acceptodds
Under review as a conference paper at ICLR 2027

Incorporate Self-Prediction to Enhance Dynamics of Spiking Neurons

Abstract

Spiking Neural Networks (SNNs) offer the potential for energy-efficient computation through event-driven, sparse processing, but their training is challenged by spike non-differentiability and trade-offs among performance, efficiency, and biological plausibility. Predictive coding motivates exploring how a neuron's own input–output history can modulate its subsequent dynamics. Inspired by this, we propose a self-prediction enhanced spiking neuron framework that generates an internal prediction current from its input–output history to modulate membrane potential, with separate coefficients for input memory and spike feedback. This design introduces additional gradient pathways through the auxiliary state and provides a local feedback mechanism inspired by dendritic modulation and neuronal self-feedback. Experiments with the reference feedback rules show performance gains in most evaluated configurations across diverse architectures, neuron types, time steps, and tasks. LIF ablations show that learning the spike-feedback coefficient further improves accuracy in two architectures.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Self-prediction, Spiking Neuron, Spiking Neural Networks

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_AVbROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a self-prediction enhanced spiking neuron framework that augments an existing spiking neuron with an auxiliary recurrent state. The auxiliary state integrates a weighted combination of the neuron’s historical input and its own spike output, and is then fed back into subsequent membrane integration.

Concretely, the method introduces an auxiliary state of the form

mp[t]=(1−τp)mp[t−1]+τp(ax[t]−bs[t]).m_p[t] = (1-\tau_p)m_p[t-1] + \tau_p(a x[t] - b s[t]).

which is added to the input current at the next time step. The coefficients aa and bb control input memory and spike feedback, respectively. The authors evaluate this mechanism across multiple neuron models (IF, LIF, PLIF, and CLIF), network architectures, image classification tasks, sequential CIFAR-10, and reinforcement learning environments.

The paper reports improvements in many CIFAR-10 configurations, modest gains on ImageNet, improvements on sequential CIFAR-10, and generally positive results on the tested reinforcement learning tasks. Additional ablations examine input memory, spike feedback, gradient detachment, and learnable feedback coefficients.

Critical strengths
  1. Broad empirical evaluation.
    The method is tested across several neuron types, network architectures, time steps, classification tasks, sequential classification, and reinforcement learning. This breadth provides useful evidence that adding the auxiliary state can improve performance in multiple settings.

  2. Reasonably thorough ablations.
    The paper separately studies input memory, spike feedback, gradient detachment, and learnable feedback coefficients. These ablations are valuable because they expose which components are actually contributing to the observed gains.

Critical weaknesses
  1. It is unclear what is actually being “predicted.”
    The proposed mechanism does not explicitly predict a future input, future spike, or latent target, nor does it optimize a prediction-error objective. Mathematically, the auxiliary state is an exponentially weighted history of the input minus an exponentially weighted history of the neuron’s own spikes. This appears closer to recurrent memory, adaptation, or autaptic feedback than to predictive coding or self-prediction. The paper should justify more clearly why the proposed mechanism should be interpreted as prediction rather than as an additional recurrent state.

  2. The main performance gain may come from input memory rather than the proposed spike-feedback mechanism.
    In the path ablations, input-memory-only already improves substantially over the base LIF neuron. Moreover, adding spike feedback provides only a small gain for SEW ResNet-18 and actually reduces the final-five-epoch accuracy relative to input-memory-only for Spiking ResNet-18. Therefore, the evidence does not clearly establish that the input-output residual or self-feedback component is necessary.

  3. The learnable spike-feedback coefficient is not consistently beneficial.
    Although the main text highlights gains from learning the spike-feedback coefficient on two architectures, the broader exploration over 20 architecture-neuron combinations shows improvements in 11 cases and decreases in 9 cases, with an average gain of only 0.0465 percentage points. This makes the claimed general benefit of adapting the feedback strength considerably less convincing.

  4. The method increases the neuron state dimension, but the paper does not provide a sufficiently strong capacity-matched baseline.
    A standard LIF neuron is augmented with an additional continuous recurrent state. Therefore, part of the improvement may simply come from increasing the temporal state capacity. A key missing baseline would be a simple auxiliary leaky state driven only by the input, with comparable parameter count and state dimension, but without the proposed input-output residual or predictive-coding interpretation.

  5. The relationship to existing adaptive and recurrent neuron models is not sufficiently disentangled.
    The proposed mechanism appears closely related to existing ideas such as adaptation currents, recurrent/self-feedback dynamics, autaptic connections, multi-timescale neurons, and auxiliary membrane states. Although these works are discussed in the related-work section, the empirical comparisons do not directly show why the proposed formulation offers a distinct advantage over representative alternatives.

  6. Some of the reported improvements are small or unstable.
    The ImageNet-1K improvement is only 0.20 percentage points, and several coefficient ablations show architecture-dependent gains or losses. This weakens the claim that the mechanism provides a generally reliable improvement.

  7. The parameterization can become poorly conditioned.
    In the broader coefficient experiments, the learned feedback coefficient can become extremely large while the memory-update coefficient becomes very small, suggesting a scale-compensation or identifiability issue. This raises questions about the interpretability and numerical stability of the learned parameters.

Questions for authors
  1. What exactly is being predicted by the proposed “self-prediction” mechanism? Since the formulation does not explicitly predict a future quantity or optimize a prediction error, why should it be interpreted as predictive coding rather than as an auxiliary recurrent/adaptive state?

  2. Could the authors compare against a simple capacity-matched baseline that adds an additional leaky input-memory state but does not include spike feedback? This would help determine whether the gains come from the proposed input-output residual or simply from adding another recurrent state.

  3. In the feedback-path ablations, input-memory-only is already strong, and full feedback can even underperform input-memory-only depending on the architecture. Can the authors clarify what empirical evidence demonstrates that the spike-feedback term is an essential part of the method?

  4. The broader learn-bb experiments improve 11 of 20 configurations and degrade 9, with a very small mean gain. How should this result be reconciled with the claim that learning the spike-feedback strength generally improves performance?

  5. Could the authors provide direct comparisons with representative adaptive/recurrent neuron models, such as ALIF/adLIF, TC-LIF, TS-LIF, STC, or other comparable auxiliary-state neuron models, under the same backbone and training setting?

  6. Some learned coefficients become extremely large while the corresponding memory-update coefficient becomes very small. Does the model have a scale-identifiability issue between bb and τp\tau_p? Would directly parameterizing the effective quantity τpb\tau_p b be more stable and interpretable?

  7. Could the authors report runtime and memory overhead introduced by the additional recurrent state during training and inference?

1 comment
AuthorsOpenReview ↗

Mechanistic Framing, Capacity-Matched Controls, and Conditioning

Thank you for your structured review, insightful questions, and recognition of our broad evaluation and diagnostic ablations. Your feedback directly helps us sharpen our terminology, highlight our capacity-matched controls, and clarify parameter conditioning.

1. Reframing "self-prediction" We agree that mpl[t]m_p^l[t] does not explicitly predict future stimuli or optimize a prediction-error loss; rather, it modulates membrane integration using an exponentially weighted trace of input drive minus emitted spikes: Il[t]=xl−1[t]+mpl[t−1]I^l[t]=x^{l-1}[t]+m_p^l[t-1], mpl[t]=(1−τpl)mpl[t−1]+τpl(alxl−1[t]−blsl[t])m_p^l[t]=(1-\tau_p^l)m_p^l[t-1]+\tau_p^l(a^l x^{l-1}[t]-b^l s^l[t]). In the revision, we will clarify that "self-prediction" serves as an intuitive label for this history-dependent anticipatory drive, while framing the mechanism rigorously as neuron-local input-output memory with subtractive feedback.

2. Capacity-matched baseline in Table 9 Table 9's "Input memory only" variant (mpl[t]=(1−τpl)mpl[t−1]+τplxl−1[t]m_p^l[t]=(1-\tau_p^l)m_p^l[t-1]+\tau_p^l x^{l-1}[t], added to Il[t]I^l[t]) serves as this exact capacity-matched baseline. Verified in our implementation, it disables only the spike-feedback term while retaining the identical auxiliary state dimension and learnable τpl\tau_p^l parameter count as fixed full feedback, achieving Last-5 accuracies of 94.998% (SEW ResNet-18) and 94.798% (Spiking ResNet-18).

3. Complementary roles of input memory and spike feedback Comparing full feedback against capacity-matched input-only memory yields Last-5 accuracies of 95.170% (+0.172 pp) on SEW ResNet-18 and 94.572% (-0.226 pp) on Spiking ResNet-18. Together with our single-neuron and sequence analyses, this cleanly disentangles two complementary functions: input memory supplies persistent drive for accuracy gains, whereas subtractive spike feedback acts as a homeostatic regulator—curbing strong-input firing from 19 spikes back to 10 spikes in Fig. 2, eliminating post-stimulus firing across 36/36 settings in Fig. 12, and enabling long-horizon credit assignment on Sequential CIFAR-10 (Table 3).

4. Positioning fixed rules vs. learnable blb^l We appreciate this point. While learning blb^l improves LIF accuracy in Table 5 (+0.22/+0.33 pp Best; +0.118/+0.420 pp Last-5), the broader 20-configuration sweep in Table 10 shows 11 gains and 9 losses (mean +0.0465 pp). Accordingly, we will position the tuning-free fixed reference rule (al=1,bl=τla^l=1, b^l=\tau^l) as our primary recommended method and present learnable blb^l as architecture-dependent adaptation, while flagging historical comparisons and single-run observations (including ImageNet-1K's +0.20 pp).

5. Comparison with related adaptive/recurrent neurons Thank you for highlighting ALIF/adLIF, TC-LIF, TS-LIF, and STC. We will expand Section 2 to systematically contrast state dimensions, driving signals, and feedback injection points—highlighting how our state couples pre-synaptic input current with output spikes while preserving base membrane/reset mechanics. Controlled benchmarks under matched backbones, time steps, and training budgets (reporting both accuracy and overhead) are outlined as an immediate follow-up study.

6. Parameter conditioning and effective step size Defining κl=τplbl\kappa^l = \tau_p^l b^l rewrites the update as mpl[t]=(1−τpl)mpl[t−1]+τplalxl−1[t]−κlsl[t]m_p^l[t]=(1-\tau_p^l)m_p^l[t-1]+\tau_p^l a^l x^{l-1}[t] - \kappa^l s^l[t]. Thus, even when bl≈86201.5b^l \approx 86201.5 is paired with τpl≈1.13×10−4\tau_p^l \approx 1.13\times 10^{-4}, the immediate spike-triggered decrement remains bounded at κl≈9.756\kappa^l \approx 9.756. Because τpl\tau_p^l also governs memory retention (1−τpl)(1-\tau_p^l), this is not exact global scale non-identifiability, yet extreme pairs indicate ill-conditioning; we will discuss directly parameterizing κl\kappa^l in Appendix F.3 as a cleaner, well-conditioned alternative.

7. Analytical vs. empirical system footprint Storing one auxiliary scalar state per neuron requires 4N4N bytes of persistent FP32 memory for NN neurons, alongside elementwise updates per step and BPTT intermediate storage. We will clarify in the revision that this represents an analytical complexity bound rather than empirical profiling, noting that wall-clock latency, peak GPU memory, and SOPs depend on implementation, batch size, and TT, and warrant dedicated benchmarking.

Reviewer_LS4FOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a self-prediction-enhanced spiking neuron model that uses historical inputs and output spikes to modulate membrane potential dynamics. The method introduces an auxiliary state that combines input memory and spike feedback, with separate coefficients controlling their contributions. It is applied to four existing neuron models: IF, LIF, PLIF, and CLIF. The authors analyze the resulting neuronal dynamics and gradient propagation and investigate the effects of learning the feedback coefficients. Experiments cover CIFAR-10, ImageNet-100, ImageNet-1K, sequential CIFAR-10, and four MuJoCo reinforcement learning tasks. The reported results show accuracy improvements in most evaluated configurations, although the benefits of learnable feedback coefficients vary across architectures and neuron types.

Critical strengths

1. Broad evaluation across neuron models and tasks

The authors evaluate the proposed feedback mechanism across four neuron models (IF, LIF, PLIF, and CLIF), multiple network architectures, and classification and reinforcement learning tasks. On CIFAR-10, the fixed-feedback variants improve accuracy in 35 of 40 configurations. The method also shows gains on sequential CIFAR-10 and most evaluated MuJoCo configurations. These results suggest that the mechanism is applicable beyond a single neuron model or task, although the magnitude of improvement varies considerably.

2. Simple formulation with useful ablations

The method introduces a neuron-local auxiliary state without changing the underlying membrane integration and reset rules. The authors provide explicit update equations, analyze the additional gradient pathways, and separately examine input memory and spike feedback. These ablations help clarify which components contribute to the observed improvements, even though the benefits of the complete feedback mechanism are not consistent across architectures.

Critical weaknesses

1. Limited methodological novelty

The main idea is closely related to existing neuron-level feedback and adaptation mechanisms. The paper itself discusses ALIF/LSNN, STC (ICML 2024), and adLIF (Nature Communications, 2025), which already incorporate spike history, adaptive states, or neuronal self-feedback. The proposed method mainly combines input history and spike feedback in an additive auxiliary state, with separately adjustable coefficients.

Although this formulation differs from prior methods, the paper does not establish why this particular combination constitutes a substantial methodological advance. More importantly, it lacks direct experimental comparisons with these closely related approaches. The term "self-prediction" is also questionable, since the method does not explicitly predict future neuronal activity or optimize a prediction error.

2. Insufficient large-scale evaluation and limited performance gains

The experimental evaluation is broad in terms of configurations but less convincing in terms of scale and statistical reliability. The only ImageNet-1K experiment, using SEW ResNet-18 with PLIF neurons, improves accuracy from 64.25% to 64.45%, a gain of just 0.20 percentage points. No repeated runs or uncertainty estimates are reported for this result.

Given the limited methodological novelty, stronger empirical evidence is needed to establish the practical value of the proposed modification. The current large-scale results do not demonstrate a substantial or reliably reproducible advantage.

3. Inconsistent benefits of the proposed feedback mechanism

The ablation results raise further questions about whether the added spike-feedback mechanism is necessary. In Table 9, full feedback performs worse than input-only memory on Spiking ResNet-18 (94.572% vs. 94.798% in the Last 5 metric). In Table 10, independently learning the feedback coefficient improves accuracy in only 11 of 20 configurations, with an average gain of just 0.0465 percentage points over the historical fixed-feedback results.

These findings suggest that the additional feedback mechanism does not consistently improve performance. Moreover, the method introduces an auxiliary state and additional computation, but the paper does not adequately quantify the resulting memory, latency, or energy costs. It remains unclear whether the proposed mechanism offers a meaningful advantage over simpler alternatives.

Questions for authors
  1. What is the main methodological novelty compared with existing neuron-level feedback mechanisms?

The Related Work section discusses several closely related methods, including ALIF/LSNN, STC (ICML 2024), and adLIF (Nature Communications, 2025), which already incorporate spike-history-dependent adaptation, auxiliary states, or neuronal self-feedback.

What is the fundamental methodological contribution of this paper? Can the authors explain why this formulation constitutes a substantive advance over these existing approaches, rather than an incremental modification?

  1. What motivates the specific design of the proposed feedback mechanism?

Feedback and information reuse are already well-established in biological neural circuits and artificial neural networks. Biological neurons exhibit self-feedback through autapses and dendritic mechanisms, while recurrent SNNs and adaptive neuron models have long incorporated history-dependent feedback. Residual networks also demonstrate the utility of reusing information, although their skip connections are not equivalent to temporal feedback.

Given these existing mechanisms, what is the specific motivation for combining historical inputs and output spikes through the proposed residual (δ[t]=ax[t]−bs[t]\delta[t] = ax[t] - bs[t])? Is this formulation derived from a biological principle, a predictive coding objective, or a theoretical analysis, or is it primarily a heuristic design choice? In particular, why is subtractive spike feedback appropriate, and what makes this mechanism self-prediction rather than another form of recurrent adaptation?

2 comments
AuthorsOpenReview ↗

Core Contribution, Mechanistic Motivation, and Calibrated Claims

Thank you for your constructive review and for recognizing our broad evaluation across neuron models and tasks as well as our diagnostic ablations. Your feedback helps us distill our core methodological contribution and calibrate our empirical claims.

1. Core methodological contribution Our primary contribution is a unified, plug-and-play joint input-output memory module: Il[t]=xl−1[t]+mpl[t−1]I^l[t]=x^{l-1}[t]+m_p^l[t-1] and mpl[t]=(1−τpl)mpl[t−1]+τpl(alxl−1[t]−blsl[t])m_p^l[t]=(1-\tau_p^l)m_p^l[t-1]+\tau_p^l(a^l x^{l-1}[t]-b^l s^l[t]). By decoupling incoming-current memory (alxl−1[t]a^l x^{l-1}[t]) from spike-triggered subtractive feedback (−blsl[t]-b^l s^l[t]) while keeping each base neuron's membrane and reset dynamics intact, the exact same formulation and path ablations transfer seamlessly across IF, LIF, PLIF, and CLIF. Rather than claiming priority over neuronal feedback, our work establishes how this specific decoupled residual design generalizes across diverse base neurons, architectures, and tasks.

1a. Relationship to ALIF/LSNN, STC, and adLIF We will expand Section 2 with a structured comparison of state variables, driving signals, and feedback injection points. Specifically, adLIF couples membrane voltage to an adaptation state and ALIF modulates thresholds using spikes alone, whereas our state integrates pre-synaptic input current alongside output spikes and feeds additively into subsequent integration—establishing a surrogate-free temporal gradient pathway through xl−1[t]x^{l-1}[t]. We will also outline matched-backbone comparisons against these adaptive models as a focused follow-up study.

2. Mechanistic motivation for the input-output residual The residual δl[t]=alxl−1[t]−blsl[t]\delta^l[t] = a^l x^{l-1}[t] - b^l s^l[t] balances persistent input integration against output-dependent homeostatic regulation. Crucially, mpl[t]m_p^l[t] persists across membrane resets, preserving temporal context that hard/soft reset erases. When driven by sustained input, mpl[t]m_p^l[t] accumulates to facilitate firing; once a spike occurs (sl[t]=1s^l[t]=1), subtractive feedback decrements mpl[t]m_p^l[t] by τplbl\tau_p^l b^l to prevent excessive bursting—cutting strong-input firing from 19 spikes (input-only) back to 10 spikes (matching base LIF) in Fig. 2. We will frame this clearly as a dynamical design principle.

2a. Clarifying "self-prediction" We fully agree that our module does not optimize an explicit future-target prediction error or derive from a specific biophysical circuit. In the revision, we will clarify that "self-prediction" is used intuitively for history-dependent internal drive, and define the mechanism rigorously as neuron-local input-output memory with subtractive feedback.

3. Insights from component ablations Table 9 directly provides a capacity-matched input-only control (bl=0b^l=0, matching state dimension and trainable parameter count). Comparing input-only memory against full feedback (94.998% →\to 95.170% Last-5 on SEW ResNet-18; 94.798% →\to 94.572% on Spiking ResNet-18) cleanly isolates their distinct roles: input memory drives a substantial share of the static accuracy gain (with full fixed feedback achieving +0.490/+1.122 pp Last-5 over base LIF in Table 5), while subtractive spike feedback provides homeostatic firing regulation (Fig. 2 & Fig. 12) and long-horizon sequence stability (Table 3).

4. Scope of learnable blb^l and large-scale results Your point is well taken. In the revision, we will designate the tuning-free fixed reference rule (al=1,bl=τla^l=1, b^l=\tau^l) as our primary method, while framing learnable blb^l (11 wins/9 losses, +0.0465 pp mean in Table 10) as configuration-dependent adaptation rather than a universal gain. Likewise, we will explicitly present the ImageNet-1K improvement (64.25% →\to 64.45%) as a single-run observation without multi-seed error bars.

5. Analytical vs. empirical system cost Architecturally, our module adds one continuous state per neuron and elementwise operations without new dense weight matrices (plus one scalar per layer when learning blb^l). We will revise the text to clearly separate this analytical operation count from empirical hardware profiling (network-wide SOPs, latency, and peak GPU memory), highlighting full system benchmarking as an important next step.

The main concern remains unresolved. The paper evaluates the proposed method across several neuron models, including IF, LIF, PLIF, and CLIF, and reports improvements in many settings. However, it does not directly compare against closely related adaptive neuron models, such as ALIF, LSNN, and adLIF. Without these comparisons, it remains unclear whether the proposed method offers an advantage over existing approaches. This is an important gap in evaluating its novelty and practical value.

Reviewer_CuNyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose a self-prediction enhanced spiking neuron framework, in which each neuron maintains a neuron-local auxiliary state that integrates a weighted input–output residual, i.e., input memory weighted by a coefficient a and subtractive spike feedback weighted by a coefficient b . This auxiliary state acts as an additional prediction current that modulates the subsequent membrane potential integration. The framework is general enough to be attached to IF, LIF, PLIF, and CLIF neurons, and the reference feedback rules are recovered as special cases of the coefficients. The authors provide analyses from the perspectives of dynamics, gradient pathways, and biological motivation, and conduct experiments on CIFAR-10 (5 architectures × 4 neuron types × 2 time steps), ImageNet-100, ImageNet-1K, Sequential CIFAR-10, and four MuJoCo continuous-control tasks with TD3. Ablation studies on the learnable coefficients and on detaching the feedback spike gradient are also included. The overall idea is interesting and the experimental workload is considerable.

Critical strengths

1.Simple and plug-and-play design. The proposed module adds only one scalar coefficient and one auxiliary state per spiking layer. It can be integrated into various neuron models and architectures with very little engineering effort, which I think is practically valuable.

2.Broad evaluation coverage. The experiments span five architectures, four neuron models, two time steps, static and sequential image classification, large-scale ImageNet experiments, and reinforcement learning tasks. The amount of experimental work is quite large and is appreciated.

3.Honest reporting of negative results. The authors clearly report that 4 out of 40 configurations degrade and 1 ties, that learning the feedback coefficient improves only 11 out of 20 combinations, and that Hopper-v4 slightly regresses under the PT framework. Such honesty is relatively rare in submissions and I respect it very much.

4.Careful single-neuron analysis. The illustration in Figure 2, together with the exact history decomposition, the one-step causal intervention, and the sensitivity grid in the appendix, gives a clear and trustworthy picture of the forward dynamics of the module.

5.Detailed appendix. Training protocols, augmentation pipelines, hyperparameters, and the code-to-notation mapping are documented, which is helpful for reproduction.

Critical weaknesses
  1. Limited novelty relative to existing neuron-level mechanisms.

    Mathematically, the auxiliary state is a first-order exponential moving average of the weighted input minus the emitted spikes, added to the input current. The spike-feedback term is very close to the adaptation current in adLIF (Baronig et al., 2025), the threshold adaptation in ALIF/LSNN (Bellec et al., 2018), and the gated terms in GLIF (Yao et al., 2022). The input-memory term is similar to a slow synaptic current. Combining these known ingredients with two scalar coefficients is a reasonable engineering contribution, but the conceptual novelty appears modest. More importantly, no experiment compares against any of these closely related methods, so it is difficult to position the reported gains.

    In addition, I would like to suggest the authors to also discuss and, if possible, compare with the following closely related work, which is currently missing: Tao Chen, Shukai Duan, Lidan Wang, "Exploiting memristive autapse and temporal distillation for training spiking neural networks," Knowledge-Based Systems, vol. 305, 112627, 2024. This work (MA-SNN) is also inspired by cortical autaptic structures, uses memristor-based feedback-connected autapses to adaptively regulate spikes, and is evaluated on CIFAR-10/100 and neuromorphic datasets (DVS-CIFAR10, N-Caltech101). Since the present paper also motivates its design from neuronal self-feedback/autapses and targets training enhancement of SNNs, this seems to be a direct and necessary baseline for positioning the contribution.

  2. The predictive-coding motivation does not fully match the mechanism. In predictive coding, the residual should be the difference between the actual input and a prediction of it. Here, however, the residual is x−b⋅s , where s is the neuron's own output spike rather than a prediction of x . This is more like autaptic self-feedback plus input smoothing than prediction in the PC sense. Furthermore, the cited biological evidence (Yin et al., 2018) concerns excitatory autapses, while the proposed feedback is subtractive, as the authors themselves acknowledge in Appendix E.2. I would gently suggest reconsidering the framing to avoid overstating the biological connection.

  3. Statistical reliability of the main claims is weak. All main classification results (Tables 1–3) come from a single run with "best" accuracy and no error bars. The paper's own data show the magnitude of seed noise: the same fixed-feedback configuration on SEW ResNet-18 with LIF at T=4 is reported as 95.04 (Table 1 / historical run in Table 10) and 95.36 (Table 5), a 0.32 pp gap between two runs under the shared protocol. Therefore, many reported improvements in the 0.1–0.3 pp range (including the ImageNet-1K gain of +0.20 pp and the coefficient-learning gains of +0.22/+0.33 pp) are unfortunately indistinguishable from run-to-run noise. Also, Table 10 compares new learn-b runs against "historical" fixed runs from earlier experiment batches, which is not a well-controlled comparison.

  4. The claimed contributions are not fully closed by the experiments. (a) The learnable-coefficient contribution — one of the main contributions — is actually contradicted by the broader exploration in Table 10 (11 improvements vs. 9 decreases, descriptive mean +0.0465 pp). (b) The gradient-pathway claim is only weakly supported: detaching the feedback spike gradient changes Last5 accuracy by merely 0.032 pp on SEW ResNet-18, while the Sequential CIFAR-10 ablation suggests it consistently helps; these two results are not well reconciled. (c) The introduction motivates the work by training difficulty and surrogate-gradient limitations, but the module mainly reshapes forward dynamics; the loop from problem to module to evidence is partially broken.

  5. The energy-efficiency dimension is not examined. The single-neuron simulations clearly show that input memory increases firing (e.g., 10 → 19 spikes under strong input). Since sparsity is the core selling point of SNNs, the trade-off between accuracy gain and increased spike rate / SOPs / energy, as well as the extra memory and compute cost of the auxiliary state during BPTT, should be quantified.

  6. Some baselines appear under-tuned, which may amplify gains. For example, the LIF base on ImageNet-100 (65.80%) and on Spiking ResNet-34 at T=8 (87.71%) are noticeably low, and the corresponding gains (+3.26 pp, +6.11 pp) are the largest in the paper. Clarification on baseline tuning would be helpful. Also, neuromorphic benchmarks (DVS-CIFAR10, N-MNIST, etc.) are absent, which are the standard testbeds for dynamics-enhancement methods.

  7. Minor issues. The theoretical analyses (boundedness in Eq. 10, the Jacobian in Eqs. 17–18) are descriptive and do not lead to actionable conclusions; the exponential parameterization can produce pathological coefficient pairs ( e.g., b with tau_p, Appendix F.3) without discussion; figures use raw TensorBoard tags as legends (Figure 3) and the overall presentation still feels unfinished.

Questions for authors
  1. Could the authors kindly provide comparisons with adLIF, ALIF/LSNN, GLIF, STC, and MA-SNN (Chen et al., KBS 2024, memristive autapse + temporal distillation) under the same architecture and training protocol? If the framework is claimed to be general, a mathematical clarification of its relationship to the adaptation current of adLIF would also be very valuable.

  2. Could the authors clarify the precise meaning of "self-prediction"? Since s[t] is not a prediction of x[t] , in what sense does x−b*s constitute a prediction error? How does the mechanism correspond to dendritic prediction of somatic activity (Urbanczik & Senn, 2014)?

  3. Would it be possible to report mean ± std over at least 3 seeds for the main tables? This is especially important for gains below 0.3 pp, such as ImageNet-1K (+0.20 pp) and the coefficient-learning results.

  4. Could the authors report average firing rates, SOPs/energy estimates, and the wall-clock/memory overhead introduced by the auxiliary state?

  5. Regarding the pathological parameter pair in Appendix F.3 (b^tau_p): does the exponential parameterization have an ill-conditioned direction, and is the learned effective strength tau_p*b well-behaved across layers?

  6. Could the authors explain how the baselines with unusually low accuracy (e.g., ImageNet-100 LIF at 65.80%, Spiking ResNet-34 LIF T=8 at 87.71%) were tuned?

  7. How should readers reconcile the detach ablation difference of only 0.032 pp on SEW ResNet-18 with the consistent improvements from retaining the gradient on Sequential CIFAR-10?

  8. As suggestions: please consider softening the predictive-coding narrative and positioning the method honestly as a neuron-local memory mechanism combining input smoothing and spike negative feedback; consider replacing "best" accuracy with last-k mean plus multi-seed statistics; adding neuromorphic benchmarks would also strengthen the paper considerably.

1 comment
AuthorsOpenReview ↗

Mechanism Reframing, Analytical Insights, and Revision Plan

Thank you for the thorough, fair, and deeply constructive review, and for appreciating our transparent reporting and single-neuron dynamical analysis. Your critique has directly helped us sharpen our framing and align our claims strictly with our empirical evidence.

1. Positioning and related adaptive models Thank you for recommending MA-SNN (Chen et al., KBS 2024). We will cite and discuss MA-SNN alongside an expanded mathematical comparison with adLIF, ALIF/LSNN, GLIF, and STC in Section 2. Unlike adLIF—which couples membrane voltage to an adaptation state driven solely by spikes—our auxiliary state mpl[t]m_p^l[t] preserves the base neuron's membrane and reset equations while jointly integrating pre-synaptic drive and emitted spikes: for al=0a^l=0, −mpl[t]-m_p^l[t] acts as a filtered spike trace; for al>0a^l>0, it retains incoming-current history and creates a surrogate-free temporal gradient pathway (∂mpl[t]/∂xl−1[t]=τplalI\partial m_p^l[t]/\partial x^{l-1}[t] = \tau_p^l a^l \mathbf{I}). We will also outline matched-backbone benchmarks (disentangling autapse vs. distillation gains for MA-SNN) as an immediate next step.

2. Reframing "self-prediction" and biological inspiration We fully accept your suggestion. Operationally, mpl[t]m_p^l[t] provides history-dependent modulation rather than supervised future prediction, and δl[t]=alxl−1[t]−blsl[t]\delta^l[t] = a^l x^{l-1}[t] - b^l s^l[t] is an input-output residual. In the revision, we will foreground the precise mechanistic description—neuron-local input-output memory with subtractive spike feedback—and explicitly clarify that Urbanczik & Senn (2014) and Yin et al. (2018) serve as conceptual inspiration for local feedback rather than biophysical derivations.

3. Statistical scope and reporting clarity We agree that separating multi-seed uncertainty from within-run metrics is essential. In the revision, we will explicitly label the 95.04% and 95.36% entries as separate experiment batches, identify all historical comparisons in Table 10, and frame the +0.20 pp ImageNet-1K gain as a single-run observation. Alongside Best accuracy, we will highlight within-run Last-5 summaries (e.g., +0.118/+0.420 pp when learning blb^l in Table 5) while noting multi-seed evaluation as a clear limitation.

4. Firing regulation and computational footprint In the strong-input simulation (Fig. 2), base LIF, input-only memory, and full feedback emit 10, 19, and 10 spikes, illustrating how subtractive spike feedback curbs the excess firing caused by unopposed input memory. In the revision, we will clearly distinguish this single-neuron dynamical regulation and our analytical complexity count (one continuous state per neuron and elementwise updates, plus one scalar per layer when learning blb^l) from network-level SOPs, latency, and peak-memory profiling, which we highlight for future benchmarking.

5. Parameter conditioning and effective step size Your observation on Appendix F.3 is spot on. Rewriting the update shows that κl=τplbl\kappa^l = \tau_p^l b^l sets the immediate spike-triggered decrement while 1−τpl1-\tau_p^l governs memory retention. Thus, in the recorded pair (bl≈86201.5b^l \approx 86201.5, τpl≈1.13×10−4\tau_p^l \approx 1.13\times 10^{-4}), the effective jump remains finite (κl≈9.756\kappa^l \approx 9.756). While altering τpl\tau_p^l changes retention (preventing exact scale invariance), extreme pairs indicate ill-conditioning; we will discuss direct parameterization of κl\kappa^l in Appendix F.3 as a clean stabilization remedy.

6. Matched baseline protocols Both base and augmented models share identical training hyperparameters (e.g., SEW ResNet-34 on ImageNet-100: T=4T=4, 100 epochs, batch size 32, lr 0.001). We will explicitly state in Section 5 that reported gains quantify improvements under matched standard recipes rather than exhaustive per-baseline hyperparameter sweeps.

7. Reconciling gradient-path ablations Comparing static SEW ResNet-18 (T=4T=4, +0.032 pp Last-5) with Sequential CIFAR-10 (T=32T=32, +5.78/+0.95/+0.49/+1.24 pp Best across IF/LIF/PLIF/CLIF, Table 3) reveals a clear task-horizon dependence: propagating surrogate gradients through −blsl[t]-b^l s^l[t] becomes particularly impactful over longer sequences without spatial shortcuts. We will refine Section 4.3 to frame our Jacobian analysis as establishing gradient-pathway availability that benefits long-horizon tasks.

8. Manuscript polish We will replace all raw TensorBoard screenshots (Fig. 3 and Appendix) with clean publication-grade figures, position the tuning-free fixed reference rule as our primary method (framing Table 10's 11 wins/9 losses, +0.0465 pp mean as architecture-dependent adaptation), and note neuromorphic datasets as next-step validation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.