", "where": "Zenodo doi:10.5281/zenodo.7615975 (BayesianRingAttractor v1.0, commit a9db6f9), data_processed/figure4g.npz with Figure4g,h.ipynb and network_filter.py", "proposed": null }, { "claim": "figs2-performance-independent-of-n", "statement": "The Bayesian ring attractor's inference accuracy depends little on its number of neurons.", "source": "SI Appendix, Fig. S2 legend; Results, Neural Network Implementation, paragraph 6", "quote": "Network inference performance is mostly independent of the number of neurons N in the Bayesian ring attractor network.", "statistics": [], "kind": "finding", "inferred": "", "outcome": "not_reproduced", "blocked_by": null, "notebook": "reproduce_kutschireiter_2023.ipynb", "notes": "Of the 160 values in figureS2.npz (7 network sizes, 23 rates, one missing), one is more than 0.02 from the quadratic circKF reference, N 16 at 0.1/s (0.041 below). Every other value, N 3 and 4 included, is within 0.019. Both readings include N 16, so neither holds, and the criterion requires both (changed from partially_reproduced after verification round 1). The deviation is not sampling error. With 10,000 runs per rate, the standard error of a difference of two independent scans at 0.1/s is 0.0096, and N 16 is 4.3 of them below the reference, with no other value beyond 3, and 0.029 below the mean of the other six sizes at that rate. The 0.02 bound is near the sampling error at low rates (about 1.2 of the values with N of 8 and more expected beyond it by chance alone), but this value is far beyond it. One value (N 128 at 100/s) is missing (NaN) in the stored array. Whether \"mostly independent\" allows single exceptions is asked of a reader (questions/000289.yaml).", "deviations": [], "assumed": false, "disagreed": true, "data": "elsewhere", "where": "Zenodo doi:10.5281/zenodo.7615975 (BayesianRingAttractor v1.0, commit a9db6f9), data_processed/figureS2.npz and vonMises_q_stats.npz with FigureS2.ipynb", "proposed": null }, { "claim": "prediction-bump-amplitude-cue-reliability", "statement": "The paper predicts that the bump amplitude in the HD system is higher in the presence of reliable external HD cues, such as salient visual landmarks. The paper does not test this.", "source": "Discussion, paragraph 7", "quote": "Second, we predict that bump amplitude in the HD system should vary dynamically, with higher amplitudes in the presence of reliable external HD cues, such as salient visual landmarks.", "statistics": [], "kind": "finding", "inferred": "", "outcome": "blocked", "blocked_by": "methods_underspecified", "notebook": "reproduce_kutschireiter_2023.ipynb", "notes": "Tested on the 15 flies of the dandiset's cue-contrast experiment, high (more reliable) against low contrast, with bump amplitude the length of the population vector at the bump's spatial frequency across the PB midline ROIs (frequency 2, the spectral peak in every fly). The readings disagree, and the criterion blocks the claim when they do. With dF/F, amplitude is higher at high contrast in 10 of 15 flies (reading (a) holds) but the one-sided Wilcoxon test gives p 0.138 (reading (b) fails). With raw fluorescence minus baseline, 12 of 15 flies (a) and p 0.021 (b). Median high to low ratio 1.12 (dF/F) and 1.11 (raw). Variants, high against zero contrast give 12 of 15 flies (p 0.047) and 13 of 15 (p 0.006), and low contrast is not above zero contrast (7 of 15 with either signal). Without sub-2, which has one high-contrast block, dF/F gives 9 of 14 (p 0.21) and raw 11 of 14 (p 0.034). Across 18 variants of the raw signal (smoothing 0, 2 or 4 pixels, baseline at the 5th, 10th or 20th percentile, F - F0 or (F - F0) / F0), p is below 0.05 in all nine F - F0 variants (0.021 to 0.042) and in none of the nine divided by F0 (0.11 to 0.89), so the split between the readings follows the division by the baseline. Question for the paper's authors (finding below). The Discussion gives no experimental method for the prediction (which activity signal, which cue conditions are compared, which test). With these data the answer depends on the signal and the test, and no reading has more support in the paper than another.", "deviations": [ "The dandiset holds data of a separate study, not of this paper; the paper does not test its prediction.", "Raw fluorescence is read from 10% of the frames (25 volumes every 250), averaged over planes and smoothed in the plane (Gaussian SD 4 pixels) before averaging in each ROI mask; the baseline is each ROI's 10th percentile. The projection and the smoothing were chosen as those that best reproduce the stored dF/F (median correlation over ROIs 0.93, against 0.66 without smoothing). Smoothing 0 and 2 pixels, baselines at the 5th and 20th percentile and the baseline-divided signal are shown as variants." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null } ], "findings": [ { "kind": "metadata_error", "statement": "The dandiset's related resources name the bioRxiv preprint \"Bayesian inference in ring attractor networks\" (10.1101/2021.12.17.473253 version 4) as the paper it supplements, but its description says the data belong to \"Uncertainty and learning in a biological ring attractor\" (Basnak et al., in preparation). The linked paper, published as Kutschireiter et al. 2023 (PNAS), is a theory paper that analyses none of these data, so the dandiset links no paper that uses them.", "found": null, "status": null }, { "kind": "metadata_error", "statement": "imaging_rate is 9.18 Hz in every file, but sub-4 and sub-10 (cue-contrast experiment) are imaged at 7.87 Hz (0.127 s frame interval, 9446 frames in 1200 s) with 14 planes, against 9.18 Hz and 12 planes in the other 13 sessions.", "found": null, "status": null }, { "kind": "other", "statement": "In sub-2 the trials table has five blocks (high, zero, low, zero, low) where every other fly of the cue-contrast experiment has six with two of each contrast. The last block lasts 405 s against about 200 s in the protocol, so it may hold a low and a high block labelled as one.", "found": null, "status": null }, { "kind": "metadata_insufficient", "statement": "The processing behind the stored dF/F is not described (which planes, the extent of each ROI, the baseline). The ROI masks are 2D, 3 to 11 pixels each (median 7), and the dF/F agrees with the raw imaging averaged in those masks only after spatial smoothing (median correlation over ROIs 0.66 without, 0.93 with a 4-pixel Gaussian, sub-1). The RoiResponseSeries unit is \"na\".", "found": null, "status": null }, { "kind": "methods_unavailable", "statement": "Question for the paper's authors. The Discussion predicts higher bump amplitude with reliable external HD cues but gives no experimental method. Which activity signal (dF/F or raw fluorescence), which comparison of cue conditions and which test would test it? On this dandiset's cue-contrast experiment, high contrast exceeds low contrast in most flies with either signal, but a one-sided Wilcoxon test over 15 flies gives p 0.021 with raw fluorescence and p 0.138 with dF/F.", "found": null, "status": null }, { "kind": "methods_unavailable", "statement": "Question for the paper's authors. Fig. 4B sorts neurons \"by their spatial position as a proxy for their preferred HD\". Which position was used? The hemibrain names give only the PB glomerulus for EPG and PEN1 neurons. This attempt sorted by it and counted any synapse as a connection.", "found": null, "status": null }, { "kind": "code_mismatch", "statement": "The released code does not run as published. RNN_filter_Run (circularFiltering.py) uses z but has no z parameter, and Figure2h.ipynb and the scan scripts pass kappa_y to functions whose parameter is kappa_v. One NaN remains in figureS2.npz (N 128 at 100/s).", "found": null, "status": null }, { "kind": "code_mismatch", "statement": "In the released single-population network (RNN_filter_Run) at the Bayesian ring attractor's parameters, additive neural noise lowers the mean bump amplitude (from 1.00 to 0.42 to 0.83 after 5 s, N 64 and 256, noise SD 1 and 2). The Results and SI Appendix section 5 say the noise causes a positive drift of the amplitude. With the inhibition set to the value a noise-free bump of the decoded amplitude gives, which keeps the restoring force and removes only the noise's own contribution to the rectified sum that drives the inhibition, the mean amplitude ends at 1.002 to 1.045, above the noise-free value, so the fall comes from that contribution.", "found": null, "status": null } ], "reconciliation": null, "verifications": [ { "date": "2026-10-30", "kind": "agent", "by": "Independent agent run by CatalystNeuro (same harness, separate context, did not write the notebook)", "reviewed": "Round 1. The first executed notebook with the conclusion cells and the summary removed, the helper module and the claims file, with the paper. The verifier recomputed the prediction's p-values from the printed table, tested bra.fourier on a synthetic bump and checked the asset count of the dandiset version. Its report had no attempt record and gave its own outcome for each claim; 11 of 13 agree with the record, and fig4b-connectome-motifs and figs2-performance-independent-of-n do not.", "agreed": 11, "of": 13, "disagreements": [ { "claim": "fig4b-connectome-motifs", "outcome": "partially_reproduced", "reason": "Under reading (a) the offsets have opposite signs (left +0.56, right -0.48 on the same side) and all 42 Delta7 reach more than 8 EPG glomeruli for input and output at 1 synapse or more. Reading (b), which the criterion requires, cannot be formed because no instance name holds an EB or gall position. The Delta7 breadth depends on the synapse threshold (output in 42, 7 and 0 of 42 at 1, 3 and 5 synapses); the connectome is v1.2, not v1.2.1 (B4)." }, { "claim": "figs2-performance-independent-of-n", "outcome": "not_reproduced", "reason": "N 16 at 0.1/s deviates by 0.041, so both readings (N of 8 and more, N of 16 and more, within 0.02) fail; it is the only value above 0.02, and N 3 and 4 stay within 0.0186. The bound is close to the run-to-run variability and the notebook gives no Monte Carlo SE (B2). One NaN value for N 128 is skipped by nanmax and its rate is not printed (B8)." } ] }, { "date": "2026-10-30", "kind": "agent", "by": "Independent agent run by CatalystNeuro (same harness, separate context, did not write the notebook)", "reviewed": "Round 2. The full executed notebook of the first revision with its conclusions and summary, the helper module, the README, the attempt record and the claims file, with the paper. The verifier checked each outcome, each assumption and each sentence of the conclusions against the printed outputs, and whether its round-1 points were resolved. It agreed with all 13 outcomes.", "agreed": 13, "of": 13, "disagreements": [] } ] } ], "entries": [ { "name": "000289-kutschireiter-2023", "type": "reproduction", "title": "Bayesian inference in ring attractor networks (Kutschireiter et al. 2023)", "description": "Tests the findings of Kutschireiter, Basnak, Wilson and Drugowitsch (2023), a theory paper. Its simulation findings are tested on the paper's released code and processed outputs (Zenodo), its connectome finding on the hemibrain connectome, and its prediction that bump amplitude is higher with a reliable cue on dandiset 000289, which holds two-photon imaging of EPG neurons in flies shown a visual cue at three contrasts, from a separate study by the same lab.", "datasets": [ "dandi:000289" ], "papers": [ "10.1073/pnas.2210622120" ], "derived_from": [], "notebooks": [ { "path": "reproduce_kutschireiter_2023.ipynb", "title": "Testing Kutschireiter et al. (2023) with dandiset 000289", "executed": "notebooks/000289-kutschireiter-2023/reproduce_kutschireiter_2023.html" } ] } ] }