Controlling Multiple Errors Simultaneously with a PAC-Bayes Bound

Current PAC-Bayes generalisation bounds are restricted to scalar metrics of performance, such as the loss or error rate. However, one ideally wants more information-rich certificates that control the entire distribution of possible outcomes, such as the distribution of the test loss in regression, or the probabilities of different mis-classifications. We provide the first PAC-Bayes bound capable of providing such rich information by bounding the Kullback-Leibler divergence between the empirical and true probabilities of a set of $M$ error types, which can either be discretized loss values for regression, or the elements of the confusion matrix (or a partition thereof) for classification. We transform our bound into a differentiable training objective. Our bound is especially useful in cases where the severity of different mis-classifications may change over time; existing PAC-Bayes bounds can only bound a particular pre-decided weighting of the error types. In contrast our bound implicitly controls all uncountably many weightings simultaneously.

Paper

References (47)

Scroll for more · 35 remaining

Similar papers

Peer review

Reviewer fvxp7/10 · confidence 3/52024-07-10

Summary

The authors suggest an approach for controlling errors of various types (simultaneously) within a PAC-Bayes framework. They derive a high probability bound on the KL divergence between the empirical and distribution risk vectors. This generalizes earlier work in the binary case by Maurer, 2004 and Begin et al 2016. The authors propose a method for using this bound as a (differentiable) training objective via considering a linear combination of the elements of the risk vector and inverting the resulting binary KL divergence.

Strengths

- The main result seems to be a creative and potentially useful extension of the PAC-Bayesian framework. - Results are generally introduced with some explanation, and given more context after the result, which improves readability.

Weaknesses

- The framework considered would be better motivated by providing a specific example (ideally with an experiment) indicating the usefulness of the method (and illustrating why existing approaches with a union bound are insufficient). - The authors included the NeurIPS checklist from last year (and not from this year). - The proof of theorem 2 is very long and either an outline/proof sketch should be provided at the beginning, or the proof should be broken into lemmas/propositions to improve readability.

Questions

- Is it really not possible to upper bound arbitrary linear combinations of different types of risk with a union bound? I would think that bounding finitely many would place linear constraints on other linear combinations, allowing at least some bounds to be derived. It isn’t immediately obvious to me how sharp the resulting bounds would be, or if indeed this approach leads to non-vacuous bounds, but I’d appreciate the authors commenting on this and adding some discussion on the topic if the answer to my question is that it is possible. - How would $\ell$ be chosen in practice when defining the objective function? It seems the whole purpose of the framework is to provide useful bounds simultaneously for many types of predictive errors. But it also seems the choice of $\ell$ will have a large impact on how well a method trained with proposed approach will perform on tasks that weight certain types of errors heavily.

Rating

7

Confidence

3

Soundness

4

Presentation

3

Contribution

3

Limitations

- The authors acknowledge limited experiments as a limitation of the current work. I agree that this is a limitation. In particular, I wonder whether similar results can be obtained using held-out data, and whether or not the bounds derived this way are competitive with this approach. In the one experiment provided, half of MNIST was used to define the prior, while have was used for optimization of the bound. It seems possible that a similar breakdown of data between a training and validation set (trying directly via cross-entropy loss) might lead to similar model performance and tighter bounds on the risk (as in the cases considered in Foong et al). It isn’t immediately obvious whether it is easy to directly invert the multinomial CDF to derive tight bounds on various datasets as in the binary case using holdout data. But I’m curious to hear the authors’ thoughts on this approach. - The authors point out that the bound is limited to discrete cases. This seems reasonable, and there are many discrete problems where the bound might be useful. I think finding a concrete problem (dataset, with a plausible decision depending on balancing error types) where this bound might be useful would be more convincing to address this limitation than extending the work to continuous problems. Minor points: - Line 147: $S_{m,n}^{>0}$ isn’t the interior of $S_{m,n}$ in a topological sense, assuming $S_{m,n}$ is considered with the discrete topology, which seems the natural choice. Otherwise, I don’t know what is meant by interior. - Line 198: "Different different"

Reviewer 4oQc7/10 · confidence 3/52024-07-11

Summary

The notion of a set of "error-types", introduced by this work, is a user-defined partition of the product space of predictions and responses, generalizing well-known summaries of the erring behavior of predictors. This work presents a PAC-Bayes bound on the divergence between and empirical and true distributions over a set of "error types" for a choice of posterior prediction scheme. The phrasing of the bound in terms of KL divergence naturally generates bounds on arbitrary linear combinations of the true error type probabilities, which hold simultaneously. The authors present a training objective founded on the minimization of the worst-case value of a pre-defined weighting of error types consistent with the empirical error-type profile of a given choice of posterior over predictors.

Strengths

The paper strikes a good balance between practicality and theory. The problem setting is well-motivated, and the generalization of the confusion matrix through error types allows for the use of the bounds in a variety of non-standard settings. The PAC-Bayes bound is novel. While PAC-Bayes has been previously used to control the spectral norm between between empirical and true confusion matrices, the bounding distributions over error types in terms of KL divergence straightforwardly generates bounds on the expected cost (over the true distribution over error types) which hold simultaneously, something of notable practical use. The paper is generally well-written.

Weaknesses

One of the main strengths of this work the simultaneous control of arbitrary linear combinations of the error type probabilities. However, the ability of the previous literature to generate some semblance of these guarantees is not particularly well-exposited. Having not previously read any of this line of work, I think an explicit example of how one can construct a bound on a single linear combination might do the paper some good (and might motivate this work past the simultaneous control of all linear combinations, given that the task seems somewhat tricky given a spectral norm bound).

Questions

Is there a quick example of to control a single linear combination of confusion matrix probabilities using a previous bound? In line 331, what is an "empirically tighter" bound?

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes

Authorsrebuttal2024-08-13

We are pleased to read that you found the experiment clarifying! Thank you again for taking the time to write a careful review and for your insightful comments, which we believe have improved the work.

Reviewer cfTA6/10 · confidence 3/52024-07-12

Summary

This paper introduces a novel PAC-Bayes bounds that extend classical kl-based bounds to vector-valued losses that can control several error-types simultaneously. The bound is converted into a differentiable minimization objective and details for the practical implementation of the bound are provided in the Appendix.

Strengths

1- The paper is clearly written and theoretically solid. 2- The contributions have a potential impact in scenarios where simultaneous control over multiple error types is needed.

Weaknesses

1- The proof techniques and the main contribution (Proposition 4) are straightforward extensions of scalar kl-bounds to the multidimensional case. I fear this contribution alone might not be significant enough. 2- The experiment in Section 7 is too simple, further empirical evaluation of the bounds (e.g. simultaneous minimization of Type I and Type II errors in more realistic scenarios) would be very beneficial for the impact of the paper. 3- Wu, Y. S., & Seldin, Y. (2022). Split-kl and PAC-Bayes-split-kl inequalities for ternary random variables. Advances in Neural Information Processing Systems, 35, 11369-11381. is significant related work not discussed in the paper. In general, I don't think this is a bad paper, but the theoretical contributions are not very original and I feel that the paper needs more experimental backup to be a well-rounded contribution.

Questions

See weaknesses.

Rating

6

Confidence

3

Soundness

4

Presentation

3

Contribution

2

Limitations

The limitations are properly discussed in Section 6.

Reviewer R3U15/10 · confidence 4/52024-07-12

Summary

In the standard statistical learning theory control of generalization error typically means studying deviations of the empirical risk from the risk (for instance, with high probability over the sample). In the language of statistics such control doesn't make a difference between type I & II errors. This paper goes beyond as it looks at a multi-objective risk, i.e. 'deviation' of the vector or M empirical risks from the corresponding risks. In the context of the paper, the vector of empirical risks lives on a probability simples and the paper shows how to control a KL divergence between such a vector and a risk probability vector. Then, the paper goes on and studies a PAC-Bayes formulation of this problem (Theorem 1). In section 4 the paper proposes how to construct a differentiable objective for this problem.

Strengths

This is an important problem, especially in the context of multiobjective errors which are commonplace in practice. To some extent these results also extend results of Maurer (2004) (PAC-Bayes bounds on little-kl) to simpleces (while original results were shown for Bernoulli case).

Weaknesses

The price for having such a multiobjective setting is M log(M), whereas the standard union bound would lead to a log(M) cost --- note that there is a (significant) gap. However, it should be noted that this would give a bound on a different objective (not kl() between points on the simplex). I think that the paper should argue why such kl() is interesting or at least design analytical instances and/or some experiments which would demonstrate that there exist instances where kl() leads to a better/tigther results (perhaps a performance on a downstream task?).

Questions

Not a question really, but I'd be interested to hear authors' thoughts on the weakness against a union bound. E.g. how would one approach this comparison problem / experiment design.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Seems to be adequate.

Authorsrebuttal2024-08-13

We are glad you found the evaluations interesting and are pleased that you are in favour of accepting the paper! We will indeed expand the evaluations in the final version if it is accepted. Thank you for your detailed review and suggestions, which we believe have improved the paper!

Reviewer cfTA2024-08-11

Thank you for your clarifying comments, with my doubts regarding the theoretical contributions solved, having read the rest of the rebuttals and with more realistic experiments, I am happy to increase my rating to 6. Including these extra experiments in the camera-ready version will make for a solid contribution.

Authorsrebuttal2024-08-13

We are glad that you found the additional experiment useful! Thank you for raising your score in light of this, and for your suggestions which we believe have improved the work.

Reviewer fvxp2024-08-12

The authors have addressed the primary concerns raised in my review. I have increased my score by a point, as I think the paper should be accepted.

Authorsrebuttal2024-08-13

We are glad we have addressed the important points you raised, and to read that you believe the paper should be accepted. Thank you again for taking the time to write such a thorough review, and for your suggestions and insights which we believe have allowed us to improve the paper!

Reviewer R3U12024-08-13

Thanks for your explanations, and empirical evaluations which are indeed interesting. I guess, for the final revision it would be interesting to see a more complete empirical suite (even if it's synthetic) to see where the proposed approach "breaks" and the union bound becomes better. Otherwise it is hard to see a full picture. Overall, I think this is interesting and proofs have some new ideas, so I'm in favor of accepting this paper.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC