acceptodds
Withdrawn from ICLR 2027

Towards an Energy-Efficient Anomaly Detector: When Multimodal Anomaly Detection Meets Spiking Neural Network

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent advances in multimodal anomaly detection (MAD) primarily rely on more sophisticated and powerful artificial neural networks (ANNs) to improve detection performance. However, less attention has been paid to the fact that ANN-based paradigms often require high energy consumption to achieve these performance gains. This gap raises a key challenge: Can we significantly reduce energy consumption while maintaining competitive performance? Spiking neural networks (SNNs), which employ binary spikes and event-driven computation, have emerged as a promising energy-efficient alternative to conventional ANNs. Motivated by this, we propose SpikingMAD, a spike-driven MAD framework designed to achieve both low power consumption and high performance. Specifically, we develop a spike-driven multimodal reconstruction network that restores abnormal features to normal ones through spike-sparse computation and multiplication-free operations. Furthermore, given that dot-product attention fails to measure the similarity between RGB and depth features encoded by spiking neurons, we introduce a spike-driven Hamming cross-attention mechanism to capture cross-modal feature similarity and facilitate multimodal feature fusion. Extensive experiments on MVTec 3D-AD, Real-IAD D3, and Eyecandies demonstrate that SpikingMAD achieves competitive performance in terms of accuracy, memory usage, frame rate, and energy consumption, while also showing strong potential in few-shot scenarios. Code: https://anonymous.4open.science/r/SpikingMAD.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Anomaly Detection, Unsupervised Industrial Anomaly Detection, Spiking Neural Network

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_EnG1OpenReview ↗
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
  1. 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.

  2. 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
  1. 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.

  2. 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
  1. 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.

  2. 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.

  3. 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.

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

This paper investigates low-power multimodal anomaly detection based on Spiking Neural Networks (SNNs) and proposes the SpikingMAD framework, which aims to reduce the computational and energy costs associated with traditional ANN-based anomaly detection models while maintaining high detection accuracy. The method comprises two core modules: (1) the Spike-driven Multimodal Reconstruction Network (SMRN) and (2) Spike-driven Hamming Cross-Attention (SHCA).

Critical strengths
  1. This paper proposes a novel SpikingMAD method to steer the architecture with superior performance and lower energy consumption.
  2. This paper designs a spike-driven multimodal reconstruction network (SMRN) with a residual topology for abnormal image restoration.
  3. This paper proposes a spike-driven Hamming cross-attention (SHCA) to overcome the limitation of dot-product attention in measuring similarity between RGB and depth features encoded by spiking neurons
Critical weaknesses
  1. The hardware deployment of this hybrid ANN-SNN network has not been adequately discussed.
  2. The methodological innovation is relatively limited. SMRN primarily employs existing spiking neuron models, residual connections, and reconstruction network structures, while SHCA leverages established concepts of Hamming similarity and binary feature matching. The overall framework largely represents a combination and application of existing techniques to the task of multimodal anomaly detection, lacking breakthroughs in spiking computation mechanisms or network architecture innovation.
  3. There is a lack of performance comparison with ANN models sharing the same architecture. It is recommended to conduct a rigorous comparison between ANN and SNN versions of the same architecture to further isolate the actual benefits derived from the spiking computation mechanism.
  4. Comparisons with current SNN-based backbones are also missing.
Questions for authors

see above

Reviewer_SBvBOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SpikingMAD for energy-efficient multimodal anomaly detection. EfficientNet extracts RGB and depth features, and a spiking network reconstructs normal features with feature perturbations during training. A Hamming cross-attention module combines the two modalities. Experiments on MVTec 3D-AD, Real-IAD D3, and Eyecandies report detection and localization performance, estimated energy, memory usage, and frame rate. SpikingMAD improves over the tested SNN baselines, but remains below stronger ANN methods in detection accuracy.

Critical strengths
  1. The figures and tables are carefully prepared. The evaluation covers three datasets, different modality settings, two time-step settings, and few-shot training.
  2. The energy-accuracy trade-off is worth studying. Reporting memory usage and frame rate alongside detection performance makes the comparison more useful.
Critical weaknesses
  1. The motivation mainly relies on the general energy advantage of SNNs. The paper does not explain clearly which difficulty in multimodal anomaly detection this design addresses, or how it compares with a lightweight ANN.
  2. SHCA closely follows the Hamming attention in SpikeVideoFormer (ICML 2025). The binary-embedding argument, centered Hamming formulation, and computation rearrangement already appear there. The technical addition beyond applying them to RGB and depth features is not clearly established.
  3. Proposition 1 assumes a shared Gaussian random projection. The implemented embeddings come from two learned reconstruction branches.
  4. The energy estimate needs a breakdown for the EfficientNet encoders, reconstruction blocks, SHCA, and other inference operations. SHCA converts zero spikes to −1, so its cost cannot be inferred from spike sparsity alone.
  5. The largest reported energy saving comes with a clear accuracy loss. Table 7 gives 89.6% I-AUROC for SpikingMAD, versus 95.1% for 3D-ADNAS and 95.4% for CFM. Without a matched ANN version, the results do not show whether spiking offers a better trade-off.
  6. Feature jittering, residual reconstruction, segmentation-branch removal, and spiking computation are introduced together. Table 8 compares fusion operations, but does not separate the effects of these other changes.
  7. The visual analysis does not explain the method sufficiently. There is no matched comparison showing how the reconstruction changes, when cross-modal fusion helps, or where the proposed method fails.
Questions for authors

see weaknesses

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

Related papers

Discussion (0)

Sign in to comment.