- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper investigates the trade-off between representational capacity and computational cost in multi-level spiking neural networks and proposes QuantLIF. It constrains spike amplitudes to sums of at most D signed power-of-two terms, allowing synaptic multiplication to be implemented through a bounded number of shifts and accumulations. Low-amplitude spikes are randomly suppressed during training and deterministically removed during inference to reduce the number of events. Classification experiments on CIFAR10/100, ImageNet-1K, and CIFAR10-DVS demonstrate promising accuracy, while estimates of synaptic arithmetic energy consumption also indicate some advantages.
Critical strengths
-
The method is well motivated. Directly connecting the representational capacity of multi-level spikes with an operation budget addresses an important challenge. The appendix also provides theoretical proofs with rigorous reasoning.
-
The paper is clearly written and well structured. The empirical results are strong, and the ablation studies are comprehensive.
Critical weaknesses
-
Incomplete baseline comparisons. Table 3 reports only the 77.33% TET result on CIFAR10-DVS, omitting the 83.17% result achieved with data augmentation, despite QuantLIF also using augmentation and TET loss. EventRPG reports 85.55% with Spike-VGG11 and 10 timesteps, exceeding QuantLIF's 84.10%, although its augmentation-based approach is not a directly controlled neuron comparison. Furthermore, PIT is included in Tables 2–3 but omitted from the controlled neuron comparison in Table 1. These omissions weaken the completeness of the experimental comparisons and require clarification.
-
Limited validation of practical energy efficiency and generalizability. Although QuantLIF demonstrates promising classification performance, its generalizability beyond classification remains untested. More importantly, Appendix F estimates synaptic arithmetic energy while accounting for the actual decomposition terms, but excludes quantization, decoding, communication, and memory-access costs. These additional costs could affect the claimed efficiency advantages. A more comprehensive energy analysis, ideally supported by hardware measurements, would strengthen the conclusions.
-
Minor presentation and terminology issues.Some figures, particularly Figure 4, contain small text. On page 6, the phrase "same D timesteps during training" is ambiguous because denotes the decomposition-term budget rather than simulation timesteps . The "(No Title)" reference entry on page 12 should also be corrected.
Questions for authors
-
Could the authors complete the comparisons across datasets and methods, including the latest relevant approaches where possible? After adding these comparisons, how should the scope of the paper’s SOTA claim be defined?
-
Could the authors provide comparative results on segmentation or tracking tasks to better demonstrate the effectiveness and generalizability of QuantLIF? Regarding energy estimation, please also explain the energy consumed by processing steps outside the neural network.
-
Why does a reference contain “(No Title)”?