Should We Learn Most Likely Functions or Parameters?

Standard regularized training procedures correspond to maximizing a posterior distribution over parameters, known as maximum a posteriori (MAP) estimation. However, model parameters are of interest only insomuch as they combine with the functional form of a model to provide a function that can make good predictions. Moreover, the most likely parameters under the parameter posterior do not generally correspond to the most likely function induced by the parameter posterior. In fact, we can re-parametrize a model such that any setting of parameters can maximize the parameter posterior. As an alternative, we investigate the benefits and drawbacks of directly estimating the most likely function implied by the model and the data. We show that this procedure leads to pathological solutions when using neural networks and prove conditions under which the procedure is well-behaved, as well as a scalable approximation. Under these conditions, we find that function-space MAP estimation can lead to flatter minima, better generalization, and improved robustness to overfitting.

Paper

Similar papers

Peer review

Reviewer wSt37/10 · confidence 2/52023-06-22

Summary

This paper investigates an important yet neglected question in machine learning: should we train the model in the parameter space or function space? The authors show the benefits and shortage of function space MAP estimation which could provide a guide for practitioners. Last but not least, the authors propose a scalable approximation for function space MAP estimation for deep learning.

Strengths

- Thorough analysis of the pros and cons of parameter space MAP estimation and function space MAP estimation - Propose a scalable approximation for large neural networks which extends the application area of function space MAP estimation - Well written and easy to follow

Weaknesses

- I wonder how universal are the pros and cons demonstrated in the paper, considering they're being shown in a relatively simple toy data set - I understand the author is not claiming L-MAP works well in all cases, but I am still interested to see the performance of L-MAP on tasks other than classification

Questions

See weakness above

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

See weakness above

Reviewer Zoj66/10 · confidence 4/52023-06-30

Summary

This work investigates the task of performing inference over model functions rather than model parameters. In particular, the authors propose an objective function *function-space MAP (FS-MAP)*, which intuitively is the usual MAP objective where the prior term is taken over functions rather than parameters. Defining this function-space prior is fundamentally challenging, and to do so, the authors generalize the techniques of Wolpert (1993) to consider functions observed on (possibly) infinite sets. The authors perform an in-depth empirical evaluation of this proposed objective on several synthetic and real-world benchmarks. In addition, a scalable approximation to the objective is proposed in Section 4.3.

Strengths

- The paper is overall very well-written and clear in its aims and conclusions. - The key ideas of the paper are likely of interest to the community and the findings have the potential to be impactful for future work on function-space inference. - I found the discussion relating the function-space geometry to the sampling distribution in Section 3.2 fascinating, and this was a novel connection that I had not considered before. - The experimental evaluation is generally sound and the conclusions drawn are supported by the evidence provided throughout the paper. - The proposed objective is a non-trivial extension of previous work in this area, and the scalable approximation in Section 4.3 is similarly non-trivial.

Weaknesses

- There are several places in the paper lacking rigor and where I believe there to be undressed latent fundamental issues (see the questions section). As noted on Line 96, a fundamental difficulty in function-space inference is defining an appropriate notion of a prior density over infinite dimensional spaces. The methods of this work are obtained via discretization of the functions followed by a heuristic passage to the limit, but it is unclear how this relates to the underlying function-space distributions, which are the true objects of interest. - The empirical results on real-world data (Table 1) do not show benefits of using the FS-MAP objective over the standard PS-MAP objective. - Note that I do not view this necessarily as a weakness of the work itself but rather a potential limitation of the methodology itself, and an honest evaluation of the proposed methodology is valuable. - The paper needs to be better contextualized within the literature, e.g. there is no related work section. See e.g. [1-4] for some potentially relevant works on function-space inference which are potentially of interest. [1] Function-Space Regularization in Neural Networks: A Probabilistic Perspective, Rudner et al., ICML 2023 [2] Understanding Variational Inference in Function-Space, Burt et al., AABI 2020 [3] Generalized Variational Inference in Function Spaces: Gaussian Measures meet Bayesian Deep Learning, Wild et al., NeurIPS 2022 [4] Tractable Function-Space Variational Inference in Bayesian Neural Networks, Rudner et al., NeurIPS 2022

Questions

- Under what conditions is Equation (10) well-defined? It was not clear that the integral in this expression is finite without e.g. further assumptions on the model class. (It is straightforward to give examples where the integral is infinite) - Can the infinite-observation objective (Equation (11)) be justified as an MAP estimate? Equation (8) follows from the change-of-variables formula, but it was not clear to me if (or how) Equation 11 was being justified in the same manner, or merely as a heuristic.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

The authors appropriately discuss the limitations of their work within the submission.

Authorsrebuttal2023-08-21

Thank you again for supporting acceptance of our submission! Your feedback and suggestions were very valuable and have helped us further improve our submission, and we will include the additional results and clarifications from our rebuttal in the revised manuscript. If you agree that the suggested clarifications and new results improved our submission, we would be grateful if you would consider raising your score to reflect these additions. Thank you for your time and effort!

Reviewer PZwV6/10 · confidence 2/52023-07-07

Summary

This paper analyzes the question: should we find the parameter that maximizes $p(\theta|D)$ or $p(f_\theta|D)$, given a data distribution $D$. The former is the classic MAP, or PS-MAP ("parameter space") in this work, and the latter is called FS-MAP ("function space"). The paper shows that neither is universally more desirable than the other, and describes pros and cons of both methods: - FS-MAP has the advantage of directly optimizing the quantity we care about (i.e. function value), is invariant to reparameterization, and can be more robust to label noise. - On the other hand, PS-MAP tends to be more stable, and while it is not invariant to reparameterization, it's unclear how much this would be a disadvantage in practice. Moreover, PS-MAP can be closer to the Bayesian model average than FS-MAP. The findings are supported by both theory and experiments.

Strengths

- The paper has extensive comparisons on the limitations and applicability of both methods. - The paper addresses practical considerations: - The problem of having a (near) singular Jacobian is addressed by adding perturbations. - For scalability, the problem of requiring two forward steps is addressed by zeroth-order approximation, which requires 1 forward pass only.

Weaknesses

While I appreciate the discussions, I'm not sure the evidences are clear enough. - Fig 2 shows that there's no difference between FS-MAP and PS-MAP when the number of samples is around 10k, a sample size that most practical settings can afford. - In Table 1, the results from both methods are very close, often within the margin of standard deviation. Overall, it seems that both methods compare similarly, and it's unclear how much we would lose by simply always choosing PS-MAP.

Questions

- Sec 3.3: the choice of prior is important; is there guidance on how to choose a prior in practice? - Is there guidance on how to choose between FS-MAP and PS-MAP given a practical problem? How much would the choice matter (the experiments seem to show that the choice doesn't matter much)? - Fig 2b: why is there an increase in curvature a the number of samples increases? - also, is there a reason for measuring flatness as the average eigenvalue (rather than e.g. max)?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The paper clearly discusses the limitations. There is no direct societal impact.

Reviewer VTcx7/10 · confidence 2/52023-07-21

Summary

The paper compares the widely-used parameter estimation of machine learning against function estimation for the maximum a posterior (MAP) estimation. The authors provide detailed analysis and mathematical variation why there are significant difference in results for function vs. paramter MAP, and introduce conditions where the function MAP can avoid its failure modes. To be honest, I am not an expert in the area, and cannot make an assessment how novel the proposed method is against relevant literature. I would rely on other reviewers on the aspect of the evaluation.

Strengths

- The paper is fairly well-written in most of cases, and grounds strong motivation and intuition for the proposed idea. Even though I am not expert in the area, I enjoyed reading through the paper.

Weaknesses

- The paper is heavy on math. Even though they authors did a decent job to make the equations approachable, some of important definitions are missing, and many of details are deferred to the supplementary material. It is suggested to double-check the completeness of the formulation in the main paper. If it cannot be self-contained, I am afraid the publication should head to another venue with a longer length limit (maybe a journal). For example, P was never introduced, but could be only inferred as the parameter dimension. Also, $\theta_R$ and $\theta_L$ was not clearly defined in Figure 1, which is not desirable considering that it should serve as a strong motivating example. - The results are not very competitive. Especially for Table 1 for MNIST and CIFAR-10. Is there any intuitions why it is the case? On the other hand, graphs in Figures appear quite promissing. I would appreciate more discussion in the paragraph on benchmark performance (line 297-298).

Questions

- This might be simply my ignorance. But what is the difference between N samples ($x_\mathcal{D}^{(i)}$) vs. M samples ($\hat{x}$) used in Equation 5? Is it necessary to derive different samples from the space, or can we use the same samples? Roughly how big number would you choose for N and M, respectively? The paper writes condition based on M ($MK \geq P$). But is there any condition on N?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors clearly stated limitations in extensive aspects. However, it would be informative to have a detailed numbers on the increased numerical complexity or time/ memory requirement for using FS-MAP (or LS-MAP) over PS-MAP for some of famous architecture.

Reviewer NXLr5/10 · confidence 3/52023-07-25

Summary

This paper explores the question of learning most likely functions vs most likely parameters in the context of MAP estimation, a common setting including minimizing log likelihood with L1 or L2 regularization, as these regularizations correspond to imposing a prior. Differences in parameterization affect the MAP through change of measure for the prior, and potentially in very unhelpful ways. Because of these potential problems with parameterizations, authors explore doing MAP not on parameters as we usually do (referred to as PS-MAP), but on function outputs for a set of inputs, an idea from Wolpert (1993). Authors refer to this as function-space MAP (FS-MAP). The contributions are as follows: - While Wolpert considered function outputs on a finite set of inputs, authors generalize to function outputs on a measure on input space. Authors emphasize that this measure of input space can be specified according to the task at hand (e.g. restricting to natural images, rather than all images). - Authors empirically compare FS-MAP and PS-MAP on a simple example and empirically find that FS-MAP finds flatter minima, has better generalization in some settings, and produces less overfitting. Authors also discuss how FS-MAP and PS-MAP compare with the posterior mean in variations on this simple setting. Also, FS-MAP’s test performance depends on the measure of inputs that are considered important. - Authors discuss problems with FS-MAP: FS-MAP can be computationally very expensive, and can behave in pathological ways under certain conditions. Authors also provide conditions under which it is well-behaved, as well as a scalable approximation to FS-MAP, called L-MAP, that also behaves ok even when FS-MAP would behave pathologically. Authors empirically compare FS-MAP and PS-MAP on neural networks for image classification. L-MAP performs similarly to PS-MAP, but is better calibrated in one experimental setting, and also finds flatter minima.

Strengths

1. Authors raise and discuss an interesting question that I had not thought about. 2. There is a substantive discussion of PS-MAP and FS-MAP, including weaknesses of both approaches. 3. Authors provide a computationally feasible, well-behaved approximation to a computationally too-expensive and poorly-behaved objective (L-MAP, vs FS-MAP).

Weaknesses

1. There is a tension between statements like “parameters have no meaning independent of the functional form of the models they parameterize” (lines 22-23) and the fact that the difference between best parameters vs best functions is a matter of parameterization of the prior, and the prior is a prior on the parameters (Equation 1). This tension is central to the question of parameterization for MAP and is not addressed. 2. This tension continues into the experiments, many of which emphasize flat minima. Authors cite Dinh’s “Sharp minima can generalize for deep nets”, which points out that the flatness / sharpness of minima is a function of parameterization. 3. Function-space MAP and priors on functions are defined much later in the paper than ideal. For example, in Figure 1, I do not know how FS-MAP is calculated. p_X (important to FS-MAP) is also not specified. 4. The terminology of “most likely functions” and “more probable functions” obscures the dependence of these terms on the choice of p_X (e.g. line 37, 184). 5. The connection to variational inference does not seem particularly strong. MAP as VI would involve choosing the best q, while the construction in Section B.9 does not do that. It’s also unclear to me why we care about these constructions; e.g. why do we care about q, besides that it ultimately leads to (13) and (B.13)? 6. Also, (B.8) looks like a KL divergence, not a variational lower bound / ELBO. Maybe I’m missing something here. 7. The experiments and discussion around experiments could be more thorough. See questions. 8. If someone told me they were looking at most likely functions, rather than most likely parameters for MAP, I would be curious about priors on functions in place of priors on parameters. (This isn't a weakness / question / limitation; it's more of a comment.)

Questions

1. On line 135-136, “the distance between two function evaluations is measured locally by [...]” : what two function evaluations are being referred to? 2. Why is it important to have infinitely many evaluation points? I assume that’s not the appeal but that’s how it was advertised. (Lines 127-128) 3. What is p_X=N/A in Figure 3b? 4. Can you provide intuition for why FS-MAP archives better generalization, in the settings where it does achieve better generalization? Can you also provide intuition for why it is less prone to overfitting? Can you discuss when this wouldn’t be the case? (Section 3.3) 5. The numbers in Table 1 are all very similar. What do you think is the reason for this? Should we be surprised by the similarities between the different choices of p_X, given how much p_X mattered in Section 3.3? 6. “For example, while it is 331 natural to assume that FS-MAP more closely approximates a Bayesian model average”: Why is it natural to make this assumption? (lines 330-331) 7. Could you briefly discuss why the posterior mean is particularly important (line 254)? The citation that follows is a textbook. 8. Why do we want to constrain eigenvalues of the Jacobian? I understand that you can run into problems if they are quite large. Is that the only reason? Is there a risk of over-constraining? 9. Why is calibration better with FS-MAP for CIFAR-10 and p_X=CIFAR-100 compared to PS-MAP, but not other settings? 10. Can you explain further why L-MAP is more robust to training with label noise? The explanation in line 325 is Section 3.3, which appears to be about input space, not about label noise 11. Should I expect these comparisons between FS-MAP and PS-MAP in other settings? In what settings should these results hold vs not (beyond satisfying the assumptions)? 12. While Wolpert’s evaluation on finite sets of points is a special case of the authors’ formulation, are they not effectively the same in practical settings? 13. Authors discuss a measure on which to consider function outputs. I am curious about the connections to distribution shift.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The experiments and discussion around experiments could be more thorough. See questions. It's not clear to me which experimental results should generalize, and under what conditions.

Reviewer VTcx2023-08-14

Thanks for the reply. It cleared most of my concerns.

Authorsrebuttal2023-08-16

Thank you for your feedback and for recommending acceptance of our submission! We will take your comments and suggestions into account when updating the manuscript.

Authorsrebuttal2023-08-21

Thank you again for supporting acceptance of our submission! Your feedback and suggestions were very valuable and have helped us further improve our submission, and we will include the additional results and clarifications from our rebuttal in the revised manuscript. If you agree that the suggested clarifications and new results improved our submission, we would be grateful if you would consider raising your score to reflect these additions. Thank you for your time and effort!

Reviewer wSt32023-08-14

Thank you for the response, it answered my questions and concerns.

Authorsrebuttal2023-08-16

Thank you for your feedback and for recommending acceptance of our submission! We will take your comments and suggestions into account when updating the manuscript.

Authorsrebuttal2023-08-21

Thank you again for supporting acceptance of our submission! Your feedback and suggestions were very valuable and have helped us further improve our submission, and we will include the additional results and clarifications from our rebuttal in the revised manuscript. If you agree that the suggested clarifications and new results improved our submission, we would be grateful if you would consider raising your score to reflect these additions. Thank you for your time and effort!

Reviewer NXLr2023-08-14

Thanks for your response; it addresses most of my concerns and I am also now better informed. I see that there is much discussion about the performance of FS-MAP vs PS-MAP, but not of how to make sense of the question of "best functions" vs "best parameters" given that the setting has a prior on the *parameters*. Perhaps this is not the point of the paper, but I am still wondering.

Authorsrebuttal2023-08-16

Thank you! We are glad to hear your concerns are addressed. While the details of the distinction between most likely functions and parameters does depend on the model parameterization, the existence of this distinction is generic, regardless of the parameterization. Moreover, considering a prior on parameters does not limit the scope of our work. Indeed, any prior in function space can be equivalently expressed as a prior in parameter space using the identity in Equation 6 (or its generalization we presented in Section 3.2).

Authorsrebuttal2023-08-21

Thank you again for supporting acceptance of our submission! Your feedback and suggestions were very valuable and have helped us further improve our submission, and we will include the additional results and clarifications from our rebuttal in the revised manuscript. If you agree that the suggested clarifications and new results improved our submission, we would be grateful if you would consider raising your score to reflect these additions. Thank you for your time and effort!

Reviewer PZwV2023-08-18

Thank you very much for the detailed response and for the additional empirical results! I still think that the paper can be further strengthened if the practical takeaway is clearer (since the comparisons are close and mostly at a small scale), but I think the current results are interesting enough for an acceptance, so I'm raising my score.

Authorsrebuttal2023-08-18

Thank you for your feedback and for recommending acceptance of our submission! We will take your comments and suggestions into account when updating the manuscript.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC