{"id":"18b8dc5c-3970-435a-a00b-4438b2f3b530","arxiv_id":"2505.10371","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ILIF, a leaky integrate-and-fire neuron with membrane potential and current inhibition, lowers firing rates and improves accuracy on CIFAR and neuromorphic benchmarks, though theoretical support is incomplete.","lead":"This paper introduces ILIF, a spiking neuron model with two inhibitory units that reduce excessive firing and preserve gradient flow during training. If the results hold, SNNs can use larger surrogate gradient widths without overactivation, improving accuracy and energy efficiency.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3's gradient-shortcut identity (Eq. 23) is not a valid chain-rule collapse; the theoretical support for ILIF's gamma-dilemma claim is therefore unproven.","rationale":"The reader's weakest assumption identifies exactly the same bottleneck: Eq. (23) in Theorem 3 is an unsupported chain-rule cancellation. This concern is load-bearing because the abstract's theoretical contribution is the claim that ILIF overcomes the gamma dilemma by preserving gradient flow, and Theorem 3 is the only theorem dedicated to that claim. If Eq. (23) fails, the central explanation for ILIF's advantage is unsupported, although the empirical results remain relevant. The paper does provide code, a reproducible experimental setup, and ablations, which are positive evidence for the empirical side. However, the accuracy gaps over CLIF are small (e.g., 95.49 vs 95.41 on CIFAR10, 78.51 vs 78.36 on CIFAR100), and no multi-seed variance is reported, so the empirical case alone is not enough to fully establish the theoretical narrative. The energy inconsistency on DVSGesture in Table 5 further weakens the abstract's energy-efficiency claim. For these reasons, the reader's CONDITIONAL verdict is appropriate: the paper should be accepted only if the gradient-flow proof is corrected or the theoretical claims are scaled back to the observed empirical phenomena. My stress-test does not change that verdict.","tokens_in":18374,"tokens_out":7552,"duration_ms":70412,"concrete_test":"Independently re-derive the temporal gradient of a single ILIF neuron for T=3 from Eq. (17), or compute it by automatic differentiation of the exact computation graph, and compare it with the shortcut formula in Eq. (23). For example, choose a scalar loss equal to the final membrane potential Ul[3], set λ_U=1 and λ=1, and check whether the product ∂Ul[2]/∂Ul[1] · ∂Ul[1]/∂Sl[1] · ∂Sl[1]/∂Ul[1] equals ∂Ul[2]/∂Sl[1] · ∂Sl[1]/∂Ul[1]. If the two disagree, or if the recursive BPTT expansion of Eq. (37) contains ε=0 factors while Eq. (23) predicts a nonzero term for γ=V_th, then the no-attenuation claim is not demonstrated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim is that ILIF preserves gradient flow through a no-decay shortcut, and on this basis Theorem 3 concludes that ILIF resolves the gamma dilemma. The proof reduces to Eq. (23), where a product of intermediate Jacobians from time T-1 down to t is collapsed to ∂Ul[T-1]/∂Sl[t] by appending ∂Ul[t]/∂Sl[t]·∂Sl[t]/∂Ul[t]. But the ILIF dynamics in Eq. (17) are not a simple composition along one variable chain: future membrane potentials also depend on Sl[t] through \\bar ml[t], ml[t], and the inhibitory accumulator Ul[t], so ∂Ul[T-1]/∂Sl[t] is not generally equal to the telescoped product times ∂Ul[t]/∂Sl[t]. The intermediate partial derivatives do not cancel as written. Consequently Eq. (23) does not establish a no-attenuation path. Appendix C's own recursive expansion, Eq. (37), still multiplies every additional term by products of λε_l[t], the same decay factors that vanish for small γ, and Appendix D asserts ∂ml[t']/∂Ul[t'] < 0 without deriving it from Eq. (17). If Eq. (23) is invalid, the theoretical guarantee that ILIF overcomes the gamma dilemma is unsupported; what remains is the empirical firing-rate and accuracy evidence, which is suggestive but lacks error bars and, on DVSGesture, Table 5 shows ILIF consuming more energy than LIF (1034.6 vs 961.1 µJ), contradicting the abstract's energy-efficiency claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper identifies a trade-off in the surrogate gradient support width γ: large γ leads to overactivation and excessive firing, whereas small γ causes vanishing temporal gradients. The authors propose an ILIF neuron model with two interconnected inhibitory units, MPIU and CIU, and report experiments on CIFAR10, CIFAR100, DVSCIFAR10, and DVSGesture showing higher accuracy, lower firing rates, and in most cases lower energy consumption than LIF. They further claim a theoretical analysis (Lemmas 1–2 and Theorems 1–3) proving that ILIF overcomes the γ dilemma by reducing firing rates while preserving gradient flow.","tokens_in":18860,"tokens_out":7383,"duration_ms":75413,"significance":"The empirical part of the paper is valuable: the reported accuracies are competitive with recent SNN methods, the ablation in Table 2 supports the contribution of each inhibitory unit, Table 3 suggests the mechanism transfers to PLIF, and the code is made available. If the theoretical claims were adequately supported, the paper would make a clear conceptual contribution to SNN neuron design. However, the theoretical core as written does not establish the central claims: the proof of the no-decay gradient shortcut is invalid, Lemma 1 contains a variable slip and an unsupported weight-norm argument, Theorem 2 assumes the conclusion through the definition of its auxiliary functions, and the energy-efficiency claim is contradicted by the paper's own Table 5 on DVSGesture. The manuscript is therefore not acceptable in its current form, but the strengths of the empirical study justify a major revision rather than rejection.","major_comments":[{"comment":"The no-decay shortcut identity in Eq. (23) is not a valid chain-rule collapse for the ILIF dynamics. In Eq. (17), the membrane potential at a later time step depends on S_l[t] not only through U_l[t] but also through \\bar m_l[t], the MPIU accumulator, and the CIU state, so ∂U_l[T−1]/∂S_l[t] is not equal to the telescoped product multiplied by ∂U_l[t]/∂S_l[t]; intermediate Jacobians do not generally cancel. Appendix C's expansion in Eq. (37) and Appendix D's Eq. (43) show that every additional term is still multiplied by products of λε_l[t], the same attenuation factors that vanish for small γ, so no attenuation-free pathway is exhibited. Consequently, Theorem 3's claim that ILIF preserves gradient flow through a direct shortcut is not established by the provided derivation.","section":"Section 4.3, Eq. (23)"},{"comment":"The proof of Lemma 1 has a variable slip and a non-sequitur. It starts with γ_2 > γ_1 but then states that 'As λ increases from λ1 to λ2'; if λ is intended, it is never linked to the γ comparison. More importantly, the proof asserts without derivation that increasing the support width causes previously inactive neurons to re-enter the update interval and that this increases the equilibrium weight norm. The sign annotations in Eq. (12) do not supply the missing argument: the factor (Y_L[t′] − \\hat Y[t′]) is not nonnegative in both cases, and no monotonicity of the equilibrium weight norm in γ is shown. Since Lemma 1 is used to define the core 'dilemma of γ', this needs to be repaired.","section":"Section 4.1, Lemma 1"},{"comment":"Theorem 1 relies on two unproven sign properties: the statement that the additional gradient b(W) 'shares the same sign as W', and the claim in Appendix D, Eq. (43), that ∂m_l[t′]/∂U_l[t′] < 0. These properties are not derived from Eq. (17). Because the notation in Eq. (17) overloads U_l[t] for both the membrane potential and the MPIU accumulator, the sign of ∂m_l[t′]/∂U_l[t′] depends on which variable is meant and on the value of S_l[t′]; it is not negative in general. The assertion that a(W) 'typically' opposes the sign of W is also heuristic rather than a proof. Thus the inequality ∥W″∥ > ∥W′∥ is not established.","section":"Section 4.3, Theorem 1 and Appendix D"},{"comment":"The proof of Theorem 2 assumes the conclusion through the construction of the auxiliary functions. The functions F, R, and J are defined with exactly the properties needed to force r″ < r′: R and J are strictly increasing, J(r′) = 0, and H(r) = R(F(∥W″∥) − J(r)) is therefore decreasing. These properties are stipulated rather than derived from the ILIF equations, so the theorem restates the intended outcome as an assumption. In particular, J(r′) = 0—that inhibition vanishes exactly at the LIF equilibrium firing rate—is not established by Eqs. (13)–(17). The theorem therefore does not provide an independent analysis of the ILIF dynamics.","section":"Section 4.3, Theorem 2"},{"comment":"The abstract claims that ILIF 'improves energy efficiency by reducing firing rates', but Table 5 reports that on DVSGesture ILIF consumes 1034.647 µJ versus 961.133 µJ for LIF, i.e., ILIF is less energy-efficient on that dataset. The text in Appendix H acknowledges this, but the abstract and conclusion do not. Since energy efficiency is presented as a principal contribution, the claim must either be restricted to the datasets and settings where it holds, or the paper must quantify when the doubled MAC cost of ILIF is not offset by reduced spike counts.","section":"Section 5.4, Table 5 and Abstract"}],"minor_comments":[{"comment":"The same symbols U_l[t] and I_l[t] are used for both the membrane potential and the input current and for the MPIU and CIU accumulators, respectively. This overloading makes the derivations in Section 4.3 and Appendix D very difficult to follow; please use distinct symbols for the inhibitory accumulators.","section":"Equations (13) and (17)"},{"comment":"Please replace 'λ increases from λ1 to λ2' with 'γ increases from γ1 to γ2', or define λ explicitly if it is meant to be a different quantity.","section":"Section 4.1, Lemma 1 proof"},{"comment":"The underbrace annotations in Eq. (12) should be corrected: the factor (Y_L[t′] − \\hat Y[t′]) is a signed difference, not a nonnegative quantity, and the sign analysis needs to be written out separately for the cases \\hat Y[t′] = 1 and \\hat Y[t′] = 0.","section":"Eq. (12)"},{"comment":"The pseudocode is titled 'Main Fire Procedure for LIF Model', which appears to be a leftover from a LIF version; it should be titled for ILIF. The heading 'Pseudocode of ILIF mdoel' also contains a typo.","section":"Algorithm 1 and Appendix E"},{"comment":"The text states 'with λ = 1 and τ = 1.1', but τ does not appear elsewhere in the experimental setup and is not defined. Please either remove it or state what it denotes.","section":"Section 5.1 and Table 4"},{"comment":"Only single-run results are reported, with no standard deviations or number of seeds. Given that several gains over baselines are small (e.g., 95.49 vs. 95.41 on CIFAR10) and that DVSGesture ties at 97.92, adding repeated-run statistics or at least stating that results are single runs would strengthen the empirical claims.","section":"Tables 1–3 and Figures 3–5"}],"recommendation":"major_revision","confidential_remarks":"The paper's theoretical claims are the main obstacle. The empirical study is promising, but the current proofs do not support the advertised 'theoretical analysis demonstrates ILIF effectiveness' narrative. I recommend the editor request a major revision in which the theory is either substantially corrected or explicitly downgraded to a heuristic discussion, and in which the energy-efficiency claim is made consistent with Table 5."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The empirical core of this paper is worth engaging with. The ILIF neuron, with its separate membrane-potential inhibitory unit (MPIU) and current inhibitory unit (CIU), is a genuine variation on the CLIF/TC-LIF/PLIF line, not a copy. The ablations in Table 2 show both units contribute, and the consistent firing-rate reductions across layers and datasets are real evidence that the mechanism does something useful. The accuracy numbers are competitive, and the code is available. I would be comfortable citing this for the architectural idea and the empirical firing-rate results.\n\nThe soft spots are in the theoretical apparatus, and they are not minor. Lemma 1 slips between gamma and lambda, and the weight-norm argument is heuristic rather than a proof. Theorem 2 essentially assumes monotone functions F, R, J whose properties encode the conclusion, so it does not explain why ILIF works. The bigger problem is Theorem 3: Eq. (23) collapses a product of Jacobians into a single partial derivative, but the ILIF dynamics do not create a simple chain where those intermediate derivatives cancel. The MPIU and CIU introduce dependencies through m-bar, m, and the inhibitory accumulator, so the shortcut is not valid as written. That means the central claim that ILIF provably overcomes the gamma dilemma is unsupported. The empirical section does not depend on this theory, but the paper sells the theory as a contribution, so it needs to be fixed or explicitly downgraded.\n\nTwo smaller issues: there are no error bars or multi-seed results, and Table 5 shows ILIF consuming more energy than LIF on DVSGesture (1034.6 vs 961.1 µJ), which contradicts the abstract's energy-efficiency claim. The energy story should be reworded or that number addressed.\n\nWho is this for? Researchers working on directly-trained SNNs who care about neuron-level inhibition and firing-rate reduction. The empirical part is useful and reproducible, and the theory, once corrected, may add value. I would send it to a serious referee, but with the expectation of major revision: the proofs need to be rewritten or removed, and the empirical claims need error bars. The paper is not a take-down; it is a solid empirical contribution wrapped in over-claimed theory.","headline":"Empirically solid neuron model with a genuinely new inhibitory architecture, but the theoretical proof that it resolves the gamma dilemma does not hold up as written.","tokens_in":19278,"tokens_out":1187,"would_cite":true,"duration_ms":14337,"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":"Two temporally coupled inhibitory units added to the leaky integrate-and-fire neuron are claimed to dissolve the surrogate-gradient trade-off between overactivation and vanishing gradients, cutting spike counts while raising accuracy on…","keywords":["spiking neural networks","surrogate gradients","leaky integrate-and-fire","neuronal inhibition","overactivation","gradient vanishing","energy efficiency","temporal dynamics"],"falsifier":"Compute the exact backward pass for a single ILIF neuron over $T$ time steps (or a tiny two-neuron network) and compare the true gradient $\\partial L/\\partial U^l[t]$ with the no-decay product asserted in Eq. (23); if the ratio of the two shrinks as $T$ grows, the shortcut does not exist. A cheaper experiment: train ILIF with $\\gamma = V_{\\mathrm{th}}$, the regime where Lemma 2 kills every temporal LIF gradient via $\\epsilon = 1 - V_{\\mathrm{th}}/\\gamma = 0$; if ILIF accuracy also collapses there, its inhibitory pathways do not actually bypass the $\\gamma$-limited product.","tokens_in":18222,"feed_emoji":"🧠","tokens_out":13623,"duration_ms":120602,"temperature":0.7,"pith_summary":"Spiking neural networks trained with surrogate gradients face a hidden dial, $\\gamma$, the width of the gradient's support around the firing threshold: turn it up and neurons over-fire, wasting energy; turn it down and gradients vanish across time. This paper claims the dilemma can be escaped by changing the neuron rather than the dial. The proposed ILIF neuron adds two interconnected inhibitory units, one that accumulates post-spike membrane potential to damp long-term excitability and one that feeds a neuron's own spike back to suppress incoming current, so firing naturally settles below the overactivation level while gradient flow is preserved. If the claim holds, energy efficiency and learnability need not be traded against each other: the model reports fewer spikes, more stable training, and higher accuracy (95.49% on CIFAR10, 78.51% on CIFAR100, 78.60% on DVSCIFAR10, 97.92% on DVSGesture).","feed_headline":"Inhibitory neuron makes spiking nets fire less, learn better","feed_subtitle":"Two temporal inhibitory units dissolve the trade-off between overactivation and vanishing gradients in spiking networks.","key_machinery":"The carrying object is the ILIF neuron: a vanilla leaky integrate-and-fire core wrapped in two recurrent inhibitory units. The membrane-potential inhibitory unit (MPIU) integrates post-spike potential into a slowly decaying state $\\bar{U}^l[t] = \\lambda_U\\bigl(\\bar{U}^l[t-1] + S^l[t]\\cdot\\bar{m}^l[t]\\bigr)$ and subtracts a bounded sigmoid of that state, $S^l[t]\\cdot\\sigma\\bigl(\\bar{U}^l[t]\\bigr)$, from the reset potential, mimicking afterhyperpolarization for long-term suppression. The current inhibitory unit (CIU), $\\bar{I}^l[t] = \\lambda_I\\bigl(\\bar{I}^l[t-1] + S^l[t]\\cdot I^l[t]\\bigr)$, decays quickly and subtracts the previous time step's current from the incoming synaptic current, mimicking fast retrograde feedback. These units do two jobs at once: forward, they suppress firing once activity accumulates; backward, their interconnections add the extra gradient terms $\\varphi^l[t]$ to the backpropagation-through-time expansion and, per Theorem 3, provide a shortcut intended to carry temporal gradients without decay.","core_discovery":"The paper's central claim is that temporal inhibition, not surrogate tuning, resolves the $\\gamma$ dilemma. With the two inhibitory units in place, a network can settle at equilibrium weights of larger magnitude than vanilla LIF (Theorem 1) while its equilibrium firing rate is strictly lower (Theorem 2), because inhibition grows with firing history and offsets the stronger excitatory drive. The paper further claims these units create direct backward shortcuts through time, so gradients travel from late to early time steps without the attenuating factor $1 - V_{\\mathrm{th}}H'(U)$ that otherwise vanishes at small $\\gamma$ (Theorem 3). Empirically, ILIF matches or exceeds prior neuron models on four datasets while firing less, including a more-than-30% reduction in continuous firing rate on DVSGesture, which the paper reads as evidence that the inhibition mechanism, not gradient tuning, carries the improvement.","pith_inferences":["Editorial extension: ILIF front-loads spiking activity (the first time step fires more, later steps less), so a testable consequence is that inference could be truncated after the first few time steps with little accuracy loss, enabling early-exit SNNs the paper does not discuss.","Editorial extension: the inhibition acts as an activity-dependent self-regulator; freezing or ablating the MPIU saturation curve would test whether stability comes from the bounded sigmoid or from the gradient shortcut.","Editorial extension: a LIF baseline with its threshold raised to match ILIF's firing rate would isolate the firing-rate effect from the gradient-pathway effect, separating regularization from the claimed shortcut mechanism.","Editorial extension: because ILIF's benefit is claimed to grow as inhibition accumulates, the mechanism should matter most for long sequences and deep networks; measuring accuracy and firing-rate gaps versus sequence length is a direct stress test."],"forward_implications":["If ILIF works as claimed, SNN training no longer needs to tune $\\gamma$ against the firing-versus-vanishing trade-off: accuracy stays stable across $\\gamma$ values while firing rates stay low.","The accuracy gains (95.49% CIFAR10, 78.51% CIFAR100, 78.60% DVSCIFAR10, 97.92% DVSGesture) come with fewer spikes, which the paper's energy model translates into lower estimated consumption per forward pass on three of the four datasets.","The inhibitory units are neuron-local and add no learnable parameters beyond the two decay constants, so they can be dropped into other spiking neurons; the paper's IPLIF variant improves on PLIF across all four datasets.","Theorems 1 and 2 predict a decoupling of weight magnitude from firing rate, equilibrium weights grow while firing rates fall, which is a distinctive signature that other overactivation controls such as adaptive thresholds or normalization would not produce.","Training stabilizes: loss curves are smoother and converge lower than LIF, consistent with the claimed extra gradient pathways and with the reduction in overactivation."],"supporting_citations":[{"why":"Provides the analysis that small $\\gamma$ risks gradient vanishing that Lemma 2 extends, and supplies CLIF, the closest accuracy baseline ILIF is compared against.","marker":"[Huang et al., 2024]"},{"why":"Defines surrogate-gradient learning, the training setting in which the $\\gamma$ dilemma arises and which ILIF modifies.","marker":"[Neftci et al., 2019]"},{"why":"One of the biological sources for afterhyperpolarization-style inhibition that the MPIU unit instantiates.","marker":"[Bellec et al., 2018]"},{"why":"Supplies the retrograde, GABAergic feedback inspiration for the CIU current inhibition.","marker":"[Zenke and Ganguli, 2018]"},{"why":"Provides PLIF, the neuron model used to show the inhibitory units transfer beyond LIF through the IPLIF results.","marker":"[Fang et al., 2021b]"},{"why":"Establishes the $\\gamma \\ge V_{\\mathrm{th}}$ convention whose regime bounds $1 - V_{\\mathrm{th}}/\\gamma$ between 0 and 1 in the Lemma 2 vanishing-gradient proof.","marker":"[Wu et al., 2019]"},{"why":"Backpropagation through time is the training algorithm whose temporal gradient expansion carries the entire theoretical analysis.","marker":"[Werbos, 1990]"}],"fun_headline_variants":["Temporal inhibition cracks SNN firing-rate dilemma","Inhibitory neuron halts overactivation, boosts gradient flow","Spiking neurons get a brake: ILIF cuts firing, preserves learning","Double inhibition resolves SNN's gamma trade-off","Temporal inhibitory neuron trims SNN firing, not accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theoretical resolution of the gradient-vanishing half of the $\\gamma$ dilemma rests on Theorem 3's assertion that the backward chain through the inhibitory shortcut collapses to the single product $\\frac{\\partial U^l[T-1]}{\\partial S^l[t]}\\frac{\\partial S^l[t]}{\\partial U^l[t]}$ with no attenuation; that chain-rule cancellation is stated, not derived, and is not generally valid, so if it fails the claim that ILIF preserves gradient flow is no longer established by the theory, and the $\\gamma$ dilemma is resolved only empirically.","fun_headline_variants_meta":{"raw":{"variants":["Temporal inhibition cracks SNN firing-rate dilemma","Inhibitory neuron halts overactivation, boosts gradient flow","Spiking neurons get a brake: ILIF cuts firing, preserves learning","Double inhibition resolves SNN's gamma trade-off","Temporal inhibitory neuron trims SNN firing, not accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000565,"raw_usage":{"total_tokens":2670,"prompt_tokens":930,"completion_tokens":1740,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":1658}},"tokens_in":546,"tokens_out":1740,"duration_ms":12172,"temperature":1.0,"reasoning_tokens":1658,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:10:27.497670+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the exact backward pass for a single ILIF neuron over $T$ time steps (or a tiny two-neuron network) and compare the true gradient $\\partial L/\\partial U^l[t]$ with the no-decay product asserted in Eq. (23); if the ratio of the two shrinks as $T$ grows, the shortcut does not exist. A cheaper experiment: train ILIF with $\\gamma = V_{\\mathrm{th}}$, the regime where Lemma 2 kills every temporal LIF gradient via $\\epsilon = 1 - V_{\\mathrm{th}}/\\gamma = 0$; if ILIF accuracy also collapses there, its inhibitory pathways do not actually bypass the $\\gamma$-limited product.","supporting_citations":[{"cited_title":"Clif: Complementary leaky integrate-and- fire neuron for spiking neural networks.ICML,","cited_arxiv_id":null,"evidence_quote":"Provides the analysis that small $\\gamma$ risks gradient vanishing that Lemma 2 extends, and supplies CLIF, the closest accuracy baseline ILIF is compared against."},{"cited_title":"Long short-term memory and learning-to-learn in net- works of spiking neurons.NeurIPS, 31,","cited_arxiv_id":null,"evidence_quote":"One of the biological sources for afterhyperpolarization-style inhibition that the MPIU unit instantiates."},{"cited_title":"Superspike: Supervised learning in multi- layer spiking neural networks.Neural computation, 30(6):1514–1541,","cited_arxiv_id":null,"evidence_quote":"Supplies the retrograde, GABAergic feedback inspiration for the CIU current inhibition."},{"cited_title":"Backpropagation through time: what it does and how to do it.Proceedings of the IEEE, 78(10):1550–1560,","cited_arxiv_id":null,"evidence_quote":"Backpropagation through time is the training algorithm whose temporal gradient expansion carries the entire theoretical analysis."}],"review_version":1}