{"id":"6a4d8cd8-95f2-418a-8c04-9181d6320f90","arxiv_id":"1908.03886","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A simulation study reports that recurrent networks with synapses randomly assigned as excitatory or inhibitory can encode time-varying inputs in population firing rate, and that intermediate recurrent probability and I/E strength ratios optimize encoding fidelity.","lead":"This paper simulates a recurrent spiking network in which each neuron's outgoing synapses are randomly assigned excitatory or inhibitory, rather than fixing each neuron's type. It finds that such 'undetermined-type' networks encode time-varying stimuli in population firing rate, with intermediate recurrence and moderate inhibition-to-excitation ratio giving best fidelity.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 2.1 avoids same-target E+I input, so the random 80/20 sign assignment tests per-synapse sign, not co-release; the claimed biological advantage over fixed-E/I networks is therefore not established.","rationale":"The reader's weakest_assumption is on target. The paper's central claim as stated in the abstract and conclusion goes beyond 'a random-sign network encodes well'; it claims the biological phenomenon of co-release of excitatory and inhibitory receptors makes such a network rational. However, the model's only deviation from a standard E/I network is that the sign of each directed synapse is drawn independently, and the model explicitly excludes the defining same-target feature of co-release. Thus the reported high Q values and the Fig. 10 advantage over a fixed-E/I network are real computational observations but cannot be attributed to co-release without an additional model variant. This is a logical gap rather than a numerical error; the internal consistency of the simulations is not in question. The appropriate disposition is the reader's CONDITIONAL: accept only if code/data are made available and this co-release assumption is addressed. My stress-test does not change that verdict, so UNCHANGED.","tokens_in":14078,"tokens_out":9651,"duration_ms":108745,"concrete_test":"Add a true co-release variant to the Section 3.3 comparison (same HH parameters, same Prc, R, D, gexc, tau): for each directed connection, with probability p the presynaptic spike activates both an excitatory conductance (Vsyn=0 mV) and an inhibitory conductance (Vsyn=-75 mV) on the same postsynaptic neuron, with the number of single-sign connections reduced so the marginal 4:1 E/I balance is preserved; simulate p=0 (current model), p=0.25, and p=0.5 and recompute Q for the R sweep. If Q does not increase with p, or if the advantage over the fixed-E/I network (Fig. 10) vanishes, then the paper's co-release conclusion is not supported by the model. Separately, rerun the main R sweep with PE_rc=0.7 and 0.9 to check that the claimed optimality of the 80/20 ratio is not an artifact of that single choice.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and introduction frame the work as motivated by receptor-determined synaptic polarity and by co-release of excitatory and inhibitory receptors in the same synapse (refs. 41-42), and the Conclusion says the results make it 'rational to consider' co-release. But the implemented network in Section 2.1 assigns each recurrent connection an independent 80% excitatory / 20% inhibitory sign and explicitly states that 'one postsynaptic neuron receives the excitatory and inhibitory synapse from the same presynaptic neuron at the same time is avoided.' That is a random per-synapse sign model ('undetermined-type'), not a co-release model: co-release requires the same presynaptic terminal to activate both receptor classes on the same postsynaptic neuron. No mechanism is given for why a neuron should be excitatory to some targets and inhibitory to others with exactly a 4:1 probability, and no sensitivity analysis varies this ratio. The paper's own conclusion admits the model is 'not very biological enough.' Consequently, the central comparison in Section 3.3 / Fig. 10 may demonstrate only that this particular random edge-coloring outperforms a source-determined E/I coloring; it does not demonstrate that receptor-determined polarity or co-release improves population-rate coding.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies population-rate coding in a recurrent Hodgkin-Huxley network in which each directed recurrent synapse is randomly assigned an excitatory (80%) or inhibitory (20%) sign independently of the presynaptic neuron's identity, which the authors call an 'undetermined-type' network. The stimulus is a half-wave-rectified Ornstein-Uhlenbeck process, and the encoding quality Q is the maximal correlation between the stimulus and the population firing rate at an optimal lag. Parameter scans over the I/E strength ratio R, recurrent probability Prc, noise intensity D, excitatory strength gexc, and synaptic time constant reveal an optimal regime around Prc ≈ 0.1 and 15 < R < 40 with Q above 0.9. The paper concludes that this undetermined network encodes better than a standard fixed-E/I network and interprets this as evidence that co-release of excitatory and inhibitory receptors should be considered.","tokens_in":14454,"tokens_out":7719,"duration_ms":80239,"significance":"If the computational results hold, the paper provides a systematic computational demonstration that edge-level random E/I sign assignment can be as good as or better than source-determined E/I assignment for population-rate encoding in recurrent Hodgkin-Huxley networks. The parameter sweeps are extensive and the qualitative inverted-U dependencies are coherent. The significance is, however, limited by the mismatch between the implemented model and the claimed biological motivation, and by the absence of quantitative uncertainty in the comparison to the determined network. The work would be a solid contribution if presented as a study of 'random-sign' or 'undetermined-type' connectivity rather than as evidence for co-release.","major_comments":[{"comment":"The implemented network assigns each recurrent connection an independent 80/20 E/I sign and explicitly states that one postsynaptic neuron never receives both E and I from the same presynaptic neuron; this is a per-synapse sign model, not a co-release model in which one terminal activates both receptor classes on the same target. Consequently, the abstract's and conclusion's claim that the results make it 'rational to consider the co-release of inhibitory and excitatory receptors' is not supported by the simulation. Please either simulate co-release explicitly or replace the co-release interpretation with the more modest claim about random, undetermined-type sign assignment, and add a sensitivity analysis for the 4:1 E/I ratio.","section":"Section 2.1 and Section 4"},{"comment":"The central claim that the undetermined network outperforms the determined network is presented without error bars, trial counts, or statistical testing, and the figure legend in 10(b) is internally inconsistent ('gexc=0.08 gexc=0.1 gexc=0.08 gexc=1'). Given that the stated differences are small (roughly 0.85-0.9), the reader cannot assess whether the advantage is reliable. Please report mean ± SD over at least the 10 trials used elsewhere and state the exact parameters for both models.","section":"Section 3.3, Fig. 10"},{"comment":"There is an inconsistency in how the population firing rate is defined: Eq. (7) sets Δt=1 ms, while the correlation in Eq. (8) is computed with a 10 ms window sliding by 1 ms. Also, Eq. (7) as written gives spikes/ms, but Fig. 2 labels the ordinate in Hz. Since Q is a correlation coefficient it is invariant to this scale, but the manuscript should state the exact time bin used to generate the figures and provide the conversion factor; otherwise the simulations are not fully reproducible.","section":"Section 2.4, Eqs. (7)-(8)"},{"comment":"The text states that 'the product of R and gexc has a positive relationship with the encoding quality,' but the surrounding discussion and the figure indicate a non-monotonic (inverted-U) relationship, with both too-large and too-small inhibitory strength reducing Q. Please correct this statement to reflect the described optimum at intermediate values of R*gexc.","section":"Section 3.2.2, Fig. 7"}],"minor_comments":[{"comment":"There are numerous typos and cross-reference errors, including 'thses', 'hypthesis', 'In other word', and a reference to 'Fig. 5(b)' that should be 'Fig. 6(b)'.","section":"Throughout"},{"comment":"The synchronization measure k_{i,j} is not normalized by the number of time bins, and this quantity is never used in the Results; please either correct the formula or remove it to avoid confusing an undefined measure.","section":"Section 2.4.3, Eq. (10)"},{"comment":"The phrase 'the relationship of Q versus Prc around R≈0.1' appears to be a typo for 'Q versus R around Prc≈0.1,' since the curves shown in Fig. 4(c) are Q versus R for fixed values of Prc.","section":"Section 3.1"},{"comment":"The legend text 'gexc=1' is presumably a typo for 'gexc=0.1'; this should be corrected because the comparison of synaptic strength is a central element of the argument.","section":"Section 3.3, Fig. 10(b)"}],"recommendation":"major_revision","confidential_remarks":"The main computational finding may become publishable after reframing the claims from co-release to random-sign connectivity and after adding error bars to the comparison with the determined network. As submitted, the biological interpretation is overstated and the key comparison is underpowered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing you should know: the novel twist is that each neuron can have both excitatory and inhibitory outgoing synapses, assigned randomly with an 80/20 split. That is a genuine variant of the standard fixed-E/I network, and the paper shows it can encode time-varying stimuli in population firing rate with high fidelity. But the abstract and introduction frame this as motivated by receptor-determined polarity and co-release, and that framing is not what the model does. Section 2.1 explicitly says the same postsynaptic neuron does not receive both excitatory and inhibitory synapses from the same presynaptic neuron. So the simulation is a random per-synapse sign assignment, not a co-release model. The stress-test note lands on reading the paper, and the authors themselves admit the architecture is \"not very biological enough.\"\n\nWhat the paper does well: it runs a systematic scan over the I/E strength ratio, recurrent probability, noise intensity, excitatory strength, and synaptic time constants, using Hodgkin-Huxley neurons with conductance-based synapses. The qualitative results—intermediate recurrence around Prc ≈ 0.1, moderate I/E ratio, and intermediate noise promoting encoding—are internally consistent and extend known results from fixed E/I networks. The direct comparison with a determined-E/I model is useful, even if the conclusions drawn from it are overreach.\n\nThe soft spots, in order of seriousness. First, the co-release claim. Because the model avoids same-target E+I input, Fig. 10 only demonstrates that this random edge-coloring outperforms a source-determined coloring at particular parameters. It says nothing about whether co-release improves coding. That needs to be retracted or heavily qualified. Second, the paper is not reproducible from the manuscript alone: no code, no data, and Fig. 10 curves lack error bars. For a computational parameter study, that is a meaningful gap. Third, Eq. (7) defines population rate as N(Δt)/Δt with Δt = 1 ms, which is spikes/ms, while the figures label Hz with no conversion stated. That is minor but should be fixed. Fourth, the optimal parameter ranges are read from the same surfaces used to phrase the conclusions; that is mild circularity, common in this kind of study, but a sensitivity analysis on the 4:1 E/I ratio would strengthen the claims.\n\nNone of this sinks the core modeling result. The paper is a modest but legitimate contribution to the population-rate-coding literature. It deserves a serious referee, but the referee should insist on code/data, units clarification, a toned-down biological interpretation, and at least one sensitivity run on the sign-assignment ratio. If those are addressed, it is publishable in a specialty journal; if not, it is a desk-reject candidate.\n\nMy recommendation: send to peer review, but with a clear request for major revisions.","headline":"A competent but modest parameter study of a random E/I edge-coloring network, whose stated biological motivation (co-release) is not actually implemented; the encoding trends are plausible but the paper overclaims the model's relationship to co-release.","tokens_in":14889,"tokens_out":2045,"would_cite":false,"duration_ms":24279,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92C20","92B20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A recurrent network whose neurons can send both excitatory and inhibitory synapses represents time-varying stimuli in population firing rate with peak correlation above 0.92 and outperforms a fixed-type E/I network.","keywords":["population rate coding","recurrent neuronal network","undetermined-type neurons","excitatory-inhibitory balance","Hodgkin-Huxley model","co-release","encoding fidelity","synaptic strength ratio"],"falsifier":"Run the identical encoding protocol on a matched network in which inhibitory postsynaptic effects are delivered only by a fixed set of 20 inhibitory neurons, keeping the same number of excitatory and inhibitory synapses, and compare Q; if the undetermined network no longer shows higher encoding fidelity, the paper's central comparison fails. Similarly, if allowing one presynaptic neuron to excite and inhibit the same postsynaptic cell causes Q to drop below the paper's exclusion rule, the biological motivation would be contradicted.","tokens_in":13885,"feed_emoji":"🧠","tokens_out":7549,"duration_ms":79984,"temperature":0.7,"pith_summary":"This paper tries to establish that a recurrent spiking network in which a neuron's outgoing synapses are randomly typed excitatory or inhibitory rather than the neuron itself being typed can faithfully represent a time-varying stimulus in its population firing rate, and does so better than a traditional determined-type network. It reports an optimal working point: intermediate recurrent probability near Prc = 0.1 and a moderate ratio of inhibitory to excitatory synaptic strength, roughly 15 < R < 40, give the highest encoding fidelity, with peak correlation values above 0.92. The same intermediate-tuning pattern appears for background noise, excitatory synaptic strength, and synaptic time constant, where moderate noise, moderate excitatory strength, and fast synapses all help. A sympathetic reader would care because the result suggests that receptor-determined and co-released synaptic polarity, a physiologically observed arrangement, can be a workable architectural principle for rate coding rather than a biological detail to be averaged away.","feed_headline":"Undetermined-type neurons beat fixed E/I networks at rate coding","feed_subtitle":"In Hodgkin-Huxley simulations, letting one neuron send both excitatory and inhibitory signals raises coding fidelity.","key_machinery":"The key machinery is the undetermined-type assignment rule: each synapse is independently labelled excitatory with probability 0.8 or inhibitory with probability 0.2, so synaptic sign is treated as a property of the activated receptor rather than of the presynaptic neuron. The argument is carried by the ratio R = g_inh/g_exc, which controls whether the recurrent population is in an excitation-dominated or inhibition-dominated regime, together with the recurrent probability Prc, which fixes how many mixed-sign feedback paths exist. The encoding quality Q, defined as the maximum of the cross-correlation between the input signal and the population firing rate, is the measure that exposes the optimum: Q is high only when R, Prc, noise, and excitatory strength place the population in an intermediate dynamical state rather than at either extreme.","core_discovery":"The central discovery is that population rate coding survives and is sharpened by removing the assumption that each neuron has a fixed excitatory or inhibitory type. In the proposed undetermined-type network, each recurrent connection is independently assigned as excitatory with probability 0.8 or inhibitory with probability 0.2, so one presynaptic neuron may excite some postsynaptic targets and inhibit others. Using the maximum correlation Q between a half-wave-rectified Gaussian input and the population firing rate as the measure of encoding quality, the paper finds Q reaches values above 0.92 for intermediate inhibitory-to-excitatory strength ratios (about 15 < R < 40) and recurrent probability near 0.1. The optimal values of noise intensity, excitatory synaptic strength, and synaptic time constant are likewise intermediate, meaning the network encodes best in a state that is neither too excitable nor too suppressed. Under matched parameters, this undetermined network attains higher Q than a conventional network with 80 excitatory and 20 inhibitory neurons, which the paper takes as evidence that co-release of excitatory and inhibitory receptors is a plausible and beneficial coding mechanism.","pith_inferences":["If the random-sign assignment is interpreted as receptor-determined polarity, the model predicts that target-specific differences in receptor expression, rather than presynaptic cell type, determine whether a connection excites or inhibits; this could be tested by comparing networks with different correlations between target identity and synapse sign.","The optimum near Prc = 0.1 may reflect a balance between averaging over many neurons and preserving stimulus transients: denser recurrence smooths away fast signal changes, while very sparse recurrence leaves too little shared activity to encode reliably.","A natural extension the paper does not simulate is true co-release, where one presynaptic neuron simultaneously excites and inhibits the same postsynaptic cell; whether opposing conductances cancel or sharpen the code is an open question that could raise or lower the reported advantage.","The 80/20 excitatory-to-inhibitory ratio is fixed throughout; testing sensitivity to other ratios, such as 70/30 or 90/10, would show whether the qualitative optimum is tied to the biological balance or is an artifact of the chosen 4:1 mix."],"forward_implications":["Population firing rate can remain a high-fidelity code even when individual neurons have no fixed excitatory or inhibitory identity, as long as the collective mix of synapses is balanced.","Encoding quality is non-monotonic in inhibitory strength: both too-weak and too-strong inhibitory feedback degrade the population code, so an intermediate E/I balance is required for faithful encoding.","Sparse recurrence around Prc = 0.1 is better than dense recurrence for this rate-coding task, suggesting that too many recurrent connections wash out the stimulus-related signal.","Moderate background noise and fast synaptic kinetics improve fidelity, while strong noise and slow synapses degrade it.","At identical parameters, the undetermined-type network outperforms the standard fixed 80/20 E/I network, implying that receptor-determined synaptic polarity is a rational modeling choice for population rate coding."],"supporting_citations":[{"why":"Defines the Hodgkin-Huxley neuron model whose equations drive all simulated neuronal dynamics.","marker":"[43]"},{"why":"Supplies the gating-variable rate functions used in the Hodgkin-Huxley simulations.","marker":"[44]"},{"why":"Provides the cross-correlation measure C(tau) from which the encoding quality Q is defined.","marker":"[12]"},{"why":"Supplies the population coherence measure used to quantify synchrony and characterize network state.","marker":"[45]"},{"why":"Establishes fast propagation of firing rates through noisy layered networks, the rate-coding framework this paper extends to recurrent mixed-sign networks.","marker":"[6]"},{"why":"Studies population rate coding in recurrent networks with unreliable synapses, the closest prior recurrent-network setting that this work modifies.","marker":"[28]"},{"why":"Reviews GABA receptor co-localization and cross-talk with other receptors, cited as physiological evidence that synaptic effect is receptor-determined.","marker":"[41]"},{"why":"Reviews glutamate and GABA receptor cross-talk, cited to support co-release and the importance of E/I balance in the brain.","marker":"[42]"}],"fun_headline_variants":["Undetermined neuron types sharpen rate coding","Mixed-signal synapses boost population coding","Flexible E/I roles improve coding fidelity","Co-release of receptors enhances rate coding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that randomly labelling each outgoing synapse as excitatory or inhibitory in a fixed 4:1 ratio, and never letting the same presynaptic neuron give both signs to the same postsynaptic cell, is a fair stand-in for receptor-determined co-release; if the real biological rule for synaptic sign differs, the reported encoding advantage belongs to the random-labelling rule rather than to the biology it is meant to capture.","fun_headline_variants_meta":{"raw":{"variants":["Undetermined neuron types sharpen rate coding","Mixed-signal synapses boost population coding","Flexible E/I roles improve coding fidelity","Co-release of receptors enhances rate coding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000257,"raw_usage":{"total_tokens":1612,"prompt_tokens":1011,"completion_tokens":601,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":549}},"tokens_in":627,"tokens_out":601,"duration_ms":7113,"temperature":1.0,"reasoning_tokens":549,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:59:35.491216+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical encoding protocol on a matched network in which inhibitory postsynaptic effects are delivered only by a fixed set of 20 inhibitory neurons, keeping the same number of excitatory and inhibitory synapses, and compare Q; if the undetermined network no longer shows higher encoding fidelity, the paper's central comparison fails. Similarly, if allowing one presynaptic neuron to excite and inhibit the same postsynaptic cell causes Q to drop below the paper's exclusion rule, the biological motivation would be contradicted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the cross-correlation measure C(tau) from which the encoding quality Q is defined."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes fast propagation of firing rates through noisy layered networks, the rate-coding framework this paper extends to recurrent mixed-sign networks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Studies population rate coding in recurrent networks with unreliable synapses, the closest prior recurrent-network setting that this work modifies."},{"cited_title":"Character randomized benchmarking for non-multiplicity-free groups with applications to subspace, leakage, and matchgate randomized benchmarking","cited_arxiv_id":"2011.00007","evidence_quote":"Reviews GABA receptor co-localization and cross-talk with other receptors, cited as physiological evidence that synaptic effect is receptor-determined."},{"cited_title":"Kantamneni, Cross-talk and regulation between glutamate and gabab receptors, Frontiers in Cellular Neuroscience 9 (2015) 135","cited_arxiv_id":null,"evidence_quote":"Reviews glutamate and GABA receptor cross-talk, cited to support co-release and the importance of E/I balance in the brain."}],"review_version":1}