acceptodds
Under review as a conference paper at ICLR 2027

DNASNet: Decoupled Neural Architecture Search Driven by Task-Conditioned Spike Timing

Abstract

Spiking neural networks (SNNs) promise energy-efficient computation, but differentiable search over temporally unrolled SNNs is costly. DNASNet decouples weight-sharing NAS: spatio-temporal backpropagation (STBP) learns shared weights, while an STDP-inspired rule updates architectures from task-induced spike timing without architecture-level backpropagation. With network state and timing evidence fixed, the STDP utility is the gradient of an explicit concave temporal-balance potential, and the normalized update equals entropy mirror ascent on node-wise probability simplices. DNASNet produces stable architectures across random runs, achieves near-state-of-the-art performance on CIFAR-10, CIFAR-100, and DVS-Gesture, and yields competitive results on DVS-CIFAR-10 and ImageNet1K. Compared with bilevel optimization, it reduces peak search memory by up to 22% and search time by up to 51%. Ablations further confirm the importance of timing information for architectural consistency and retraining performance.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spiking Neural Networks, Neural Architecture search, Spike-Timing-Dependent Plasticity

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_NWbAOpenReview ↗
Rating
3
Confidence
4(+2)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper concerns architecture search for spiking neural networks, where the differentiable formulation couples the architecture update to spatio-temporal backpropagation and therefore requires an unrolled backward graph. The authors decouple the two: shared weights are trained with STBP, while routing probabilities over candidate edges and operations are updated from spike timing statistics collected in a forward pass, with no backward path through the architecture parameters. The update takes a multiplicative STDP form, which the authors characterize as entropy mirror ascent on the node-wise simplices, driven by the gradient of an explicit concave potential at a frozen network state. Each iteration alternates a supervised lookahead step, a gradient-free routing update on held-out data, and a weight update under the revised routing. Stability, convergence and regret results are stated with their assumptions, including the caveat that the convergence result does not cover the AdamW optimizer used in the experiments. Experiments span CIFAR-10, CIFAR-100, ImageNet-1K, DVS-CIFAR-10 and DVS-Gesture, plus a speech dataset used only in the ablation; accuracy on the two CIFAR benchmarks is close to the strongest published SNN-NAS results, and search cost is lower than SNN-DARTS. In the timing ablation, replacing the timing rule with a Hebbian or rate-only rule produces architectures composed almost entirely of pooling and skip connections.

Critical strengths
  1. The timing ablation is the strongest part of the paper, and it is set up so that it could have failed. Replacing the STDP utility with a Hebbian or a rate-only rule, holding the rest of the pipeline fixed, drops CIFAR-10 retraining accuracy from 96.51 to 87.90 and 91.17. The reason is visible in what the search returns rather than only in the final number: both alternative rules produce cells made almost entirely of average pooling and skip connections, while the timing rule retains convolutions. I have not seen the rate-versus-timing contrast demonstrated this concretely for SNN architecture search, and it is a mechanism-level result rather than another benchmark entry.
  2. The search-cost reduction is measured against a matched baseline, and the paper reports where it disappears. With identical search hyperparameters, DNASNet uses roughly a fifth less peak memory and between a quarter and a half less wall-clock search time than SNN-DARTS on CIFAR-10 and CIFAR-100. On DVS-Gesture the paper states that peak memory is the same for every method, because at long sequence lengths the temporal backward graph is the bottleneck rather than the architecture update. Reporting the negative case next to the positive one is what makes the efficiency claim credible.
Critical weaknesses
  1. The accuracy claims either fall within run-to-run noise or rest on unmatched comparisons. The CIFAR-10 margin over NeuEvo is 0.08 points, smaller than the spread across the authors' own runs. The ImageNet gain comes from retraining a transferred CIFAR-10 cell against baselines that were searched on ImageNet, so it does not isolate the search method. On DVS-CIFAR-10 the method trails NeuEvo by 2.64 points, and SNN-DARTS was not searched on that dataset at all. The accuracy table is therefore not what supports the performance claim.
  2. The abstract claims near-state-of-the-art performance on DVS-Gesture, which the paper does not support. That benchmark appears only in the rule ablation and a retraining curve, and every number for it is measured against SNN-DARTS or the authors' own rule variants. No external baseline is listed for it anywhere in the submission, and the configuration used there is deliberately small. Either supply external baselines under a stated protocol or drop the claim.
Questions for authors
  1. The ImageNet result comes from retraining the best CIFAR-10 cell rather than from searching on ImageNet, while the baselines appear to be searched results, and the SEW-ResNet entry is quoted from NeuEvo. Can the authors say what the comparison looks like under matched protocols, or restate the claim as one about transfer? This is the number I would otherwise weigh most heavily, and at present I cannot tell how much of the 4.53-point gap comes from the search method and how much from the transfer setup.
  2. On CIFAR-10 the margin over NeuEvo is 0.08 points and the proposed method runs vary by 0.10. Do the authors regard the two methods as separated by these results? If not, how would you describe the accuracy contribution, as opposed to the search-cost contribution?
  3. SNN-DARTS was not searched on DVS-CIFAR-10, so the only in-house comparison there is a random-architecture baseline, and the remaining entries are published numbers from other protocols. Would it be feasible to search SNN-DARTS on that dataset, or to mark the comparison explicitly as cross-protocol in the table?
3 comments

We appreciate your valuable feedback.

To Q1/W1:

We agree with the reviewer and have explicitly framed the ImageNet-1K experiment as evaluating cross-dataset cell transferability from CIFAR-10 in the revised PDF (Section 4, lines 326–329).

All search baselines on ImageNet in Table 1 report results obtained by transferring and retraining the best cell searched on CIFAR-10. Following the standard cell-transfer protocol established in DARTS and adopted by SNN-NAS baselines such as SpikeDHS and NeuEvo, we search for the cell topology on CIFAR-10, scale up the network depth and channels(Appendix C.3 shows the retrained network configuration after extension on ImageNet), and train the resulting target architecture from scratch on ImageNet-1K (whereas the remaining ImageNet-1K entries in Table 1 are handcrafted architectures without search).

Under this matched CIFAR-10-to-ImageNet-1K cell-transfer protocol, DNASNet (73.27%, 22M) outperforms both NeuEvo (68.74%, 26M) and SpikeDHS (68.64%, 58M) by more than 4.5 percentage points with fewer parameters. Note that the NeuEvo paper (from which the SEW-ResNet-34 baseline accuracy of 67.23% is cited) does not list the parameter count for SEW-ResNet-34.

To Q2/W1:

We agree with the reviewer: given the ±0.10%\pm 0.10\% run-to-run variation, we regard DNASNet and NeuEvo as comparable on CIFAR-10 (we reworded this contribution in the revised edition, lines 067–069) rather than "performance gain". Rather, we characterize DNASNet's accuracy contribution by its substantial margins over NeuEvo on CIFAR-100 (+0.85 points) and ImageNet-1K transfer (+4.53 points), as well as its consistent gains over SNN-DARTS within the matched search space.

To Q3/W1:

We ran an additional SNN-DARTS search on DVS-CIFAR-10 using the T=10T=10 protocol described in Appendix C.4, and retrained the selected architecture under the same evaluation protocol. In this search run, the normal cell contains only skip connections (visualization results are supplemented in Figure 7(e), lines 1542–1552):

Genotype(
    normal=[
        ('skip_connect', 1), ('skip_connect', 0),
        ('skip_connect', 1), ('skip_connect', 2),
        ('skip_connect', 0), ('skip_connect', 1),
        ('skip_connect_back', 0), ('skip_connect_back', 3),
    ],
    normal_concat=range(2, 5),
)

The retraining results are as follows:

DatasetSearch methodTTParameters (M)Accuracy (%)
DVS-CIFAR-10SNN-DARTS101977.00
DVS-CIFAR-10Random101979.10
DVS-CIFAR-10DNASNet101981.53

This skip-dominated outcome is consistent with the DVS-Gesture example in Appendix Figure 7(c)–(d), where six feedforward operations selected by SNN-DARTS are skip connections. That model reaches 89.76%, compared with 93.92% for Random and 97.05% for DNASNet.

One possible explanation is that surrogate-gradient propagation over longer temporal dependencies weakens architecture gradients for parameterized branches, favoring skip connections in the competition. DNASNet instead updates routing using timing statistics collected during forward propagation and avoids that architecture-gradient path.

We have added the new SNN-DARTS retraining result to the revised PDF (Section 4.2, Table 1, line 414) and updated Appendix C.4 to reflect that this additional search has now been performed (Appendix C.4.1, lines 1435–1437; Figure 7(e), lines 1542–1552).

To W2:

Thanks to the reviewer for pointing out this formulation problem, we removed it in the revised draft (Abstract, p.1, lines020-021). However, the scale was not "deliberatly reduced". DVS-Gesture was trained using a network of similar size (0.13M) in SEW-ResNet. They obtained 97.92% accuracy at T=16. DNASNet (0.15M, T=10, Accuracy=97.05%) still showed competitive performance.

I thank for the authors' rebuttal. I will raise my confidence after this rebuttal.

AuthorsOpenReview ↗

Thank you for your feedback!

Thank you very much for taking the time to review our rebuttal and for raising your confidence in our work! We greatly appreciate your constructive suggestions during the review process, which have helped us improve the clarity and rigor of the paper.

Reviewer_zibdOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DNASNet, a neural architecture search method designed for spiking neural networks (SNNs). It separates weight optimization from architecture updates, eliminating the need for architecture-level backpropagation through temporal unrolling. The authors provide a theoretical analysis showing that the update rule corresponds to entropy mirror ascent on probability simplices. Evaluations on image, event-based, and speech benchmarks demonstrate improved accuracy compared to SNN-DARTS under a matched search space, with reduced peak memory and search time.

Critical strengths
  1. The differentiable SNN-NAS with temporal unrolling suffers from high memory and runtime costs. Decoupling the architecture update from the weight backpropagation process is a promising method for reducing the search overhead.
  2. The theoretical analysis and experimental scope are quite broad.
Critical weaknesses
  1. The potential and gradient properties only hold under frozen network state. However, in the actual three-step pipeline, weights change between Step 1 and Step 3, and architecture update uses lookahead weights w^t\hat{w}_t, not fixed wtw_t. The concave potential and exact gradient interpretation therefore do not strictly apply to the actual running algorithm. The paper acknowledges this in Appendix A.3 only as a first-order Taylor approximation, but the main text heavily emphasizes the exact potential interpretation, which can mislead readers.
  2. Computational efficiency claims are misleading. The pipeline performs three forward-backward passes per iteration: Step 1 weight update (forward + backward), Step 2 architecture forward pass, Step 3 weight update (forward + backward). Compared with standard SNN-DARTS which performs one joint forward-backward for both weights and architecture, DNASNet reduces architecture-gradient memory but increases weight-gradient computation. The paper only reports peak memory and total search time, but does not break down FLOPs or compare per-iteration compute cost. The 51% time reduction may come from simpler update operations rather than fewer computations.
  3. Discretization drop is severe and undermines efficiency claims. On CIFAR-10, soft supernetwork accuracy drops from 74.33% to 41.24% after discretization. Although better than SNN-DARTS (11.77%), the huge gap means the continuous search does not produce well-behaved discrete architectures. Users must retrain from scratch anyway, so the search-phase efficiency gain is partially offset by the need for full retraining.
  4. No hardware or neuromorphic implementation analysis. This paper implies bio-inspired STDP update is hardware-friendly, but layer-wise probability routing, softmax normalization and lookahead weight steps are not trivial to implement on standard neuromorphic hardware.
  5. The rate-only baseline only uses post-operation firing rates, ignoring pre-synaptic activity entirely; Hebb uses raw co-activation without temporal kernels. There is no ablation that controls for total activation strength and isolates pure timing contribution. It remains unclear how much gain comes from temporal ordering vs. activation magnitude.
  6. There is no comparison with a simpler two-step variant (single weight update + architecture update). It is unknown whether the intermediate lookahead step is necessary or just adds overhead.
  7. Sensitivity analysis is only performed on small DVS-Gesture; it is unknown whether performance is robust across datasets and search spaces.
  8. The best CIFAR-10 architecture is directly transferred to ImageNet. This does not demonstrate the search method itself generalizes to larger datasets; it only shows the found cell transfers. In addition, there is no test on whether architectures found by STDP-based search generalize better or worse than DARTS architectures under input noise, event corruption or timing shift.
  9. The proposed method does not outperform other methods compared with the latest advances in SNNs [1,2,3], which diminishes the significance of this work.
[1] SpikCommander: A High-performance Spiking Transformer with Multi-view Learning for Efficient Speech Command Recognition. AAAI. 2026.
[2] Stable Spike: Dual Consistency Optimization via Bitwise AND Operations for Spiking Neural Networks. CVPR. 2026.
[3] Advancing Spatiotemporal Representations in Spiking Neural Networks via Parametric Invertible Transformation. ICLR. 2026.
Questions for authors

Please provide a response to the above questions, particularly regarding the efficiency and performance advantages of the proposed method.

7 comments

We sincerely thank the reviewer for your time.

To W1:

The exactness claim concerns the conditional routing update in Step 2, including the actual lookahead state used by the algorithm. It does not require the weights to remain fixed throughout the entire three-stage iteration. At iteration tt, Step 1 constructs w^t\widehat w_t. Step 2 collects evidence at (w^t,pt)(\widehat w_t,p_t), detaches the observed coefficients A^t,v±(w^t)\widehat A^{\pm}_{t,v}(\widehat w_t) from the computational graph, and holds them fixed while updating the routing. Under these conditions, the routing update is Eq. (9); its potential and mirror-ascent interpretations are established in Section 3.3 and Appendix A.2.

These are exact identities for the conditional surrogate used in that routing step. They do not rely on a Taylor approximation or on differentiating through the dependence of the collected statistics on the network state. As the weights and observed statistics change, the algorithm uses a sequence of conditional potentials Jt(⋅∣w^t)J_t(\cdot\mid\widehat w_t). Theorem 8, Eqs. (40)–(41), in Appendix B.3 addresses this online setting under its stated assumptions. To avoid possible misdirection, we have added this distinction to Section 3.4 (p. 6, lines 294–297).

The Taylor analysis answers a different question: Section 3.4 and Appendix A.3 examine how the supervised update from wtw_t to w^t\widehat w_t changes the conditional potential itself. They do not approximate or establish the Step 2 gradient identity. Figure 2 provides empirical evidence of the resulting timing dynamics, including the newly added two-step comparison in panel (b) (lines 242–244; lines 305–308).

To W2:

Thank you for pointing out arithmetic cost and wall-clock time are different quantities. We used FlopCounterMode to test FLOPs in actual iterations of the algorithm and analyze revenue sources.

In the standard alternating bilevel implementation of DARTS , architecture parameters are updated using a validation minibatch and weights using a training minibatch. The first-order approximation still performs these separate updates; it is not a single joint forward-backward pass.

DNASNet replaces expensive spatio-temporal architecture backpropagation [1,2] with forward-only timing-evidence collection and adds a supervised weight lookahead. Architecture gradients in an unrolled SNN propagate through the participating candidate branches and temporal dependencies. We tested the operation count per iteration and listed it together with the measured search cost as follows:

DatasetMethodBatch sizeTTFLOPs/iteration (101210^{12})Search time (h)Peak memory (GB)
CIFAR-10SNN-DARTS (first)12844.3090265.5322.40
CIFAR-10SNN-DARTS (second)12846.98055510.0524.42
CIFAR-10DNASNet12843.3732622.7018.00
CIFAR-100SNN-DARTS (first)12844.3091673.3014.92
CIFAR-100SNN-DARTS (second)12846.9807856.8516.21
CIFAR-100DNASNet12843.3734212.4611.67
DVS-GestureSNN-DARTS (first)641012.1127421.1459.69
DVS-GestureSNN-DARTS (second)641019.6225831.3366.84
DVS-GestureDNASNet64109.4724310.6159.69

The FLOPs count covers one complete search iteration, including all forward passes, backward passes, and routing updates. Across the three datasets, DNASNet reduces per-iteration FLOPs by a consistent ~22% compared with first-order SNN-DARTS, while reducing wall-clock search time by 25% to 51%. This gap occurs because FLOPs counts do not translate 1-to-1 to wall-clock runtime. While FlopCounterMode mainly captures compute-bound convolutional MACs, backpropagating architecture gradients through unrolled time steps and feedback loops across all candidate branches—and retaining their intermediate states—reduces hardware execution efficiency and adds disproportionately higher runtime overhead.

[1] Che, K., et al. Differentiable hierarchical and surrogate gradient search for spiking neural networks. NeurIPS, 2022.
[2] Yan, S., et al. Sampling complex topology structures for spiking neural networks. Neural Networks, 2024.

To W3:

Retraining is part of the standard evaluation pipeline, regardless of how much performance drops when pruning immediately after search. The search supernetwork is shallow and narrow, whereas the final model is trained from scratch after discretization and scaling. Accordingly, retrained accuracy evaluates the final architecture, while accuracy immediately after pruning diagnoses the effect of discretization on the current shared weights.

Figure 4 uses the latter diagnostic, also considered in GAEA [1]. We prune the supernetwork directly using the prescribed selection rule, then evaluate it without weight fine-tuning or BatchNorm recalibration. In an SNN, pruning can change the scale of input currents relative to firing thresholds and therefore alter spike dynamics. However, pruning candidate operations with lower routing probabilities perturbs the network far less than pruning operations that still carry substantial probability mass. Under this diagnostic, DNASNet retains 41.24% accuracy, compared with 11.77% for SNN-DARTS. This demonstrates better immediate post-discretization performance in this comparison.

The final models are evaluated under the standard two-stage search and retraining protocols described in Appendix C. Because retraining is structurally required when scaling the searched cell to the deeper and wider target network, reducing the search-phase cost directly lowers the total pipeline cost.

To W4:

Routing probabilities, softmax normalization, and supervised lookahead are used exclusively during GPU-based search and are removed upon discretization. The resulting fixed architecture consists solely of standard spiking convolutions, pooling, skip connections, and LIF neurons, making it well-suited for adaptation to neuromorphic chips (e.g., Intel Loihi) for practical deployment—though we recognize that comprehensive hardware-level benchmarking remains necessary.

The mean spike count per example and mean firing rates reported in the paper are inference-activity measurements. Prior work such as AutoSNN [2] uses spike activity in energy-aware architecture evaluation. These measurements can inform energy estimates under a specified hardware and operation-cost model. DNASNet uses no explicit spike-count regularizer during search, yet the selected architectures often exhibit lower spike counts and lower firing rates in deeper layers (Table 1,2; Fig 5 (c),(d)), this supports their potential for efficient inference.

[1] Li, L., et al. Geometry-aware gradient algorithms for neural architecture search. ICLR, 2021.
[2] Na, B., et al. AutoSNN: Towards energy-efficient spiking neural networks. ICML, 2022.

To W5:

Thank you for pointing out that ablating the temporal-kernel construction under strictly controlled activation strength is crucial to the rigor of our paper. In our ablations of the components of the STDP utility update rule, the Rate-only and Hebb designs (similar designs can be found in [1] and [2]), show that replacing the STDP utility changes the selected architectures and retrained accuracy.

To examine timing more directly, we added two controls(They are specifically defined in Appendix A.1. ):

  • Destruction time correspondence : preserve the total presynaptic and postsynaptic firing activity, but shuffle their temporal correspondence when collecting architecture-update statistics.
  • Randomized timing direction : preserve each candidate operation's total timing-evidence strength while randomizing its potentiation/depression direction.
DatasetSearch ruleTTParameters (M)Spikes (K)Accuracy (%)
CIFAR-10Destruction time correspondence424.7110295.54
CIFAR-10Randomized timing direction424.1118594.25
CIFAR-10STDP (DNASNet)424.661096.51
CIFAR-100Destruction time correspondence424.8101876.05
CIFAR-100Randomized timing direction424.882576.94
CIFAR-100STDP (DNASNet)424.658978.57
DVS-GestureDestruction time correspondence100.1572694.10
DVS-GestureRandomized timing direction100.1579493.06
DVS-GestureSTDP (DNASNet)100.1548397.05

DNASNet has higher retrained accuracy and fewer inference spikes than either control on all three datasets. Together with the original ablations, these results support the usefulness of structured timing evidence beyond the retained activity or evidence strength. We have added the complete control results and precise definitions of the perturbations and preserved quantities to the revised PDF (Section 4.3,lines 423–427; Table 2, lines 433–456, and text at lines 463–474; Appendix A.1, lines 814–823). It should be noted both temporal ablations follow the original method of network forward, weight backward propagation, and search iteration budgets, with additional time permutation or evidence direction randomization. These control operations involve random number generation, index rearrangement, and tensor processing, and therefore may incur additional overhead.

[1] Najarro, E., et al. Meta-learning through Hebbian plasticity in random networks. NeurIPS, 2020.
[2] Guo, W., et al. Towards efficient neuromorphic hardware: Unsupervised adaptive neuron pruning. Electronics, 2020.

To W6:

We ran an additional control that performs a supervised weight update and an architecture update in alternation, without a separate lookahead step. The training and architecture minibatches come from Dtrain\mathcal D_{\mathrm{train}} and Darch\mathcal D_{\mathrm{arch}}, respectively; the remaining search and retraining settings are unchanged.

DatasetUpdate scheduleTTParameters (M)Spikes (K)Accuracy (%)
CIFAR-10Two-step schedule425.6107094.69
CIFAR-10DNASNet424.661096.51
DVS-GestureTwo-step schedule100.1575695.48
DVS-GestureDNASNet100.1548397.05

DNASNet improves retrained accuracy by 1.82 percentage points on CIFAR-10 and 1.57 points on DVS-Gesture. These results support the full three-stage schedule.

Analysis. The current weights already contain historical task information. Lookahead additionally conditions routing on a fresh supervised response to the current routing, after which the final weight update responds to the newly selected routing. The accuracy comparison supports this schedule in the tested settings. Figure 2(b) of the revised PDF already shows the corresponding timing trajectories (p. 5, lines 228–242; Section 3.4, p. 6, lines 305–308): without lookahead, energy and normalized deviation do not exhibit the consistent decrease observed for DNASNet. (More detail please also see our response to Reviewer 6s2z (Q2/W2) for the searched two-step cell genotypes and further discussion.)

Reviewer_6s2zOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DNASNet, a neural architecture search framework for spiking neural networks (SNNs) that decouples architecture optimization from shared-weight training. Specifically, shared weights are optimized using spatio-temporal backpropagation (STBP), while architecture probabilities are updated using an STDP-inspired rule based on task-conditioned spike timing. The method employs supervised weight lookahead to collect timing evidence and performs node-wise architecture updates without architecture-level backpropagation. The authors further establish an entropy mirror-ascent interpretation and an explicit concave potential for the proposed update rule. Experiments demonstrate competitive classification performance and reduced search costs compared with SNN-DARTS.

Critical strengths

S1. The paper addresses an important problem of reducing the computational cost of neural architecture search for SNNs. Decoupling architecture updates from temporal backpropagation is a meaningful approach to improving search efficiency.

S2. The proposed framework provides a clear integration of STDP-inspired temporal learning and supervised weight optimization. The theoretical analysis offers useful insights into the proposed update mechanism.

S3. The experimental demonstrates competitive accuracy and improved search efficiency, while the ablation studies provide evidence for the effectiveness of the STDP-based search strategy.

Critical weaknesses

W1. Although DNASNet and SNN-DARTS use the same discrete search space, they differ in both their architecture update signals and continuous parameterizations. DNASNet employs node-wise probability competition with STDP-based updates, whereas SNN-DARTS uses edge-wise architecture parameters optimized through task-loss gradients. Consequently, the observed performance improvements may arise from the different parameterizations, the update mechanisms, or their combination. While the existing ablations demonstrate the effectiveness of STDP within the proposed framework, a comparison using the same node-wise parameterization with gradient-based architecture updates would better isolate the contribution of the proposed optimization strategy.

W2. A central motivation of DNASNet is to guide architecture optimization using task-conditioned spike timing. In Section 3.4, the authors first construct supervised lookahead weights and then collect timing evidence under the updated network state. However, the shared weights already contain task information from previous supervised training iterations. The current experiments do not directly establish whether the additional lookahead step provides meaningful benefits beyond computing STDP evidence using the current weights. An ablation comparing architecture updates based on (w_t) and (\hat{w}_t), while keeping the remaining settings unchanged, would help clarify the contribution of this design.

W3. Appendix A.1, Eq. (16), appears to use the same Gaussian perturbation for both the potentiation and depression evidence. If interpreted literally, substituting this perturbation into Eq. (3), with the default (c_v=2/d_v), gives an update proportional to ((1/d_v-p_{t,v,i})). Consequently, the random update vanishes under uniform routing probabilities and may be strongly suppressed near the approximately uniform initialization used in the experiments. This could affect the interpretation of the Random baseline in Table 2. The authors should clarify whether the positive and negative evidence components use independent random samples and provide the precise random-update formulation.

Questions for authors

Q1. How would DNASNet compare with a gradient-based architecture optimizer using the same node-wise probability parameterization and discretization procedure?

Q2. What is the contribution of supervised lookahead compared with collecting STDP evidence directly from the current shared weights?

Q3. In Eq. (16), are the Gaussian perturbations for the positive and negative timing evidence sampled independently? How is the Random baseline implemented?

3 comments

Thank you for the detailed feedback.

To Q1/W1:

We implemented Node-wise SNN-DARTS, which uses exactly the same node-wise routing parameterization and discretization procedure as DNASNet but retains gradient-based architecture updates. We repeated search and retraining on CIFAR-10, CIFAR-100, and DVS-Gesture. The DNASNet results from the paper are included below for comparison.

DatasetSearch methodTTParameters (M)Spikes (K)Accuracy (%)
CIFAR-10Node-wise SNN-DARTS425.463194.82
CIFAR-10DNASNet424.661096.51
CIFAR-100Node-wise SNN-DARTS425.560074.64
CIFAR-100DNASNet424.658978.57
DVS-GestureNode-wise SNN-DARTS100.1567893.05
DVS-GestureDNASNet100.1548397.05

On DVS-Gesture, node-wise parameterization improves SNN-DARTS from 89.76% to 93.05%, but the result remains 4.00 percentage points below DNASNet. On CIFAR-10 and CIFAR-100, it is respectively 0.55 and 0.64 points below the original edge-wise SNN-DARTS baseline. These results show that changing the parameterization alone does not explain DNASNet's improvement. Node-wise SNN-DARTS still requires architecture backpropagation, with search time and peak memory similar to those of SNN-DARTS.

To Q2/W2:

The current shared weights wtw_t already contain information from earlier supervised updates. Lookahead additionally conditions the routing update on the network's supervised response to the current routing. The final weight update then responds to the routing selected within the same iteration, as described in Section 3.4. The Taylor analysis in Section 3.4 and Appendix A.3 addresses a related effect: under the stated smoothness and boundedness assumptions, the first-order change in the conditional expected potential depends on the inner product between the training-loss gradient and the negative-potential gradient. When these directions align and the step is sufficiently small, the supervised lookahead increases that potential locally and changes the timing evidence used for routing. It should be noted that this is a conditional statement, not a guarantee that every supervised update increases the potential. However, the experimental results in Figure 2 demonstrate the effectiveness of this design, illustrating the spontaneous organization of the network to route to timing balance, Figure 2 (b) of the revised paper shows the timing dynamics without lookahead: neither energy nor normalized deviation shows the consistent reduction observed with DNASNet. The excellent retraining performance demonstrates the importance of timing balance for architecture selection.

We also ran a controlled comparison against a two-step schedule that alternates supervised weight updates on Dtrain\mathcal D_{\mathrm{train}} and routing updates on Darch\mathcal D_{\mathrm{arch}}, without a separate lookahead step. All other search and retraining settings were kept unchanged, results are as follows:

DatasetUpdate scheduleTTParameters (M)Spikes (K)Accuracy (%)
CIFAR-10Two-step schedule425.6107094.69
CIFAR-10DNASNet424.661096.51
DVS-GestureTwo-step schedule100.1575695.48
DVS-GestureDNASNet100.1548397.05

DNASNet improves retrained accuracy by 1.82 percentage points on CIFAR-10 and 1.57 points on DVS-Gesture. These results support the full three-stage schedule. Figure 2(b) shows the time series trajectory with the look-ahead removed, and the new description has been added to the revised pdf (lines 242-244; 305–308).

CIFAR-10 — two-step schedule

Genotype(
    normal=[
        ('conv_3x3', 1), ('conv_5x5', 0),
        ('avg_pool_3x3', 2), ('skip_connect', 1),
        ('avg_pool_3x3', 3), ('avg_pool_3x3', 2),
        ('avg_pool_3x3_back', 0), ('avg_pool_3x3_back', 2),
    ],
    normal_concat=range(2, 5),
)

DVS-Gesture — two-step schedule

Genotype(
    normal=[
        ('skip_connect', 0), ('skip_connect', 1),
        ('avg_pool_3x3', 2), ('conv_5x5', 1),
        ('avg_pool_3x3', 3), ('avg_pool_3x3', 2),
        ('avg_pool_3x3_back', 0), ('conv_3x3_back', 3),
    ],
    normal_concat=range(2, 5),
)

To Q3/W3:

The positive and negative components are not sampled independently. They are the positive and negative parts of the same Gaussian vector, so both evidence vectors are nonnegative. The original Eq. (16) omitted this rectification. It must be emphasized, however, that this was a mistake in our writing, and that the random baseline in the experiment was achieved as described, so our reported strengths and conclusions did not change. We sincerely thank the reviewers for pointing out this writing error. The reviewer is correct that inserting the same unrectified Gaussian value into both terms would cancel the update at uniform routing under the default scaling in original Eq. (16).

Specifically, sample ξt,v∼N(0,Idv)\xi_{t,v}\sim\mathcal{N}(0,I_{d_v}) and define [x]+=max⁡(x,0)[x]_{+}=\max(x,0). The two components are

\qquad [\widehat{A}^{-}_{t,v}]_{i}=[-\xi_{t,v,i}]_{+}.$$ We have replaced Eq. (16) with this positive/negative-part construction and clarified its conditions in Appendix A.1 (lines 806–812). This corrects the omitted rectification and makes the nonnegativity requirement in Section 3.3 explicit. Substituting these rectified components into Eq. (3) at uniform routing $p_{t,v,i}=1/d_v$ with $c_v=2/d_v$ yields\$ \\widehat{g}\_{t,v,i}=\\frac{1}{d\_v}\\left([\\xi\_{t,v,i}]\_{+}-[-\\xi\_{t,v,i}]\_{+}\\right)=\\frac{1}{d\_v}\\xi\_{t,v,i}\\neq 0\$, so the current random update does not vanish or suffer from suppression near uniform initialization. ---

Dear Reviewer 6s2z,

Thank you again for your time and constructive review. As the discussion period is drawing to a close, we would like to check whether our rebuttal and revised manuscript have addressed your concerns, and we would be happy to discuss any remaining questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Brief introduction to the revised PDF.

We thank all reviewers for their time and constructive feedback. We have revised the manuscript to address the reviewers' questions by incorporating additional validation and further clarifying the presentation of DNASNet. To facilitate your assessment, we summarize the main changes below:

  1. Figure 2(b) added energy and normalized deviation trajectories for a two-step weight-architecture variant without lookahead on DVS-Gesture, showing the impact of lookahead steps on search dynamics. The searched cells and retraining performance are added in Appendix C.6 (Table 5, Figure 9).

  2. We added controlled ablation experiments for timing information on CIFAR-10, CIFAR-100, and DVS-Gesture in Table 2, and explained the corresponding ablation search rules in the text (lines 423–427).

  3. Table 1 added first-order SNN-DARTS search and retraining results on DVS-CIFAR-10, and Figure 7(e) shows the searched cell.

  4. Revised Eq 16 clarifies the Random baseline formulation.

  5. The newly added lines 294–296 further clarified the exact conditional potential sequence optimized online during search and explicitly pointed to the existing online analysis in Appendix B.3.

  6. We refined the performance claims and updated the related work (lines 020–021, 067–069, 077–080).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.