- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes the Credal Information Bottleneck (CIB) for representation learning under uncertainty over data-generating environments and model parameters. It defines environment-wise rate and distortion objectives and establishes a distortion floor for pooled-efficient IB representations. A variational multi-head implementation, VCIB, estimates aleatoric, environment-related, and model-related uncertainty and uses unlabeled test batches to reweight environments. Experiments on three colored classification benchmarks report reduced credal distortion and improved worst-case accuracy after adaptation.
Critical strengths
- Preserving environment-specific predictive information through compression is a meaningful problem. The stated separation result clearly identifies when a representation sufficient for pooled prediction may lose information needed to recover environment-specific labeling laws.
- The evaluation includes ten seeds, statistical comparisons, and ablations separating model-level credal sets from environment-specific modeling. The paper also reports limitations of its model-uncertainty estimates and selective-prediction performance.
Critical weaknesses
- The deployment gains are not fully isolated from the benefits of environment-specific prediction and batch adaptation. The rate-free VCIB ablation provides relevant evidence, but a standard multi-head classifier using the same environmental information and batch-weighting procedure would help determine whether the gains require the proposed credal bottleneck.
- The compression comparison matches per-code rates, while predictions combine multiple encoder heads. Appendix A.5 reports joint-rate bounds, but the methods are not compared at matched total rates. Such comparisons, including the rate-free variant, are needed to assess the compression advantage of the complete predictive representation.
- The uncertainty experiments mainly assess detection proxies rather than the accuracy of the claimed decomposition. Appendix D provides analytical uncertainty components and qualitative comparisons, but quantitative agreement with these components is not established. Detecting underrepresented inputs or OOD inputs does not directly validate model or aleatoric uncertainty. The weak model-uncertainty results further limit the general disentanglement claim.
- All three benchmarks use the same two-environment binary shape–color construction. Changing the underlying images provides some variation, but does not test richer uncertainty structures or scaling with the number of environments. The practical scope of the method remains unclear beyond this construction.
- Equation 16 assigns one environment-weight vector to an entire test batch. Deployment is evaluated on homogeneous batches of 100 inputs from known environments, without establishing sensitivity to batch size or mixed-environment batches. These conditions matter because inaccurate environment inference may undermine the reported adaptation advantage.
- Appendix A.7 describes the reported distortion as expected excess log-loss, whereas the theoretical distortion measures information lost through the representation. The paper should clarify whether the reported quantity includes decoder approximation error and explain how it is computed consistently across methods. Without this distinction, the results do not clearly isolate representation-level information preservation.
Minor:
- Statements about pooled representations should consistently retain the pooled-efficiency qualification required by Theorem 2. Environment-specific decoders trained on frozen pooled features would provide useful empirical context.
- Proposition 15 distinguishes control over the generating environments from control over their mixtures. The implications of this distinction for deployment under mixed or previously unseen environments deserve clearer discussion.
- Appendix A.6 replaces the learned mixture priors with moment-matched Gaussian priors for environment inference. Report how this choice affects environment identification and deployment accuracy.
- Appendix A.2 uses moving-average environment weights and tempered weight-KL terms. Clarify how these implementation choices relate to the stated variational objective and its guarantees.
- Report parameter counts and computational costs by method, including the inference cost of multiple hypotheses and batch adaptation.
- Clarify the relationship to Group DRO [1], adaptive risk minimization [2], and prior uncertainty-disentanglement benchmarking [3], particularly the distinctions in objectives, adaptation procedures, and validation.
References:
[1] Sagawa, Shiori, Pang Wei Koh, Tatsunori B. Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. In ICLR, 2020.
[2] Zhang, Marvin, Henrik Marklund, Nikita Dhawan, Abhishek Gupta, Sergey Levine, and Chelsea Finn. Adaptive risk minimization: Learning to adapt to domain shift. In NeurIPS, 2021.
[3] Mucsányi, Bálint, Michael Kirchhof, and Seong Joon Oh. Benchmarking uncertainty disentanglement: Specialized uncertainties for specialized tasks. In NeurIPS, 2024.
Questions for authors
- Does the deployment advantage remain against standard environment-specific classifiers using the same unlabeled batches and weighting procedure?
- Do the improvements persist when methods are compared at matched joint representation rates?
- How accurately do the estimated uncertainty components match the analytical quantities in Appendix D?
- How does adaptation behave with smaller or mixed-environment batches, and how sensitive is it to the post-training choice of code priors?
- Does the reported distortion include decoder approximation error? How is it computed consistently across methods?
6 comments
Response to PBer (2/6)
Q2/W2. Matched joint rates
Comment: Do the improvements persist when methods are compared at matched joint representation rates? [...] Such comparisons, including the rate-free variant, are needed to assess the compression advantage of the complete predictive representation.
Response: Thank you for the suggestion. VCIB and Ens-VIB transmit five codes, one per encoder hypothesis. is an upper bound on the information that the five codes together carry about the input, computed with the pairwise-distance bound of [1] (column of Table 1; Appendix A.5). For each dataset, we take VCIB at the of its grid whose is closest to that of Ens-VIB, the pooled method with the largest joint rate.
Table R12. Comparison at matched joint rate. : upper bound on the rate of all codes together (bits). VCIB(): rate-free reference (deterministic code, no finite rate), values of Table 1; it is not part of the significance tests. Mean over seeds. Bold: VCIB, significantly better than every baseline on both metrics (seed-paired Wilcoxon signed-rank test, Holm-Bonferroni over the three baselines, in every case).
| Data | Method | D | WC-Acc_int | |
|---|---|---|---|---|
| C-MNIST | VIB | 1.34 | 0.248 | 0.810 |
| Ens-VIB | 2.44 | 0.246 | 0.812 | |
| VCIB_mod | 2.02 | 0.243 | 0.811 | |
| VCIB () | 2.61 | 0.061 | 0.872 | |
| VCIB() | 0.043 | 0.883 | ||
| C-KMNIST | VIB | 1.75 | 0.315 | 0.774 |
| Ens-VIB | 4.08 | 0.317 | 0.778 | |
| VCIB_mod | 3.01 | 0.330 | 0.768 | |
| VCIB () | 3.56 | 0.177 | 0.821 | |
| VCIB() | 0.157 | 0.830 | ||
| C-Fashion | VIB | 1.51 | 0.314 | 0.766 |
| Ens-VIB | 2.77 | 0.316 | 0.769 | |
| VCIB_mod | 2.01 | 0.310 | 0.770 | |
| VCIB () | 2.75 | 0.174 | 0.836 | |
| VCIB() | 0.152 | 0.836 |
The results show that, at matched joint rate, VCIB keeps a lower distortion and a points higher accuracy after deployment than every pooled method. On C-KMNIST and C-Fashion, it does so at a lower or equal joint rate than Ens-VIB. The distortion of the pooled methods does not move with their joint rate (Ens-VIB on C-MNIST: at bits), consistently with Thm. 2(i), to the extent that these encoders are approximately pooled-efficient.
In the final version, we will add Table R12 to Appendix B, and the following sentence to the paragraph "Overcoming the pooling floor" of Section 5: "When all methods are compared at the same joint rate Rjoint, VCIB has a 1.8-4\times lower distortion and a points higher post-deployment accuracy than every pooled method (Appendix B). Raising the joint rate does not lower the distortion of the pooled methods, as predicted by Thm. 2."
[1] A. Kolchinsky and B. D. Tracey, "Estimating Mixture Entropy with Pairwise Distances", Entropy 2017.
Q3/W3. Agreement with Appendix D
Comment: How accurately do the estimated uncertainty components match the analytical quantities in Appendix D?
Response: Thank you for raising this point. To evaluate the uncertainty estimation quality of VCIB against the true uncertainty values, we proceed as follows:
In the colored benchmarks (outlined in Appendix D) labels and colors are sampled with known probabilities. The label equals the class of the shape with probability , while the color equals the label with probability which depends only on the environment and the class .
Hence, the labeling law of each environment is given in closed form (Eq. 67) as a function of the class and color of the image. The true credal set for test image is obtained as and its exact aleatoric uncertainty (Def. 4), , the lowest Shannon entropy (in bits) in the set. Methods without our decomposition are scored with their own AU estimate: the predictive entropy (VIB), the mean entropy of the members (Ens-VIB), and the lowest entropy in the credal set (CBDL, VCIB_mod).
Response to PBer (3/6)
Q3/W3. Agreement with Appendix D (continued)
Appendix D defines three input groups: classes 0-4, on which the two environments give the same image different labels (50% of the test set, bits); classes 5-9 whose color agrees with the label (34%, ); and classes 5-9 whose color disagrees with it, where the label itself is random (15%, ). Only the third group has a high ; as such, a good AU estimate must rank its images above the images from a low-AU environment. We measure this with the AUROC of each method's AU, with the noisy images as positives and one other group as negatives. Additionally, we compare against a fourth group thinned, from the dataset variants of Table 2.
Table R1. We show the AUROC of each method's AU for separating the label-noise images from each low-AU comparison group. Mean over 10 seeds. ▼ marks a baseline significantly worse than VCIB (paired Wilcoxon signed-rank test, Holm-Bonferroni corrected, ). Bold: best per row.
| Data | Noisy vs. | VIB | Ens-VIB | IB-IRM | CBDL | VCIB_mod | VCIB |
|---|---|---|---|---|---|---|---|
| C-MNIST | color agrees | 0.969▼ | 0.976▼ | 0.693▼ | 0.984▼ | 0.975▼ | 0.997 |
| env. disagree | 0.503▼ | 0.497▼ | 0.525▼ | 0.607▼ | 0.513▼ | 0.990 | |
| thinned () | 0.540▼ | 0.510▼ | 0.498▼ | 0.628▼ | 0.544▼ | 0.966 | |
| C-KMNIST | color agrees | 0.882▼ | 0.924▼ | 0.592▼ | 0.946 | 0.913▼ | 0.960 |
| env. disagree | 0.476▼ | 0.455▼ | 0.524▼ | 0.570▼ | 0.479▼ | 0.909 | |
| thinned () | 0.532▼ | 0.539▼ | 0.563▼ | 0.663▼ | 0.573▼ | 0.829 | |
| C-Fashion | color agrees | 0.877▼ | 0.882▼ | 0.832▼ | 0.884▼ | 0.886▼ | 0.947 |
| env. disagree | 0.425▼ | 0.405▼ | 0.520▼ | 0.395▼ | 0.439▼ | 0.860 | |
| thinned () | 0.676▼ | 0.677▼ | 0.665▼ | 0.675▼ | 0.688▼ | 0.891 |
Averaged over the three datasets, the AU estimated by VCIB separates label noise from environment disagreement with an AUROC of , against for the best baseline (CBDL). Likewise, VCIB separates label noise from missing training data with an AUROC of , against for CBDL. Thus, VCIB is the only method among baselines that is fundamentally capable of isolating label noise from the other two sources of uncertainty. Pooled models are incapable of separating environment disagreement from label noise (see Cor. 13). In stark contrast, VCIB attributes this disagreement to EU_input instead with an AUROC of (Table 2).
Moreover, we report the mean absolute error of AU against over the test set, on the full data and on the variant, whose test set and labeling laws, and hence , are unchanged.
Table R2. Mean absolute error of AU against (bits) over the test set, mean ± std over 10 seeds. ▼ marks a baseline significantly worse than VCIB (same test and correction as Table R1; in every case, as VCIB has the lower error on all 10 seeds). Bold: best per row.
| Data | VIB | Ens-VIB | IB-IRM | CBDL | VCIB_mod | VCIB |
|---|---|---|---|---|---|---|
| C-MNIST | 0.367 ± 0.010▼ | 0.362 ± 0.007▼ | 0.682 ± 0.005▼ | 0.313 ± 0.004▼ | 0.317 ± 0.007▼ | 0.078 ± 0.007 |
| C-MNIST 2% | 0.381 ± 0.013▼ | 0.373 ± 0.008▼ | 0.687 ± 0.002▼ | 0.329 ± 0.008▼ | 0.331 ± 0.008▼ | 0.091 ± 0.007 |
| C-KMNIST | 0.402 ± 0.012▼ | 0.392 ± 0.009▼ | 0.687 ± 0.002▼ | 0.325 ± 0.005▼ | 0.352 ± 0.008▼ | 0.133 ± 0.015 |
| C-KMNIST 2% | 0.401 ± 0.015▼ | 0.397 ± 0.010▼ | 0.686 ± 0.003▼ | 0.320 ± 0.007▼ | 0.346 ± 0.010▼ | 0.131 ± 0.008 |
| C-Fashion | 0.406 ± 0.010▼ | 0.406 ± 0.006▼ | 0.660 ± 0.017▼ | 0.351 ± 0.005▼ | 0.363 ± 0.006▼ | 0.151 ± 0.012 |
| C-Fashion 2% | 0.399 ± 0.011▼ | 0.394 ± 0.011▼ | 0.673 ± 0.016▼ | 0.343 ± 0.008▼ | 0.355 ± 0.005▼ | 0.145 ± 0.005 |
Averaged over the three datasets, VCIB's AU estimation error is 2.7× lower (0.121 bits against 0.330) than the best baseline (CBDL), and 3.5× lower than the average baseline. Removing 98% of the training images of classes 2 and 7 leaves the error unchanged (0.122 bits); as such, VCIB does not mistake missing training data for label noise, consistently with Table R1.
In the final version, we will add a paragraph on the ground truth of AU to Section 5, after the discussion of Table 2, and Tables R1 and R2 to Appendix B.
Response to PBer (1/6)
Q1/W1. Multi-head classifiers
Comment: Does the deployment advantage remain against standard environment-specific classifiers using the same unlabeled batches and weighting procedure?
Response: We thank you for suggesting this baseline, which we added: a standard deterministic classifier with the same encoder backbone and one output head per environment, trained on the average of the environments' losses, with no Bayesian weights, no stochastic code and no rate term. At deployment, it uses the same unlabeled batches and Eq. 16 as VCIB, with one Gaussian fitted to its features per environment. The comparative results are in Table R11.
Table R11. Multi-head classifier against VCIB at 2 bits and at the next point of its β grid (β=24). Rate: R as in Table 1 (bits); ∞: a deterministic code has no finite rate. AUROC_conflict: EUinput on classes 0−4 against 5−9 (Table 2). AUROC_OOD,model: EUmodel on the out-of-support inputs of Table 2 against the test inputs. Best pooled: best value among VIB, Ens-VIB, CBDL and VCIB_mod, which use a single epistemic score for both AUROC columns; its D, WC-Acc_int and AUROC_conflict, and those of VCIB (2 bits), are the values of Tabs. 1 and 2. Mean over 10 seeds. ▲/▼: significantly better/worse than the multi-head classifier (seed-paired Wilcoxon signed-rank test, Holm-Bonferroni over the VCIB variants of each dataset and metric, p<0.05).
The multi-head classifier performs like VCIB(β→∞): both keep one readout per environment on an uncompressed code. As such, the gain over pooled methods comes from keeping the environments separate, which is the core idea of the CIB. The rate constraint imposed on VCIB preserves most of these gains. At 2 bits, VCIB is within 0.4−1.1 points of the multi-head accuracy; at β=24 (rate 2.3−2.6 bits), its accuracy is not significantly different on any dataset. Furthermore, its credal distortion is not significantly higher on any dataset, and is significantly lower on C-MNIST. VCIB thus matches the multi-head classifier's performance at a rate of 2.3−2.6 bits, while being capable of disentangling uncertainty across model parameters and data generating distributions.
In the final version, we will add the multi-head classifier to Table 1 and Table R11 to Appendix B.