acceptodds
Under review as a conference paper at ICLR 2027

SkewSpike: A Dictionary-based Skewed-Tree Framework for Co-Firing Spike Compression in Spiking Neural Networks

Abstract

Spiking Neural Networks (SNNs) offer a promising energy-efficient alternative to conventional neural networks due to their event-driven execution. However, modern deep SNNs predominantly utilize rate coding, generating excessive spike volumes that saturate Address Event Representation (AER) communication channels on neuromorphic hardware. In this work, we introduce SkewSpike, a hierarchical dictionary-based AER spike compression framework designed to alleviate this communication bottleneck by exploiting skewed neuronal co-firing activity. SkewSpike maps recurring synchronous spike patterns into compact symbolic AER packets via a right-skewed tree dictionary, while further sparsifying residual communication traffic through stochastic spike pruning. To prevent significant accuracy degradation caused by residual spike pruning, the framework incorporates an adaptive threshold calibration mechanism that preserves rare features within the activity distribution, alongside a deterministic, model-agnostic autotuner that discovers optimal configurations for target AER compression ratios. We evaluate SkewSpike across diverse spiking architectures and datasets, including event-based neuromorphic benchmarks and full-scale ImageNet-1k. Experimental results demonstrate that SkewSpike substantially outperforms standard Huffman coding and existing AER compression baselines, achieving up to 75.41% AER event volume reduction on fully connected models with negligible accuracy loss, and maintaining 50.38% AER compression on ImageNet-1k with only a 2.05% accuracy drop.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Spiking Neural Networks, AER Compression, Energy Efficiency, Neuromorphic Computing

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_KZRJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

SkewSpike combines a skewed-tree dictionary, partial pattern matching, stochastic residual pruning, and threshold calibration to compress inter-layer spike traffic. Experiments evaluate compression accuracy trade-offs across several SNN architectures and datasets, alongside synthesized-router estimates of dynamic communication energy.

Critical strengths
  1. The paper directly targets inter-layer AER traffic, addressing a practical bottleneck beyond weight compression.
  2. The evaluation covers fully connected and convolutional SNNs on static-image and neuromorphic datasets.
  3. Component ablations and parameter sweeps expose important compression–accuracy trade-offs.
Critical weaknesses
  1. The autotuner uses task accuracy on 50% of the evaluation data, but final results are not clearly restricted to an independent, untouched test split.

  2. Dictionary replacement alone yields only 0.03% compression in Table 3; the skewed tree’s benefit over random pruning with the same threshold calibration remains unisolated.

  3. Compression accounting mixes event counts and bit lengths, while energy estimates omit codec and dictionary costs; net end-to-end efficiency gains remain unestablished.

  4. DVS128-Gesture accuracy drops by 34.73 percentage points at SS50, without deep-model uncertainty estimates; VGG11 compression also differs between Table 2 and Table 4’s counts.

  5. Appendix A.3 attributes FP-Max to an FP-Growth reference, while Appendix A.2 cites Werbos’s BPTT paper for the specific ATan surrogate, leaving these algorithm choices inaccurately supported.

Questions for authors
  1. Which samples and labels are used for dictionary construction, autotuning, and final evaluation, and can all methods be evaluated on the same untouched test split?

  2. Can you compare against random pruning with the same threshold calibration and traffic budget, and against balanced-tree or flat dictionaries, to isolate the skewed-tree contribution?

  3. What net latency and energy gains remain under common bit-level accounting that includes codec and dictionary costs? Does FSPC’s timing include pattern mining performed offline by SkewSpike?

  4. Why does VGG11/CIFAR-10 SS50 report 50.12% compression in Table 2 but approximately 54.59% from Table 4’s counts, and how stable are the deep-model results across seeds?

  5. How are empty transactions and p=0 handled, and how many spurious spikes does partial matching introduce? What denominator is used for the constant compression ratio in Figure 2(d)?

1 comment

We thank the reviewer for their critical review. We have addressed the comments with new no-dictionary baselines, held-out split validations, alternative dictionary benchmarks, multi-seed stochastic tests, and citation fixes in our revised manuscript PDF.

  1. Sample Usage, Test Split Isolation, and Generalization (Weakness 1, Q1): In our protocol, the first 50% of test batches forms calibration subset D for offline dictionary building and autotuning.

Final performance is evaluated across the full test set and validated on the strictly untouched 50% held-out split (Appendix A.9, Table 11, Page 20):

  • VGG11 (CIFAR-10): Seen 87.83%, Unseen 87.73%, Full 87.43% (Gap: 0.10 pp)
  • 4-Layer FC (N-MNIST): Seen 99.24%, Unseen 98.58%, Full 99.01% (Gap: 0.66 pp)
  • SEW-ResNet18 (IN-100): Seen 78.63%, Unseen 76.48%, Full 77.00% (Gap: 2.15 pp)
  • VGG16 (IN-100): Seen 60.04%, Unseen 54.44%, Full 57.52% (Gap: 5.60 pp) Minimal unseen gaps confirm zero test-set contamination.
  1. Isolating the Skewed Tree; No-Dict & Alternative Dictionaries (Weakness 2, Q2):

A. No-Dictionary Baseline at Equal 50% Budget (Table 3, Page 9):

  • SEW-ResNet18 (IN-100): SkewSpike 77.36% (50.01% comp); Random+Calib (w=0.4) 76.70% (50.00% comp); Pure Random (w=1.0) 69.56% (50.00% comp).
  • VGG16 (IN-100): SkewSpike 57.60% (50.00% comp); Random+Calib (w=0.6) 57.38% (50.00% comp); Pure Random (w=1.0) 23.32% (50.00% comp).

Takeaway: Calibration prevents collapse (+34.06 pp VGG16, +7.14 pp ResNet18), while the dictionary offers a modest, consistent accuracy margin (+0.22 to +0.66 pp in CNNs) and +2.94% extra structural compression on MNIST.

B. Alternative Dictionaries on SEW-ResNet18 (Appendix A.8, Table 10, Page 20):

  • Right-Skewed (Proposed): 8.98s build, |V|=5 symbols, 260.38 ms/batch, 78.58% acc.
  • Balanced Binary Tree: 10.98s build, |V|=7 symbols, 201.53 ms/batch, 79.08% acc.
  • Flat (64 clusters): 56.82s build, |V|=64 symbols, 145.09 ms/batch, 78.70% acc. Flat clustering is >6x slower with 12.8x larger symbol memory (|V|=64 vs 5) and no accuracy gain; the right-skewed tree isolates the core into just 5 symbols.
  1. Handling p=0, Spurious Spikes & Fig. 2d Denominator (Q5):
  • Empty Transactions & p=0: Eq. 8 (Page 5) now enforces I[|E_r cap P_v|>0], guaranteeing zero-spike transactions never trigger symbols. Re-tuning VGG16 (IN-100) with non-zero p selected (m=0.4, p=1.0, sigma=0.5, w=0.6), yielding 57.60% acc at 50.00% comp (+12.94 pp over Table 2's prior 44.66%).
  • Fig. 2d Denominator: Eq. 11 computes interface compression vs. incoming calibrated stream |E_calib|. Lowering w boosts rare spikes (|E_calib| > |E_orig|). Appendix A.10 (Table 12, Page 20) reports net reduction vs. |E_orig|: 50.01% at w=1.0 (69.68% acc), 30.59% at w=0.6 (78.48% acc, within 1.02 pp of baseline), and 11.00% at w=0.4 (77.28% acc).
  1. Multi-Seed Stability Across Random Draws (Weakness 4, Q4): We evaluated stochastic stability across 4 Bernoulli pruning seeds (10, 42, 60, 85) on fixed checkpoints (Appendix A.7, Table 9, Page 19):
  • SpikingVGG11 (CIFAR-10, SS50): 87.71 +/- 0.07% acc, 50.13 +/- 0.00% comp.
  • SEW-ResNet18 (IN-100, SS50): 78.84 +/- 0.32% acc, 50.01 +/- 0.00% comp. Across millions of spikes, volume variance is zero (+/-0.00%), and accuracy stays tightly bounded. DVS-Gesture drops past SS30 because plain VGG lacks skip connections against spike starvation (Sec. 4.2).
  1. Reconciliation of Table 2 vs. Table 4 (Weakness 4, Q4): Table 2 reports per-batch average compression (50.12%), whereas Table 4 reports cumulative dataset-wide counts from seed 42 (109.96M -> 49.93M spikes, 54.59% reduction) from the standalone hardware-trace export for QuestaSim/PrimeTime simulation (clarified on Page 9).

  2. Bibliographic Corrections (Weakness 5): (1) FP-Max now cites Grahne & Zhu (2003) (Page 11, Line 583; Page 15, Line 757). (2) ATan surrogate cites Fang et al. (2021) (Page 11, Line 569; Page 13, Line 696), retaining Werbos (1990) solely for BPTT.

  3. Bit-Level Accounting & Codec Overhead (Weakness 3, Q3): AER spikes and symbols are both 16-bit flits (Table 8), making event reduction equal to bit reduction. For N=10,000 and |V|=5 symbols (Table 10), dictionary SRAM requires only 6.10 KB (<156.25 KB even at |V|=128). FSPC latency (2.38-17.70s/batch) is pure online CPU inference; SkewSpike executes in 2.36-2.79 ms on GPU.

Reviewer_3CGWOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposed SkewSpike, a spike compression approach for AER in neuromorphic hardware.

Critical strengths

None. Due to the topic for this work is quite irrelevant to my own research. I therefore has several concerns below in the Critical Weaknesses and Questions column.

Critical weaknesses
  1. I believe this work would be better suited for submission to journals and conferences in the hardware field; the hardware-software co-design for AER is likely to attract interest and gain recognition within those venues. It is, however, less appropriate for ICLR. (importance: 10/10)
  2. The primary motivation of this work is to alleviate the communication bottleneck in Address-Event Representation (AER) spiking systems. Therefore, in addition to compression ratio, energy consumption, and task accuracy, the framework’s communication efficiency and net system energy-efficiency gains on actual hardware should be quantified. It is recommended to provide further details regarding the implementation and overhead of the encoder, decoder, and dictionary storage—including estimates for synthesized area, critical path, and power consumption under specified process libraries, voltage levels, and clock constraints—as well as processing latency, sustained throughput, and buffer requirements for pattern matching and spike reconstruction. (importance: 9/10)
Questions for authors

Clarification requests:

  1. Section 3.5 uses 50% of the evaluation data for autotuning, but the rationale for this calibration budget is unclear. Could the authors justify this choice and report the absolute number of calibration samples? How do the trade-off between accuracy and compression, and the tuning cost, change with smaller calibration subsets?
  2. Are the reported AER reductions calculated from event counts or actual transmitted bits/flits, including symbol identifiers, packet headers, and routing overhead? Suggestions for this and future work:
  3. The layout in Figure 1 is atrocious. There is far too much white space around the images, resulting in a significant waste of space. It fails to convey that the work itself is sufficiently substantial.
3 comments

We thank the reviewer for their comments. We address the scope, hardware accounting, calibration budget, and formatting points below.

  1. Suitability for ICLR (Weakness 1): We respectfully believe this work fits ICLR's scope. Efficient learning, sparse representations, and spiking neural networks (SNNs) are relevant to ICLR and NeurIPS (e.g., Shi et al., ICLR 2024; Bu et al., NeurIPS 2025). ICLR's scope includes hardware and infrastructure. SkewSpike contributes algorithms for exploiting neuronal co-firing, preserving rare features during pruning, and model-agnostic autotuning of compression configurations, evaluated across architectures through ImageNet-1k.

  2. Bit-Level Accounting & Codec Overhead (Weakness 2, Q2):

  • Flit Equivalence: As specified in our 45-nm Nangate synthesis setup (Section 4.4, Appendix A.6, Table 8: 500 MHz, 1.1V), both raw AER packets and compressed dictionary symbols are packed into identical 16-bit flits (B_AER = 16). Thus, event reduction translates 1:1 to bit/flit reduction.
  • Dictionary Storage: Under our right-skewed tree, the active vocabulary requires as few as |V|=5 symbols (Table 10, Page 20), occupying only 6.10 KB of on-chip SRAM for a 10,000-neuron layer (<156.25 KB worst-case at |V|=128). While full RTL physical layout of the codec itself is outside the algorithmic scope of this ML paper, our existing post-synthesis NoC evaluation quantifies net dynamic energy savings.
  1. Calibration Budget Rationale (Q1): Using 50% of the evaluation split (e.g., 2,500 images on ImageNet-100; 5,000 on CIFAR-10; 25,000 on ImageNet-1k) balances statistical stability for rare-neuron estimation against offline search runtime (1.3 min on MLPs, 10.8 min on ImageNet-100; Table 7). In our revised PDF (Appendix A.9, Table 11), we further verified generalization strictly on the unseen 50% held-out split.

  2. Figure 1 Layout: In the revised PDF (Page 4), we tightened Figure 1's bounding box and removed excess whitespace.

Reviewer_3CGWOpenReview ↗

Follow-up Reply and Concerns

I appreciate the author respond the original concerns in my review. Nonetheless, even if this work is relevant to ICLR Conference, I still have doubts on the following points:

  • The central empirical premise (skewed firing and frequent co-firing in trained SNNs) is asserted but never demonstrated clearly. There are no firing-rate distribution is shown for any evaluated layer or model (e.g., per-neuron firing-rate histograms, or the fraction of total spikes produced by the top-k% of neurons). No co-firing statistics are reported either. This make it less believable cuz the starting point isn't clear at all.

  • Why bother using Dictionary since the training process contains regularizer to prevent excessive co-firing. Appendix A.3 states that an activity regularizer is applied to the first hidden layer "to prevent excessive co-firing", which works against the very property the method relies on.

  • The Fig.1 is still too many blanks, not informative at all. I sincerely suggest the author break down the two subgraphs and place them tightly, neatly and correctly in the manuscript.

Above these concerns not addressed, it's not necessary to discuss the rest of the method and the experiments. Overall, this work should go into a more thorough solid foundation building and well-understood polishing, including the figures, the notations, the tabulars and the logic behind it.

We thank the reviewer for the follow-up feedback and for acknowledging the relevance of our work to ICLR. We have directly addressed your points regarding the empirical premise, the Appendix A.3 regularizer clarification, and the figure layout in our revised manuscript PDF.

  1. Empirical Demonstration of Skewed Firing and Co-Firing Statistics (Point 1): To remove any ambiguity regarding our starting premise, we analyzed the uncompressed test-set activations across our evaluated models to extract exact firing-rate distributions, Gini inequality metrics, and co-activation statistics (integrated into Section 3.1, Page 4 and Appendix A.11, Table 13):
  • 4-Layer FC (MNIST, N=10,000): Mean rate r=0.0633 (6.33%), Gini=0.7014. Top 1% of neurons emit 15.11% of all spikes, Top 5% emit 40.16%, and Top 20% emit 70.74% of all layer traffic, while 10.59% of neurons are completely silent. Mean synchronization = 632.63 co-firing neurons/tx.
  • SpikingVGG11 (CIFAR-10, N=16,384): Mean rate r=0.0355 (3.55%), Gini=0.4470. Top 1% emit 4.82% of spikes, Top 5% emit 17.55%, and Top 20% emit 47.21% of all spikes, with 4.46% silent. Mean synchronization = 632.06 co-firing neurons/tx.
  • SEW-ResNet18 (ImageNet-100, N=200,704): Mean rate r=0.1073 (10.73%), Gini=0.4126. Top 1% emit 3.68% of spikes, Top 5% emit 16.41%, and Top 20% emit 46.69% of all spikes. Mean synchronization = 21,526.29 co-firing neurons/tx.

Takeaways: (1) The Gini coefficients (0.41–0.70) quantify clear population inequality, with the top 20% of neurons producing ~47% of spikes in deep ConvNets and >70% in MLPs. (2) Across active transactions, neurons co-fire in large simultaneous bursts (mean sync = 632 to 21,526 neurons/tx), confirming that AER buses carry dense synchronous wavefronts rather than independent Poisson spikes.

  1. Clarification on Appendix A.3 Regularization ("Excessive Co-Firing"): We thank the reviewer for catching this phrasing confusion, which we have now corrected in Appendix A.3 of the revised PDF:
  • Scope: This regularizer applies STRICTLY to the small 100-neuron networks (784-100-10 and 784-100-60-10) in Section 4.1 / Appendix A.3, which were retrofitted from FSPC's original snnTorch code into SpikingJelly. None of our main scalability architectures (SpikingVGG11, SpikingVGG16, SEW-ResNet18, SpikingResNet18, or the 10k-5k MLPs) use this regularizer; their training recipe is standard AdamW without any activity penalty (Appendix A.5).
  • Technical Reason (Porting & FP-Max Tractability): When porting FSPC's snnTorch baseline into SpikingJelly (using ATan surrogate gradients and unregularized Adam without weight decay), Layer 1 exhibited dense firing saturation. Because FSPC relies on combinatorial FP-Max mining (O(2^N) complexity), dense cliques caused FP-Max inference to exceed hundreds of seconds per batch. We applied the ceiling penalty (10.0 * (r - 0.2)^2 for r > 0.20) strictly to match the original snnTorch sparsity regime and keep FP-Max tractable (Table 5).
  • Population Sparsity vs. Co-Firing: Even in those baseline models, capping the average rate at r <= 0.20 prevents homogeneous saturation rather than eliminating correlated subsets, preserving the sparse, skewed firing distribution that SkewSpike operates on.
  1. Redesign of Figure 1: We appreciate the visual layout feedback. We have redesigned Figure 1 on Page 4 of the revised PDF, tightening the spacing between the hardware transceiver diagram (a) and the operational flowchart (b).

We hope these empirical distributions, the clarification on why the regularizer was restricted to the snnTorch-to-SpikingJelly baseline port, and the tightened Figure 1 address your foundational concerns.

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

The paper proposes SkewSpike, a lossy method to reduce AER traffic between layers of rate-coded SNNs. Offline, neurons with firing rate of at least m are grouped with bisecting k-means into a right-skewed tree of co-firing patterns, and each pattern gets one symbol. The input current of rare neurons is also scaled by V_th/ω so they fire more. Online, a symbol is sent when at least a fraction p of its neurons fire, each remaining spike is kept with probability σ, and the receiving layer turns symbols back into spikes. A beam search autotuner picks (m, p, σ, ω) for a target event reduction of 30%, 50% or 70%. The method is compared with FSPC and Huffman coding on small MLPs and tested on 12 benchmarks up to SpikingResNet18 on ImageNet-1k, with ablations and 45-nm NoC energy estimates.

Critical strengths

AER bandwidth is a real limit in neuromorphic hardware, and on small MLPs the dictionary does help. For 784-100-10 on MNIST, σ = 0.8 alone gives about 20% while SkewSpike gives 32.16% (Table 1). The runtime in Table 5 is also promising (2.36 to 2.79 ms per batch against 2.38 to 17.70 s for FSPC), although the paper does not show that the two implementations do the same work. Threshold calibration is useful in practice too, since on SEW-ResNet18/ImageNet-100 it recovers 7.64 pp at about 50% compression (Table 3). I also appreciate that the authors report their failure cases openly.

Critical weaknesses

My main concern is that the benefit of the right-skewed dictionary on deep networks is not shown, and the scalability claim depends on exactly these networks. In Table 3 the dictionary alone reduces events by only 0.03% on SEW-ResNet18/ImageNet-100. In all 27 conv configurations the compression is within 2.5 pp of (1-σ) (Tables 2 and 6), so random spike removal can explain almost all of it. On the MLPs calibration is off (m = 0 and ω = 1 in Table 6), so the components are never shown working together. There is no baseline without the dictionary at the same event budget, so the experiments do not show that the dictionary improves the accuracy vs compression trade-off.

The dictionary construction and novelty also need more support. The k-means distance is not stated. With Euclidean distance two sparse neurons look close even if they never fire together, which matters when m = 0 as in all MLP runs. The score W_c = V_c · H_c does not use p, so a high score does not guarantee frequent matches, and there is no comparison with simpler options like a flat dictionary or a co-activation (Jaccard) similarity. On prior work, FSPC (Ganesh et al., 2025) already merges co-firing spikes into symbols with partial matching. Phase 2 is close to threshold regularization in Lee et al. (2016) and automated threshold search in Wang et al. (AAAI 2025), and neither is cited, so the authors should explain what is technically new in Phase 2.

I also have doubts about the evaluation and the energy numbers. The autotuner picks parameters using accuracy on 50% of the evaluation data, and the paper does not say if the final accuracy excludes these samples. FSPC parameters are picked on test accuracy with a different objective (App. A.3), which makes Table 1 hard to read. Only one layer interface is compressed, and in the ResNets this is layer1[0].sn1 in the residual branch, so the uncompressed identity shortcut may explain the ResNet resilience in Sec. 4.2. The conv models use one seed, although random removal changes with every Bernoulli draw. For energy, Table 4 gives the same ratio for events and energy (0.2880 on MNIST, 0.200 on N-MNIST), even though symbols route through many PEs (App. A.6) and on MNIST at the 70% target σ is 0.8 to 1.0, so symbols should give most of the reduction. The encoder, decoder, dictionary storage, compression-side buffers and random number generator are not counted either, so end-to-end savings are not shown.

Questions for authors
  1. For each row of Table 2, what share of the removed events comes from symbols and what share from random removal, and what fraction of spikes do the symbols cover? Please also report accuracy at the same event budget for random removal plus calibration without the dictionary, for FSPC patterns in place of your tree on the same layers, and for fewer timesteps or firing rate regularization (for example Bu et al., 2025). If the no-dictionary baseline gets similar accuracy, the dictionary does not improve the trade-off.
  2. Is the reported accuracy measured only on the half of the evaluation data that the autotuner did not use? Also, in Fig. 2d compression stays near 50.01% for all ω from 0.2 to 1.0, but current scaling should change the spike count. Is |E| counted before or after calibration? Event counts before calibration, after calibration and after compression would make this clear.
  3. With p = 0, Eq. (8) lets every pattern match even when none of its neurons fire, so the decoder could create spikes that were never there. Table 6 uses p = 0 for VGG16 on ImageNet-100 (SS50), and in Fig. 2b p = 0 gives about 60% compression but only about 43% accuracy. Does your code require at least one active neuron for a match?
  4. Why does the energy per event stay almost the same when symbols can take different routes than single spikes? Does H_s include delivery to every PE a symbol must reach? Please also give the energy cost of the encoder, decoder and dictionary storage.
1 comment

We thank the reviewer for the constructive and thorough critique. We have implemented the requested no-dictionary baselines, resolved the pattern-matching edge case, verified test-split isolation, tested multi-seed stochastic stability, extracted exact event counts across threshold factors, and uploaded a revised manuscript PDF.

  1. No-Dictionary Baseline & Component Isolation (Weaknesses 1 & 3, Q1) We evaluated pure stochastic Bernoulli spike removal at the identical 50% event reduction budget, with and without threshold calibration (now Table 3 for SEW-ResNet18, Page 9 in revised PDF):
  • SEW-ResNet18 (IN-100): Full SkewSpike: 77.36% acc (50.01% comp); No-Dict + Calib (w=0.4): 76.70% acc (50.00% comp); Pure Random Pruning (w=1.0): 69.56% acc (50.00% comp).
  • SpikingVGG16 (IN-100): Full SkewSpike: 57.60% acc (50.00% comp); No-Dict + Calib (w=0.6): 57.38% acc (50.00% comp); Pure Random Pruning (w=1.0): 23.32% acc (50.00% comp).
  • SpikingVGG11 (CIFAR-10): Full SkewSpike: 87.63% acc (50.14% comp); No-Dict + Calib (w=0.6): 87.45% acc (50.00% comp); Pure Random Pruning (w=1.0): 79.47% acc (50.00% comp).
  • 4-Layer FC (MNIST, 30%): Full SkewSpike: 98.47% acc (32.94% comp); Pure Random Pruning: 98.55% acc (30.00% comp).

Coverage & Removal Breakdown (Q1): On deep models at SS50, dictionary substitution accounts for 0.03%–0.14% direct structural reduction (Table 3 & VGG11), while stochastic residual pruning accounts for ~49.9%–50.0% (tracking 1 - sigma). On MLPs (MNIST), symbols directly drive 2.94% structural event reduction (32.94% vs. 30.00%).

Takeaways:

(1) Calibration prevents catastrophic collapse (+34.06 pp on VGG16, +7.80 pp on ResNet18 over pure random pruning).

(2) While the dictionary's accuracy margin on deep models is modest (+0.22 to +0.66 pp), it consistently avoids representational distortion across all architectures. It directly drives structural event reduction on fully connected networks (32.94% vs. 30.00% on MNIST). Crucially, our right-skewed tree restricts this active vocabulary to an ultra-compact |V|=5 symbols (6.10 KB SRAM; App. A.8), ensuring that symbolic aggregation introduces negligible hardware overhead.

  1. Resolution of p=0 Corner Case (Weakness & Q3) In Eq. 8 (Page 5), we added non-zero gating, guaranteeing non-firing transactions never trigger symbols. Re-tuning SpikingVGG16 (IN-100) with p in [0.1, 1.0] selected (m=0.4, p=1.0, sigma=0.5, w=0.6), achieving 57.60% accuracy at 50.00% compression (+12.94 pp higher than 44.66% reported in Table 2). Tables 2 and 6 are updated.

  2. Event Counting Denominators in Fig. 2d (Q2) In Eq. 11, compression was computed at the inter-layer interface relative to the incoming calibrated stream |E_calib|. Lowering w injects temporal redundancy into rare neurons (|E_calib| > |E_orig|). In Appendix A.10 (Table 12, Page 20), we report the full breakdown across w on SEW-ResNet18:

  • At w=1.0 (uncalibrated), |E_calib| = |E_orig|, achieving 50.01% net reduction vs. original baseline at 69.68% accuracy.
  • Selecting w=0.6 achieves 30.59% net reduction vs. original baseline with 78.48% accuracy (within 1.02 pp of baseline).
  • At w=0.4 (anchor), the network prioritizes peak accuracy (77.28%) with 11.00% net reduction. We clarified this interface vs. net trade-off on Page 5 and Appendix A.10.
  1. Validation on 50% Held-Out Split (Weakness 6, Q2) Evaluating autotuned models on the untouched 50% test split confirms zero data contamination (App. A.9, Table 11):
  • SpikingVGG11 (CIFAR-10): Seen 50%: 87.83%, Unseen 50%: 87.73% (Gap: 0.10 pp).
  • 4-Layer FC (N-MNIST): Seen 50%: 99.24%, Unseen 50%: 98.58% (Gap: 0.66 pp).
  • SEW-ResNet18 (IN-100): Seen 50%: 78.63%, Unseen 50%: 76.48% (Gap: 2.15 pp).
  • SpikingVGG16 (IN-100): Seen 50%: 60.04%, Unseen 50%: 54.44% (Gap: 5.60 pp, reflecting feedforward sensitivity without skip connections).
  1. Stochastic Stability: Across 4 Bernoulli seeds (10, 42, 60, 85), VGG11 achieves 87.71±0.07% acc (50.13±0.00% comp) and SEW-ResNet18 achieves 78.84±0.32% acc (50.01±0.00% comp) (App. A.7, Table 9).

  2. Metric, Novelty, and Hardware Overhead (Weaknesses 4, 5, 10, Q4)

  • Distance Metric: PyTorch uses L2 distance. On binary vectors, ||x - y||_2^2 = d_Hamming(x,y), directly clustering co-activation overlap without custom kernels (Page 4).
  • Novelty: Unlike global threshold sweeps (Lee et al. 2016; Wang et al. AAAI 2025), Phase 2 is an offline input-gain scale applied strictly to rare tail neurons (n not in F) (Page 5).
  • Hardware & NoC Routing: Yes, H_s for composite symbols explicitly models delivery to all destination PEs via nearest-neighbor TSP routing (App. A.6). Both AER packets and symbols are 16-bit flits (B_AER=16), so event reduction equals bit reduction. Local PE placement keeps average hops invariant, ensuring proportional dynamic energy savings (Table 4). Dictionary SRAM requires only 6.10 KB for |V|=5 (<156.25 KB even at |V|=128) for N=10,000.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.