- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes Woodbury Influence (WIF), a post-hoc score for evaluating how much an unlabeled query sample would perturb a trained model’s parameters if added to the training set. WIF combines hard pseudo-labeling with a damped, layer-wise empirical-Fisher approximation, using the Woodbury identity to compute the inverse-Hessian action in closed form. The authors validate the score against retraining on a small MNIST setup and against LiSSA on ResNet-18, then evaluate it on OOD detection, open-set recognition, and class-order selection for continual learning. The underlying question—whether influence can be useful for assessing unseen samples—is interesting. However, the paper’s empirical evidence and efficiency claims do not yet establish that WIF is a reliable or broadly practical OOD method. The reported gains are uneven, the most striking OSR result is not accompanied by a corresponding AUROC gain, and important details about the estimator’s computational cost and experimental protocols are missing.
Critical strengths
-
The sample-addition perspective is practically relevant and less explored than training-sample influence.
-
The Woodbury derivation is clear under the paper’s layer-wise empirical-Fisher approximation.
-
The paper includes useful ablations of pseudo-labeling and curvature.
Critical weaknesses
-
Direct validation against retraining is limited to a small MNIST model. The ResNet experiment compares against LiSSA, not retraining.
-
The efficiency analysis does not clearly include the cost of constructing and inverting the Gram matrix or report end-to-end memory and latency.
-
OOD results are mixed: WIF improves FPR95, but does not consistently lead in AUROC; some gains are small.
-
OSR shows a large FPR95 gain but nearly identical AUROC to the best baseline. The explanation for this gap needs stronger evidence.
-
The continual-learning evaluation is small and underspecified, making it difficult to assess whether the ordering gains generalize.
Questions for authors
-
In the OOD experiments, is WIF computed using only the final-layer block as in Eq. 10? If so, which experiments use the full layer-wise estimator in Eq. 9?
-
Does the reported timing include construction of , formation and inversion of the Gram matrix, and amortization across queries? What are the concrete values of , , and query batch size?
-
How sensitive are OOD and OSR results to the choice of damping , and was it selected using any OOD validation data?
-
Can the authors provide paired uncertainty estimates for the AUROC and FPR95 differences, particularly for OSR and the smaller OOD gains?
-
For continual learning, what data are available when the class order is selected, and how many class splits or permutations are used?
4 comments
W3 & W4
Your understanding is correct: the AUROC is relatively similar across methods, but the gain in FPR95 increases dramatically. We explain this in three points.
1) OSR, a generally harder task
OSR can be regarded as a near-OOD task, which is the more challenging setting in OOD detection. On far-OOD, a comparatively easy setting, existing methods already perform well, achieving results on par with or better than WIF. On harder settings such as near-OOD and OSR, however, existing methods struggle considerably. We believe this is how the AUROC results should be interpreted.
2) Favorable distribution shape
Sec 5.5 analyzes why WIF achieves strong performance on FPR95. FPR95 sets the threshold at the 95% point of the ID data and measures how much the OOD distribution overlaps with the ID distribution beyond it. The favorable distribution shape for minimizing this overlap concentrates most of the data at the head and becomes sparser toward the tail (the 95% point), decaying gradually. Among the compared methods, WIF comes closest to this ideal shape.
3) Amplified effect of PL depending on model performance
Table. 2 shows that WIF exhibits a notably larger FPR95 gap relative to the baselines on TinyImageNet. In all other OOD detection and OSR benchmarks, we use models that achieve about 95% accuracy on the ID test data. On TinyImageNet, however, ResNet18 cannot reach this level of accuracy, so we use a model with accuracy in the 80% range. This model has a considerably high error rate on the ID test data, which makes it more susceptible to the effect of pseudo-labeling. We therefore conjecture that the large FPR95 gap on this dataset stems from the amplified influence of pseudo-labeling in Sec 5.4
We additionally conducted experiments on the classical benchmarks CIFAR-10/CIFAR-100 and SVHN. OSR results Table has been revised as follows. The results remain consistent with the original trends.
Table 2. Open-Set Recognition results on CIFAR10/CIFAR100, SVHN, and Tiny ImageNet with ResNet-18. Our method outperforms all the baselines considered. Higher AUROC and lower FPR95 indicate better performance. All results are reported as mean ± standard deviation across 5 runs with different known/unknown class splits. Best results in bold.
| Method space | Method | CIFAR10/100 AUROC ↑ | CIFAR10/100 FPR95 ↓ | SVHN AUROC ↑ | SVHN FPR95 ↓ | Tiny ImageNet AUROC ↑ | Tiny ImageNet FPR95 ↓ |
|---|---|---|---|---|---|---|---|
| No Gradient | MSP | 0.770 ± 0.012 | 0.702 ± 0.0071 | 0.917 ± 0.008 | 0.517 ± 0.021 | 0.776 ± 0.011 | 0.826 ± 0.081 |
| No Gradient | Energy | 0.782 ± 0.017 | 0.658 ± 0.068 | 0.925 ± 0.005 | 0.472 ± 0.017 | 0.779 ± 0.020 | 0.834 ± 0.071 |
| No Gradient | ODIN | 0.755 ± 0.021 | 0.693 ± 0.072 | 0.921 ± 0.004 | 0.455 ± 0.015 | 0.782 ± 0.013 | 0.815 ± 0.070 |
| No Gradient | ASH | 0.789 ± 0.013 | 0.650 ± 0.069 | 0.927 ± 0.007 | 0.410 ± 0.021 | 0.778 ± 0.010 | 0.853 ± 0.078 |
| No Gradient | SCALE | 0.859 ± 0.011 | 0.650 ± 0.059 | 0.920 ± 0.008 | 0.423 ± 0.023 | 0.779 ± 0.012 | 0.841 ± 0.061 |
| Gradient | GradNorm | 0.763 ± 0.018 | 0.679 ± 0.061 | 0.923 ± 0.007 | 0.426 ± 0.019 | 0.773 ± 0.016 | 0.853 ± 0.082 |
| Gradient + Hessian | WIF (ours) | 0.875 ± 0.021 | 0.619 ± 0.053 | 0.935 ± 0.004 | 0.399 ± 0.016 | 0.784 ± 0.012 | 0.570 ± 0.030 |
W5
We would first like to clarify the scope of our contributions. We propose WIF, which enables label-free influence estimation in the sample-incremental regime, and we justify it theoretically and empirically. We then instantiate it for OSR, OOD detection, and order sensitivity in continual learning (CL), aiming to verify its potential and extend its applicability rather than to chase state-of-the-art numbers. Finally, we identify two mechanisms behind its separation of familiar and unfamiliar samples.
We did not study continual learning (CL) in its entirety. Instead, we restricted our scope to presenting a concrete use case, order sensitivity in CL, together with a method for addressing it. Within this scope, we used EWC and GEM, which are widely used representatives of regularization-based and replay-based approaches, respectively. However, we recognize that these two methods alone may not be sufficient. We therefore tried to add one more representative method for each category, LwF (regularization-based) and ER (replay-based), and updated the table accordingly.
We agree that claiming general effectiveness across CL would be an overgeneralization. For this reason, we deliberately narrowed our experiments to this specific problem and report results only within that scope. Broader applicability and robust methods that can compete with state-of-the-art (SOTA) approaches were mentioned only as possible directions for future work in the Appendix limitation section.
Table 4. Class-incremental CL on Tiny ImageNet with ResNet-18. The WIF-based Ben-first ordering achieves the best performance. Average Accuracy (AA, ↑) is better when higher, and Forgetting Measure (FM, ↓) is better when lower. Results are averaged over 5 seeds.
| Order | EWC AA ↑ | EWC FM ↓ | GEM AA ↑ | GEM FM ↓ | LwF AA ↑ | LwF FM ↓ | ER AA ↑ | ER FM ↓ |
|---|---|---|---|---|---|---|---|---|
| Random | 0.40 ± 1.19 | 0.60 ± 1.31 | 0.54 ± 0.87 | 0.46 ± 0.51 | 0.37 ± 0.30 | 0.63 ± 0.33 | 0.66 ± 0.14 | 0.33 ± 0.41 |
| Des-first w/ WIF | 0.37 ± 1.15 | 0.58 ± 1.37 | 0.62 ± 0.77 | 0.44 ± 0.45 | 0.33 ± 0.31 | 0.67 ± 0.44 | 0.65 ± 0.13 | 0.32 ± 0.47 |
| Ben-first w/ Logit | 0.41 ± 1.58 | 0.56 ± 1.36 | 0.72 ± 0.68 | 0.32 ± 0.46 | 0.37 ± 0.27 | 0.63 ± 0.38 | 0.75 ± 0.19 | 0.28 ± 0.28 |
| Ben-first w/ WIF (ours) | 0.42 ± 1.21 | 0.54 ± 1.66 | 0.74 ± 0.62 | 0.25 ± 0.47 | 0.40 ± 0.31 | 0.53 ± 0.38 | 0.78 ± 0.14 | 0.22 ± 0.40 |
Q1
This is already described in Section 4.2. Motivated by the experimental findings of GradNorm [3], we design WIF. Using only the last layer achieves better OOD (out-of-distribution) detection and OSR performance and even reduces computation time.
[3] Rui Huang, Andrew Geng, and Yixuan Li. On the importance of gradients for detecting distributional shifts in the wild. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan (eds.), Advances in Neural Information Processing Systems, 2021.
Q2
Addressed in W2.
Q3
The effect of λ is large enough to alter the performance ranking. Each dataset requires its own optimal λ to achieve the best performance. We heuristically varied λ from 0.1 to 0.00001 until the best performance was reached. Since this procedure is highly heuristic, we discuss it in the Limitations section.
Q4
Addressed in W3 & W4.
Q5
This is presented in Sec 4.2 (formally) and Sec. 5.7 (detailed). We evaluate on Tiny ImageNet with ResNet-18. The model is trained on 5 base classes and then incrementally learns 5 more, one class per step.
We thank the reviewer Dye2 for the careful reading. We look forward to a constructive discussion.
W1
We believe this point refers to Figure 1. We validate the approximation quality of WIF in the sample-incremental regime through two complementary experiments: one direct (retraining with small dataset and architecture) and one indirect (without retraining with larger dataset and architecture).
The left panel of Fig. 1 shows the direct validation. Following the protocol of [1], we validate against exact influence obtained via retraining, but only on a small model and dataset, since retraining once per query sample is prohibitively expensive for larger datasets and architectures. For example, if a single training run takes 10 minutes, evaluating 10,000 data points would require 100,000 minutes. This is impractical, so we adopt the same evaluation protocol as the original IF [1].
The right panel of Fig. 1 shows the indirect validation. The direct approach becomes infeasible for larger datasets and architectures because retraining is prohibitively expensive. To assess approximation quality at these scales, we instead measure the correlation between WIF and LiSSA. LiSSA is substantially cheaper than actual retraining and has been shown to approximate it closely [1].
[1] Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. In Proceedings of the 34th International Conference on Machine Learning (ICML), pp. 1885–1894, 2017.
W2
We have updated the complexity analysis in Table 3. Table 3 reports the complexities of Gram matrix construction and inverse matrix computation. Complexities retain only the dominant term, and lower-order terms are omitted, since all terms of end-to-end would be too long. For OOD detection, n is the number of training samples in the dataset (100,000). d is the dimension of the last layer (512 for ResNet18 and 1024 for ViT). Q is the number of incoming samples to be scored (Q=10,000). Batch size appears in this analysis because of the reference gradient. The closed form is constructed from all n gradients (full-batch).
Table 3. Time and Space Complexity of Influence Function methods. WIF always has complexity equal to or better than the classical methods Exact IF and LiSSA. Compared with DataInf, WIF is at a disadvantage in the Gram matrix and inverse matrix construction and in memory complexity. However, WIF gains in per-query computation, so its overall speed is faster. We denote the number of reference samples by n and the parameter dimension by d. Q denotes the number of incoming query samples. Complexities retain only the dominant term. Lower-order terms are omitted.