acceptodds
Under review as a conference paper at ICLR 2027

TA-DMP-SNN: A Topology-Aware Dual-Memory Pareto Spiking Neural Search Network for Multiobjective Reed-Muller Logic Synthesis

Abstract

The ternary polarity space of Reed–Muller (RM) logic grows exponentially with the number of variables, while area, power, delay, and reliability objectives often conflict. We propose TA-DMP-SNN, a Topology-Aware Dual-Memory Pareto Spiking Neural Network for multiobjective RM logic synthesis. Unlike approaches that use neural networks only as quality predictors, TA-DMP-SNN directly employs an SNN as a discrete search mechanism over the ternary polarity space. Circuit topology, RM-term co-occurrence, and structural relations are encoded through topology-aware synaptic gating. Reference-vector routing, fast eligibility traces, and slow nondominated-archive memory jointly regulate exploration and consolidation. Across eight nontrivial RM benchmark circuits, TA-DMP-SNN achieved the highest median four-objective hypervolume after 100 unique evaluations, with a median absolute gain of 0.143 and a median improvement of 56.0% over the strongest baseline.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Spiking Neural Network,multiobjective,logic synthesis,Pareto front,

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_vXQmOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TA-DMP-SNN, a topology-aware spiking neural network for multio-bjective Reed-Muller logic synthesis, jointly optimizing area, power, delay, and reliability. The method searches the ternary polarity space by combining topology-aware synaptic gating, reference-vector guidance, fast eligibility traces, and slow memory derived from a nondominated archive. Experiments on eight nontrivial benchmark circuits report improved four-objective hypervolume after 100 unique evaluations, suggesting better search efficiency under a limited evaluation budget.

Critical strengths

The paper addresses a meaningful problem in logic synthesis by considering trade-offs among area, power, delay, and reliability. As described by the authors, the method uses circuit structure and nondominated archive information to guide search, with reported improvements under a limited evaluation budget.

Critical weaknesses
  1. Figure 1 is inconsistent with the method described in the text. It presents TA-DMP-SNN as an objective predictor integrated into evolutionary optimization, whereas Section 4 describes it as a candidate generator. The meanings of TA and DMP also differ between the figure and the text, making the framework difficult to follow.
  2. The slow-memory ablation simultaneously removes archive replay, marginal-frequency biases, and pairwise association updates. It therefore remains unclear which mechanism accounts for the improvement. Removing topology gating, fast memory, or entropy homeostasis produces no statistically significant degradation after correction, providing limited evidence for their individual contributions.
  3. The slow-memory ablation simultaneously removes archive replay, marginal-frequency biases, and pairwise association updates. It therefore remains unclear which mechanism accounts for the improvement. Removing topology gating, fast memory, or entropy homeostasis produces no statistically significant degradation after correction, providing limited evidence for their individual contributions.
  4. The experimental setup lacks important details, including benchmark sizes, complete parameter settings, and how AlphaSyn and SNN-RET are adapted to the ternary polarity space. These omissions make reproducibility and the fairness of the comparisons difficult to assess.
Questions for authors

My questions are reflected in the concerns raised in Critical Weaknesses.

Reviewer_Leg7OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TA-DMP-SNN for searching polarity assignments in mixed-polarity Reed–Muller logic synthesis under a limited evaluation budget. It uses three competing spiking neurons per variable, interactions weighted by circuit structure, and two types of memory that learn from recent feedback and an archive of non-dominated solutions. Reference vectors and entropy control guide exploration. The method optimizes structural proxies for area, power, delay, and fault exposure. After 100 unique evaluations, it reports the highest median four-objective hypervolume on eight of nine circuits, with a tie on the remaining circuit. The gains are strongest early in the search, and the ablations show the largest loss when the combined slow-memory components are removed.

Critical strengths

[1] The three-neuron encoding fits the ternary polarity space well. Using statistics from the nondominated archive to update the synapses is also an appealing way to guide future candidates.

[2] The paper reports 30 paired runs, anytime curves, and non-parametric tests with Holm correction. The authors also report that removing topology, fast memory, or homeostasis does not lead to a statistically significant loss under the tested budget.

Critical weaknesses

[1] Could you clarify the implementation and experimental setup? Figure 1 shows the SNN predicting objective values, while Section 4.1 says it generates candidates. Please clarify which version was used in the experiments. Details on RM conversion for large circuits, key settings, and how the baselines were adapted to this problem would also help. [2] Could you test the slow-memory components separately? The current ablation removes pairwise memory, reuse of archived solutions, and biases in state selection at the same time. Testing these separately would help identify where the gains come from. A comparison with a simple non-spiking method using the same archive information would also help clarify the benefit of the SNN. [3] Could you show the trade-offs among the four objectives? The HV values of 0.830 on some circuits do not show whether the method finds diverse trade-off solutions or solutions that improve all objectives together. The raw objective values and a comparison with the exact Pareto front obtained by checking all candidates on small circuits would help clarify this.

Questions for authors

See weaknesses above.

Reviewer_n3G3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TA-DMP-SNN for multiobjective mixed-polarity Reed–Muller logic synthesis. An SNN generates ternary polarity assignments using Boolean/RM-based coupling, reward-modulated fast updates, and slower learning and replay from a Pareto archive. Experiments on nine circuits with 30 runs per setting show improved hypervolume at small evaluation budgets against three baselines, using structural proxies for area, power, delay, and fault exposure.

Critical strengths
  • The reported hypervolume gains are substantial on several circuits under a small evaluation budget (Table 1). Repeated runs and anytime comparisons provide useful evidence of the method’s early-search advantage.
Critical weaknesses
  • It is hard to tell what the SNN adds beyond archive reuse. Table 2 shows no clear gain from topology gating or fast memory, while removing slow memory also removes replay and archive statistics together; a simple archive-based baseline would help separate these effects.
  • The paper needs more detail on how the baselines were adapted and how the circuits were evaluated. Figure 1 and Section 4 also disagree on whether the SNN predicts fitness or generates candidates.
Questions for authors
  • How does the method compare with a simple non-spiking sampler using the same archive information? Can you separate the effects of replay, marginal bias, and pairwise learning in the slow-memory component?
  • How were the three baselines adapted to ternary polarity search with four objectives? How were initialization and duplicate candidates handled when counting the evaluation budget?
  • Could you provide the exact circuit versions and sizes, evaluator settings, and raw objective values or Pareto fronts for the main comparisons?
  • Which description matches the implementation: the fitness-prediction pipeline in Figure 1 or the direct candidate generator in Section 4?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.