- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
The paper studies MUAD on the setting of hundreds of categories. A series of scaling experiments show that anomaly-region reconstruction error keeps falling while normal-region error plateaus or rises. The so-called "capacity saturation" complements the usual over-generalization story. To this end, the authors propose two components: ShareMLP, an image-level (CLS-routed) dynamic FFN whose weights are soft mixtures of a weight pool shared across all decoder layers, and the Prior-Preserving Tight Bottleneck (PTB), a 32-d noisy bottleneck that re-injects positional embeddings and passes the CLS token through a separate noise-free path. The proposed method reports 94.6% I-AUROC on Real-IAD Variety (+5.5 over Dinomaly2) and 89.1% on ADNet (+7.1), with gains on Real-IAD and MANTA and parity on MVTec-AD/VisA.
Critical strengths
- The diagnosis is the paper's best contribution. Separating anomaly-side and normal-side reconstruction error under a fixed evaluation set is a clean experimental design. And the ablations are thorough: component ablation, static widening up to r=32, sparse MoE, CondConv/DY-Conv/KW-Conv, matched-parameter DynaMLP vs ShareMLP, PTB design, dropout, pool size, encoders, model scale, and plug-in to RD4AD/ViTAD/Dinomaly.
- Large and consistent gains: 160- and 380-class benchmarks, across image-, object- and pixel-level metrics. The 160-class multi-class result (94.6) nearly matches Dinomaly2's one-model-per-class number (94.9).
- Compute cost is well controlled. ShareMLP at 8× FFN parameters but ~1× FFN FLOPs beats static MLP at 8× FLOPs (92.00 vs 90.64), and total FLOPs rise only from 26.2G to 27.6G.
Critical weaknesses
- Technical novelty is modest. Input-conditioned weight mixtures (CondConv, DY-Conv) and cross-layer weight templates (Savarese & Maire; KW) are established; the contribution is choosing image-level routing and sharing for UAD, plus a narrower bottleneck.
- The "capacity saturation" evidence rests on one baseline, one dataset, one encoder size and normalized errors from single runs. The curves in Fig. 1(b) are fairly flat; it is not shown that the trend is beyond seed noise.
Questions for authors
- Were d_t, dropout, N and the stopping iteration selected on the Real-IAD Variety test split? If so, please report results with hyperparameters chosen on a held-out validation split, or on a different dataset and then transferred.
- Is it capacity or interference or optimization? Fig. 1(b) is consistent with several explanations: finite capacity, gradient interference between classes, or fewer effective updates per class. Appendix A.3 says models are trained "up to 200,000 iters"; was the per-class number of updates held fixed across C? A run of static MLP r=32 through the same 20→160 protocol would directly test whether pure width removes the normal-side saturation.
2 comments
Response to Reviewer zfFM (part2)
Q1.
Were d_t, dropout, N and the stopping iteration selected on the Real-IAD Variety test split? If so, please report results with hyperparameters chosen on a held-out validation split, or on a different dataset and then transferred.
Answer:
Thank you for this careful question. Most hyperparameters are inherited or set by default. Following your suggestion, we transferred the bottleneck d_t to a different dataset and report the results below.
-
Inherited without tuning. The dropout rate (0.4), training schedule (100k iterations), learning rate, betas, and weight decay directly follow Dinomaly2's settings on Real-IAD, without any adjustment. We report the final checkpoint at the end of the fixed schedule.
-
Not selected by performance. Performance increases monotonically with larger pool size N (Table A6), so N is not tuned for optimality. We set N=64 (8× FFN parameters) as a trade-off between accuracy and parameter count; N=96 performs better but requires more parameters.
-
Roughly set and robust. Since UAD benchmarks provide no validation split due to the unsupervised setting, d_t=32 was roughly set by observing results on Real-IAD Variety, without a fine-grained search. Performance is robust over a wide range of d_t, and we transfer the same setting to a different dataset, ADNet (Meganomaly-S). On ADNet, the variation across d_t from 16 to 48 is within 0.3% I-AUROC, and our default d_t=32 is not even the best setting.
| d_t | Real-IAD Variety | ADNet | ||||
|---|---|---|---|---|---|---|
| I-AUROC | P-AUROC | AUPRO | I-AUROC | P-AUROC | AUPRO | |
| 16 | 93.9 | 95.3 | 83.6 | 85.4 | 94.0 | 82.7 |
| 32 (default) | 94.0 | 95.5 | 84.6 | 85.5 | 94.1 | 83.2 |
| 48 | 93.8 | 95.4 | 84.6 | 85.7 | 94.1 | 83.4 |
| 64 | 93.3 | 95.4 | 84.8 | 85.1 | 93.8 | 82.9 |
Q2.
Is it capacity or interference or optimization? Fig. 1(b) is consistent with several explanations: finite capacity, gradient interference between classes, or fewer effective updates per class. Appendix A.3 says models are trained "up to 200,000 iters"; was the per-class number of updates held fixed across C? A run of static MLP r=32 through the same 20→160 protocol would directly test whether pure width removes the normal-side saturation.
Answer:
Thank you for this insightful question and the concrete experimental suggestions. Following them, we ran the static MLP (r=32) through the same 20→160 protocol (revised Fig. A1) and added a control for per-class training budget (new Fig. A3).
Pure width alleviates the saturation. As shown in revised Fig. A1, widening the static MLP to r=32 clearly alleviates the normal-side saturation: compared with the baseline, the normal-region error keeps decreasing over more categories and reaches a lower level, and the I-AUROC drop from 20 to 160 categories shrinks from 7.4% to 4.1%. Since pure width changes nothing but capacity, this directly confirms the existence of capacity saturation.
Capacity and interference are, in our view, two sides of the same coin: interference arises because all categories compete within finite shared weights, as introduced in our Abstract and Sec. 3.1: "diverse normal patterns compete and interfere within the decoder's finite, shared weights." ShareMLP, at the same parameter budget, provides far more weight combinations and lets different categories route to different combinations, thereby both expanding capacity and reducing interference. Accordingly, it sustains the decrease in normal error more effectively.
Full convergence and proportional update budget. In our original setting, for each category configuration, we trained the model for up to 200k iterations (twice the default schedule), evaluated the E_norm every 10k iterations, and selected the checkpoint at which it stopped decreasing. This ensures that every configuration is compared at its fully converged state (no anomaly labels are used, only for diagnostic analysis).
As you suggested, in new Fig. A3, we further compare checkpoints at training budgets proportional to category count ( e.g., 20k iterations for 20 categories and 160k for 160 categories), so that the per-class number of updates is held fixed across C (although we consider the original converged setting more appropriate, as it compares each configuration at its best attainable reconstruction). The same saturation trend persists under this control.
These analysis and figures will be added to the final version.
Response to Reviewer zfFM (part1)
Thank you for the thoughtful and detailed review. We sincerely appreciate your recognition of our diagnosis, clean experimental design, thorough ablations, and consistent gains at well-controlled computational cost. Your comments have been very helpful in improving our paper, and we address each of them below.
Since OpenReview does not support images in comments, the revised and newly added figures (Figs. A1, A2, and A3) have been included in the updated paper PDF. We invite you to refer to it.
W1.
Technical novelty is modest. Input-conditioned weight mixtures (CondConv, DY-Conv) and cross-layer weight templates ( Savarese & Maire; KW) are established; the contribution is choosing image-level routing and sharing for UAD, plus a narrower bottleneck.
Answer:
Thank you for this comment and for recognizing the value of our diagnosis. As you noted, our primary contribution is identifying capacity saturation as a key failure mode of large-scale MUAD. On the technical side, we do not aim to invent an entirely new mechanism; rather, we identify which design choices matter for alleviating capacity saturation and realize them in the simplest form. We have explored more complex alternatives extensively, including various dynamic weights, finer weight partitioning, and contrasting-driven routing, and found that they bring no clear improvement (Tables 4 and A4). This indicates that the current simple design is sufficient for effectively alleviating capacity saturation. In the revision, we will make this positioning explicit. We hope that our diagnosis and preliminary design can inspire future work on capacity saturation in large-scale UAD.
W2.
The "capacity saturation" evidence rests on one baseline, one dataset, one encoder size and normalized errors from single runs. The curves in Fig. 1(b) are fairly flat; it is not shown that the trend is beyond seed noise.
Answer:
Thank you for this valuable suggestion. It prompted us to extend the category-scaling diagnosis to a larger dataset, different baselines, and different encoders (new Fig. A2).
More datasets, baselines, and encoders. We repeat the scaling protocol on ADNet, scaling the training set from 40 to 380 categories while evaluating on a fixed 40-category test set, with three settings: Dinomaly2-S (ViT-S), Dinomaly2-B ( ViT-B), and RD4AD (CNN-based, WideResNet-50).
On ADNet, a more diverse and challenging benchmark spanning five domains and 380 categories, capacity saturation becomes even more pronounced:
The trends on ADNet are large, far exceeding typical seed variation in this framework. Moreover, we are running five-seed category-scaling experiments (which is time-consuming) and will post the results during the rebuttal period if time permits.