- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes SpikingMAD, a multimodal anomaly detection framework targeting a favorable accuracy–energy trade-off. It combines EfficientNet feature extraction and feature jittering with a residual spiking reconstruction network (SMRN), and introduces spike-driven Hamming cross-attention (SHCA) for RGB–depth fusion. The latter is motivated by the limitations of binary dot products and a random binary embedding argument. Validation covers MVTec 3D-AD, Real-IAD D3, and Eyecandies, including image-level detection, pixel-level localization, few-shot settings, fusion ablations, and computational efficiency. The method consistently improves over the evaluated SNN adaptations. On MVTec 3D-AD, it reports 91.7% I-AUROC at four time steps with an estimated energy cost of 49.56 mJ.
Critical strengths
-
A practically relevant objective with consistent improvements over the evaluated SNN baselines. The paper explicitly studies energy efficiency in multimodal anomaly detection. Improvements extend across three datasets and both detection and localization metrics. For example, on MVTec 3D-AD at four time steps, SpikingMAD achieves 91.7% I-AUROC versus 87.3% for SNN-CFM and 84.7% for SNN-3D-ADNAS, with lower estimated energy. This supports the utility of the proposed architecture within the evaluated SNN setting.
-
The fusion ablation provides direct evidence for SHCA’s empirical benefit. Table 8 reports 89.6% I-AUROC with SHCA, compared with 86.9% using spike-driven dot-product attention and 87.8% using addition. Improvements also hold for P-AUROC and AUPRO, supporting the contribution beyond an overall architectural comparison.
Critical weaknesses
-
The theoretical motivation does not adequately justify the implemented SHCA operator. Proposition 1 assumes a shared i.i.d. Gaussian projection followed by sign binarization; the manuscript does not establish that the actual query/key generation satisfies these assumptions. Moreover, the result approximates a nonlinear transform of cosine similarity, rather than cosine similarity itself. More directly, Eq. (10) drops the additive (1/2) in Eq. (9). Although this preserves pairwise similarity rankings, it does not preserve the subsequent unnormalized attention output: it removes the value-dependent term (\frac12\mathbf{1}\mathbf{1}^{T}V), which can change the output spikes. These gaps weaken the theoretical justification of a central contribution.
-
The evidence does not yet establish the claimed end-to-end accuracy–energy advantage. Appendix B provides operation-based estimates rather than measured deployment energy. Although Eq. (14) includes an initial MAC component, the accounting for both EfficientNet encoders, neuron-state updates, SHCA, and the remaining inference pipeline is insufficiently explicit. In particular, transforming binary spikes into (2Q-1) and (2K-1) produces dense signed representations; multiplication-free arithmetic alone does not establish sparse event-driven execution. Comparisons also lack a matched ANN version of the proposed architecture to isolate the contribution of spiking computation. Accuracy losses against stronger ANN baselines are material: on Real-IAD D3, SpikingMAD reaches 77.8% I-AUROC and 74.9% P-AUROC, versus 84.1% and 92.2% for M3DM. Consequently, the practical trade-off remains insufficiently characterized.
Questions for authors
-
How are (Q) and (K) generated in the implemented network, and which assumptions of Proposition 1 hold? Is Eq. (10) intentionally centered Hamming attention? Please explain why removing the additive term is appropriate before value aggregation and spike thresholding.
-
What components are included in the reported energy, memory, and frame-rate figures? Please specify the hardware, measurement boundaries, firing-rate calculation, and operation accounting for SHCA, including contributions from zero-valued query/key entries.
-
Can you provide a matched ANN counterpart using the same encoders, reconstruction architecture, training protocol, and comparable fusion? This would help determine whether the reported advantage comes from spiking computation or from architectural differences between methods.