{"id":"8f2104b4-3de7-480b-be68-f9539c1c236b","arxiv_id":"2509.01760","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"The H1 neuron's velocity response adapts to the contrast distribution along a dominant scaled dimension, and its integration time tracks the mean interspike interval.","lead":"This paper shows that a motion-sensitive neuron in the fly visual system adapts to the recent distribution of contrast, not just the current contrast, and that this adaptation occupies a low-dimensional space. It also finds that the neuron sets its integration time to match the mean interval between spikes, which would reduce redundant signaling.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SVD rank test uses an i.i.d. null that ignores correlated STA noise, so the low-dimensional claim is not yet established.","rationale":"The low-rank claim is the structural backbone of the abstract and the main novel contribution beyond the already-known contrast dependence of H1. The only quantitative support for rank two is the SVD test, and its null is misspecified relative to the paper's own description of the noise. The reader identified exactly this weakness in the i.i.d. null model, and the paper's Section III.A explicitly says that the STA fluctuations are larger than independent sampling would predict, making the i.i.d. null in Section III.B inconsistent with the paper's own error description. This is a real, addressable technical gap. Data and code are not available, so the surrogate test cannot be run by the reader; that justifies a conditional rather than unconditional acceptance. Other weaknesses—lack of quantitative validation for the C/clim collapse, the exponential-fit assumption, and the abstract's 'single dominant dimension' versus the body's two significant modes—are secondary but would also be clarified by the same surrogate analysis and by reporting residuals and error bars. The core phenomenological result of context-dependent adaptation (Fig. 3, same absolute contrast C=0.25 giving different responses across clim) is well supported and does not depend on the rank test. Thus the reader's CONDITIONAL verdict remains appropriate; no verdict change is needed.","tokens_in":15601,"tokens_out":11990,"duration_ms":152102,"concrete_test":"Recompute the SVD significance test using surrogate STA matrices that preserve the real noise correlations: block-bootstrap the observed spike times (or simulate the fitted LNP/GLM model) while holding the contrast time series fixed, recompute STA(tau;C) for each context, form the stacked 90x100 matrix, and compute the SVD. Repeat many times to build a null distribution of singular values that includes correlated noise. Then count how many data singular values exceed, say, the 95th percentile of this surrogate null, and report bootstrap confidence intervals for the ratios S2/S3 and S1/S2. If the count remains two and the first mode dominates, the low-rank claim is secure; if the count changes, the central abstraction must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central structural claim—that context-dependent STAs live in a low-dimensional space with a single dominant dimension—rests on the SVD significance test in Section III.B and Fig. 4. The null there is a 90x100 matrix whose elements are drawn independently from the marginal distribution of the real STA entries. This does not match the actual error structure of the data. The estimated STAs are averages over the same spike train, so sampling errors are correlated across neighboring tau bins; this is visible as smooth high-frequency wiggles in Fig. 1. The paper itself notes in Section III.A that the noise is larger than simple independent sampling because the samples are not independent. The i.i.d. null can therefore mis-estimate the noise floor of the singular-value spectrum. It is unknown whether the excess of the first two singular values would survive a surrogate null that preserves temporal and across-bin correlations. If a third mode becomes significant, the two-mode decomposition in Eq. (19) and the abstract's 'single dominant dimension' language would need revision; if only one mode survives, the proposed role of the second mode in encoding absolute contrast is unsupported. The rank determination is thus the most load-bearing unverified step in the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an experimental study of context-dependent adaptation in the fly motion-sensitive neuron H1. The authors use naturalistic synthetic stimuli with a clean separation of time scales: fast white-noise velocity, contrast c(t) varying with correlation time τc = 500 ms, and a slowly changing contrast dynamic range clim defining the context. They measure contrast-conditional spike-triggered averages STA(τ; C) and show that the response at fixed absolute contrast depends on clim. An SVD of the stacked STA matrices is claimed to reveal two significant modes, with the dominant mode scaling as C/clim and the second depending on absolute contrast. Finally, the effective integration time extracted from exponential fits is found to approximately obey r̄ τint ≈ 1 across contrast and context, which is interpreted as redundancy reduction. The main evidence is phenomenological and largely qualitative.","tokens_in":15953,"tokens_out":7311,"duration_ms":89930,"significance":"If correct, the results are significant: they document a form of distributional/context adaptation in a well-characterized neuron and propose a remarkably simple low-dimensional structure, plus a striking matching of integration time to spike interval that connects to efficient coding ideas. Strengths include carefully controlled stimulus design, long stable recordings, replication across eight flies, and an honest statement that no increase in per-spike information was found. The empirical r̄ τint ≈ 1 relation is not forced by construction and is a concrete falsifiable claim. However, the central low-rank and context-dependence claims rest on statistical and design choices that need to be strengthened before the conclusions can be accepted.","major_comments":[{"comment":"The rank determination uses an i.i.d. null model: random 90×100 matrices whose elements are drawn independently from the marginal distribution of the real STA entries. This does not match the error structure of the data. The STAs are averages over overlapping spike-triggered trajectories, so sampling errors are correlated across τ bins; the paper itself notes in the Section III.A footnote that 'the noise is a little larger because not all samples are independent.' An i.i.d. null can therefore misestimate the singular-value noise floor, and the conclusion that exactly two modes are significant is load-bearing for Eq. (19) and the abstract's 'single dominant dimension.' I request a surrogate null that preserves temporal correlations (e.g., phase randomization of STA rows or bootstrap resampling over spike trains) and/or a cross-validated rank assessment (e.g., reconstruction error for rank","section":"Section III.B, Fig. 4"},{"comment":"The six context conditions were presented in fixed order from the lowest to the highest contrast, with all main-text results from a single fly. The central demonstration of context dependence (Fig. 3) compares responses at the same absolute contrast C = 0.25 across these sequential blocks. A slow drift in recording quality, receptor state, or carryover adaptation from the preceding block is therefore an alternative explanation for the observed differences. Please state whether the order was randomized across the eight replication flies, or provide evidence of temporal stability (e.g., by repeating one context at a later time and showing the same STA). Without such a control, the 'context dependence' claim is vulnerable to a time/order confound.","section":"Appendix B; Section II.C"},{"comment":"The statement that 'almost no points significantly above this line' is not backed by a quantitative comparison. The figure appears to show error bars on r̄, while the text in Appendix D gives bootstrap estimates of τint variability, but it is not stated whether these τint uncertainties are displayed or propagated. Since the redundancy-reduction conclusion in the abstract and Discussion rests on the tightness of r̄ τint ≈ 1, please report the uncertainties in τint and provide a statistical test of the scatter about the line, for example a reduced chi-square or bootstrap confidence intervals for the product r̄ τint.","section":"Section III.C, Fig. 7"}],"minor_comments":[{"comment":"The lag convention for τ is inconsistent: Eq. (14) defines STA(τ) = ⟨v(t_i + τ)⟩, while Eq. (18) uses v(t_i − τ), and Eq. (21) is written as STAfit(−τ). Please define τ unambiguously (e.g., τ > 0 as time before the spike) and keep the sign convention uniform throughout.","section":"Eqs. (14), (18), (21)"},{"comment":"The abstract says 'a single dominant dimension,' but the analysis concludes that two modes are significant. Clarify whether 'dominant' refers to the relative size of S1 versus S2, or whether only one mode is claimed to be significant; the current wording is stronger than the two-mode decomposition in Eq. (19).","section":"Abstract; Section III.B"},{"comment":"The caption refers to 'clim = 0.15%' in the text; this should presumably be 'clim = 0.15'.","section":"Fig. 6 caption"},{"comment":"The claim that results are 'qualitatively similar' for τc = 0.25–15 s is not accompanied by data or a figure. Since the main text shows only τc = 500 ms, either show the additional conditions or soften the claim about the 'wide range of conditions.'","section":"Section III.A; Section II.A"}],"recommendation":"major_revision","confidential_remarks":"This is a strong phenomenological study with an interesting and potentially important conclusion, but the two load-bearing issues—the i.i.d. SVD null and the fixed-order context presentation—need to be resolved before publication. The first can likely be addressed with existing data by constructing correlated-noise surrogates or a bootstrap rank test; the second needs at least a clear statement about randomization or stability checks. The r̄ τint ≈ 1 relation is striking but needs proper error propagation. I do not think new experiments are necessarily required, but the revision must make the statistical basis of the central claims transparent."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a solid experimental paper. The main qualitative result — that H1's velocity STA depends on absolute contrast and, at the same contrast, on the distribution of contrast it comes from — is directly supported by the data and doesn't rely on the fancier analysis. That alone makes it a useful contribution.\n\nThe new bits are real: the explicit demonstration of context-dependent contrast adaptation in a motion-sensitive neuron, the low-rank SVD characterization of the STA family, and the observation that integration time tracks the mean interspike interval (τ_int ≈ 1/r̄). The stimulus design is careful, with a clean separation of timescales, and the replication across eight flies is reassuring.\n\nThe soft spot is the SVD significance test in Fig. 4. The null is a random 90×100 matrix with i.i.d. entries from the marginal distribution of the real STA elements. But the STA estimates are averages over the same spike train, so the sampling noise is correlated across time bins — the paper itself acknowledges this in Section III.A. An i.i.d. null can misestimate the noise floor, so the number of significant modes (two) is not yet firmly established. If a third mode survives a correlated-noise surrogate, Eq. (19) and the abstract's 'single dominant dimension' need revision; if only one survives, the second mode's proposed role in encoding absolute contrast is unsupported. This is the load-bearing unverified step.\n\nTwo smaller issues: Fig. 6's scaling collapse has no error bars, so 'a very good approximation' is hard to judge, and the abstract says 'single dominant dimension' while the body says two significant modes. The exponential fit in Eq. (21) is a modeling assumption, though the examples in Fig. 8 look reasonable. No code or data are released, so these statistical details can't be checked independently.\n\nNone of this breaks the central claims. The context dependence is visible in Fig. 3 without any SVD, and the τ_int vs. r̄ relation in Fig. 7 is independent of the rank test. This paper deserves a serious referee, and I'd be happy to discuss it in a reading group.","headline":"Solid experimental demonstration of context-dependent contrast adaptation in H1; the low-rank claim is plausible but rests on a weak SVD significance test.","tokens_in":16372,"tokens_out":2781,"would_cite":true,"duration_ms":29791,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A fly motion-sensitive neuron adapts its velocity response to the distribution of image contrast, and nearly all of that adaptation is captured by a single dominant direction in stimulus space.","keywords":["fly visual system","H1 neuron","contrast adaptation","context dependence","spike-triggered average","low-rank representation","efficient coding","sensory adaptation"],"falsifier":"Perform the same singular-value significance test on surrogates that preserve the temporal correlations in the STA residuals, for example by block-shuffling along the time axis or adding smoothed Gaussian noise with the measured power spectrum. If the second singular value falls below the surrogate noise floor, the two-dimensional description is not supported; if it survives, the claim is strengthened.","tokens_in":1653,"feed_emoji":"🪰","tokens_out":1847,"duration_ms":78375,"temperature":0.7,"pith_summary":"This paper asks whether a sensory neuron adapts not just to the current level of a stimulus but to the statistical context in which that stimulus appears. Using the fly motion-sensitive neuron H1, the authors deliver naturalistic movies in which the velocity is white noise, the contrast fluctuates on an intermediate timescale, and the dynamic range of contrast changes only slowly. They show that the neuron's velocity filter, measured as a spike-triggered average, depends on both contrast and context: the same contrast produces different responses when it comes from a narrow versus a wide distribution. The context-dependent changes are low-dimensional, dominated by a single mode, and the integration time of the response tracks the mean interval between spikes, suggesting the system is continually balancing noise reduction against redundancy.","feed_headline":"Contrast context reshapes fly motion coding along one dimension","feed_subtitle":"H1's velocity filter adapts to the contrast range it sees, matching integration time to spike interval.","key_machinery":"The central object is the contrast-conditioned spike-triggered average STA(τ;C), a matrix that records the average velocity trajectory preceding a spike when the contrast at the relevant past time falls in a small bin around C. Because velocity is white noise, the STA is proportional to the neuron's linear filter, so changes in the STA are changes in the computation itself. The analysis stacks these matrices across contrast bins and across context values into one large matrix and applies singular value decomposition; the significant singular vectors become the basis functions in which adaptation occurs. A separate exponential-with-delay fit to each STA provides the integration time τint, whi","core_discovery":"The central claim is that H1's response to visual motion adapts to the full distribution of contrast, not just to the instantaneous contrast level. By conditioning spike-triggered averages on contrast, STA(τ;C), and repeating this for several contrast dynamic ranges (clim), the authors find that a given contrast value evokes different responses depending on the context. Stacking all the contrast-conditioned STAs and performing a singular value decomposition, only two modes are significant: the dominant mode's weight depends almost entirely on contrast scaled by the dynamic range, U1(C,clim)=g(C/clim), while the second mode depends on absolute contrast. As a consequence, the neuron's highest","pith_inferences":["A natural next test is to repeat the singular-value significance analysis with surrogate noise that preserves the temporal correlations visible in the STA residuals; if the second mode then falls below the noise floor, the claim should be reduced to a single dominant dimension.","If the pattern generalizes, other sensory neurons with mixed selectivity may show similarly low-rank context dependence, with separate dimensions for distribution-scaled and absolute stimulus features.","The clean separation between a scaled dimension and an absolute dimension suggests a two-pathway model: one divisive gain-control mechanism tracking recent dynamic range, and one slower mechanism tracking absolute contrast.","A behavioral test could ask whether flies' optomotor responses track the context-dependent STA shapes rather than instantaneous contrast, connecting the observed filter changes to actual flight control."],"forward_implications":["The response to a given contrast is not a fixed property of the neuron; it is normalized by the dynamic range of contrasts recently experienced.","Because the dominant dimension scales as C/clim, efficient coding can be achieved by adapting to the distribution's range, while the weaker absolute-contrast dimension may resolve the resulting ambiguity on longer timescales.","The relation τint≈1/r̄ means H1 stays in the one-spike-per-event regime, where successive spikes carry nearly independent information and rate-code redundancy is avoided.","Context dependence persists for contrast correlation times from 0.25 s to 15 s, showing that the adaptive mechanism is not tied to a single timescale.","Together with earlier work on adaptation to velocity distributions, the results support a general principle: neural computation adapts quantitatively to the distribution of sensory inputs."],"supporting_citations":[{"why":"Supplies the identity that for white-noise velocity input the neuron's filter is proportional to the spike-triggered average, making STA matrices the response measure used throughout.","marker":"[42]"},{"why":"Early evidence that H1's response to motion depends on contrast and adapts, the phenomenon this paper extends to distributional context.","marker":"[30]"},{"why":"Early demonstration that H1's response dynamics adapt, providing precedent for stimulus-dependent integration in the same neuron.","marker":"[31]"},{"why":"Shows H1 scales its input/output range to the velocity distribution, supplying the scaling-to-dynamic-range logic generalized here to contrast context.","marker":"[32]"},{"why":"Shows H1 adapts to velocity dynamic range over many timescales and that information transmission drops before adaptation, informing the wide-timescale and efficiency interpretations.","marker":"[33]"},{"why":"Establishes contrast dependence of H1's response and the use of spike-triggered analysis for motion adaptation in this cell.","marker":"[47]"},{"why":"Provides the 'spikes as events' perspective and the one-spike-per-characteristic-time regime that the τint≈1/r̄ result connects to.","marker":"[36]"},{"why":"Shows a low-dimensional adaptive structure in salamander retinal responses, the closest comparative precedent for the rank-two result.","marker":"[64]"}],"fun_headline_variants":["Fly motion coding adapts to contrast statistics in one dimension","Context, not just contrast, sets H1's velocity filter shape","H1's response to motion depends on contrast distribution","Single dominant mode explains context-dependent fly vision","Fly H1 matches integration time to mean spike interval"],"cache_read_input_tokens":18176,"weakest_assumption_plain":"The rank-two claim depends on the noise in the spike-triggered averages being statistically independent at each time point; the actual residual noise is visibly correlated in time, so the number of genuinely significant dimensions could be one or three under a more realistic noise model.","fun_headline_variants_meta":{"raw":{"variants":["Fly motion coding adapts to contrast statistics in one dimension","Context, not just contrast, sets H1's velocity filter shape","H1's response to motion depends on contrast distribution","Single dominant mode explains context-dependent fly vision","Fly H1 matches integration time to mean spike interval"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000173,"raw_usage":{"total_tokens":1066,"prompt_tokens":648,"completion_tokens":418,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":392,"completion_tokens_details":{"reasoning_tokens":340}},"tokens_in":392,"tokens_out":418,"duration_ms":5196,"temperature":1.0,"reasoning_tokens":340,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:12:14.047523+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Perform the same singular-value significance test on surrogates that preserve the temporal correlations in the STA residuals, for example by block-shuffling along the time axis or adding smoothed Gaussian noise with the measured power spectrum. If the second singular value falls below the surrogate noise floor, the two-dimensional description is not supported; if it survives, the claim is strengthened.","supporting_citations":[{"cited_title":"de Boer and P","cited_arxiv_id":null,"evidence_quote":"Supplies the identity that for white-noise velocity input the neuron's filter is proportional to the spike-triggered average, making STA matrices the response measure used throughout."},{"cited_title":"Maddess and S","cited_arxiv_id":null,"evidence_quote":"Early evidence that H1's response to motion depends on contrast and adapts, the phenomenon this paper extends to distributional context."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Early demonstration that H1's response dynamics adapt, providing precedent for stimulus-dependent integration in the same neuron."},{"cited_title":"Brenner, W","cited_arxiv_id":null,"evidence_quote":"Shows H1 scales its input/output range to the velocity distribution, supplying the scaling-to-dynamic-range logic generalized here to contrast context."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows H1 adapts to velocity dynamic range over many timescales and that information transmission drops before adaptation, informing the wide-timescale and efficiency interpretations."},{"cited_title":"de Ruyter van Steveninck and W","cited_arxiv_id":null,"evidence_quote":"Establishes contrast dependence of H1's response and the use of spike-triggered analysis for motion adaptation in this cell."},{"cited_title":"Rieke, D","cited_arxiv_id":null,"evidence_quote":"Provides the 'spikes as events' perspective and the one-spike-per-characteristic-time regime that the τint≈1/r̄ result connects to."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows a low-dimensional adaptive structure in salamander retinal responses, the closest comparative precedent for the rank-two result."}],"review_version":1}