acceptodds
Under review as a conference paper at ICLR 2027

The Representation Cost of Conditioning

Abstract

Approximate inference in Bayesian neural networks requires choosing which weight dependencies to represent. We study shallow networks with a Gaussian prior, fixed data, Lipschitz activations and smooth convex likelihood energies of at most quadratic growth. Products over whole neurons allow dependence within each neuron; in Gaussian regression, their limiting optimum recovers the posterior mean but retains prior fluctuations. For mixtures of products over individual weights, the minimum log component count at width is to approach a strictly better whole-neuron variational value and to approximate the full weight posterior in KL, at fixed tolerances below the respective positive gaps. The factors may be non-Gaussian. For ReLU, full Gaussian weight distributions with bounded prior KL can contract one direction of limiting uncertainty while missing mean directions available to whole-neuron products. The optimal ELBO ordering changes with the labels. In the width limit, the optimal full Gaussian can give a conditional forecast in the wrong direction even after both predictive marginals are made exact. Finite fits, a covariance intervention that preserves the mean function and a separate preference-head study illustrate consequences for prediction and evidence-based prior selection.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Bayesian neural networks, variational inference, posterior approximation, product mixtures, dependence representation, Gaussian processes

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_GEDLOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This work tackles an interesting question in variational inference for Bayesian neural networks, namely, what different kinds of dependence structures one needs to accurately represent different functionals of the posterior (e.g., its mean, predictive covariance, and conditional predictions). By analyzing a single hidden layer ReLU network, the authors show that recovering the posterior mean and capturing posterior uncertainty impose substantially different representational requirements. In the infinite-width regime, the authors show that whole-neuron product variational distributions can recover the posterior mean in a Gaussian regression setting. Mixtures of scalar-factorized distributions (as defined by the authors) require exponentially many components to approximate the correct mean. The authors also characterize limitations of full Gaussian weight distributions, showing that full-covariance Gaussian weight approximations can update predictions in the wrong direction after observing another response, even when individual predictive marginals are exact.

Critical strengths

The presented results and the idea is interesting: What dependence structure is needed to represent BNN posterior distributions well and what variational families can represent these dependencies?

Critical weaknesses
  • Even after reading the whole paper, it does not become clear to me what the submission's actual contributions are, especially over the related literature. I think that the paper develops interesting results, such as a charaterization of what full-covariance Gaussian variational distributions can represent, conditional prediction failures, and rates for posterior approximation using different variational distribution families, but all of these are severely mixed up and, in my opinion, impossible to understand when reading the paper. Take the abstract as an example.
  • Consequently, the submission reads like a condensed mathematical research summary, paired with an extensive appendix, but it is almost impossible to follow the argument as a reader, as the paper does not connect the results with each other or even talks about what the results mean, why they would be interesting or how they fit into the bigger picture of the submission.
  • For instance: Eq. 6 is introduced as "combining" earlier facts but the text never explains where the term 0.5m20.5 m^2 comes from; the first term in Eq. 7 is not explained, nor is there a reference of where it comes from; rectangular cells, channel resolvability and sparse signaling (lines 181-187) are cited before Th. 2 without saying what each contributes to the construction; the start of section 5 asserts an excess log loss result for whole-neuron mixtures in one sentence but there is no formal statement, and the topic is dropped; Section 7 uses a setup that is completely different from the setup in the other sections (no hidden layer), also, η1\eta_1 is undefined in the main text; a lot of different VI families appear (I counted at least 8) without overview which makes it hard to keep track. Essentially, the main text reads like a concatenation of abstracts.
  • The experiments are disconnected from the theory, for instance the authors admit in line 411 that the L=1 experiments do not test mixture rates, so the experiments say nothing about Theorems 1-3; Section 7 is a different setting entirely.
  • The submission title seems off and generic, the topic of the paper does not follow from it, also, conditioning is never defined and it remains unclear what it refers to.
  • The code supplement contains evidence for the heavy use of an automated research pipeline with questionable human supervision, such as:
    • folder and file names that refer to round-based autonomous agent execution with different agents having different roles (execution, interpretation of results, ...)
    • later research "rounds" that pick up and respond to results in earlier rounds
    • automated verification of experiments
  • It is questionable whether the AI statement of the submission is misrepresenting to what (large) extent agents where employed for autonomous research, execution and interpretation, as the AI statement comes across as LLMs "helping" with almost every task that led to the creation of this submission, not that LLMs/agents could have essentially produced all of its contents.
  • The authors also (mis-)use non-standard and confusing terminology in the context of variational inference for BNNs that makes the submission difficult to read, for example "representation cost", "posterior contraction", and "closure".
Questions for authors
  • Can the authors please clarify their contributions, in particular with regards to the Coker et al. paper, the Hron et al. paper as well as the Foong et al. paper?
  • Can the authors clarify what stages of this submission included autonomous agent decisions and where independent human supervision was part of the process?
  • How exactly was the content of this submission verified? By means of a separate implementation, a separate agent, or an independent human assessment? Which proofs and claims received human checks?
  • The omissions.json file that is part of the code supplement specifies that "reviewer/agent conversations" were omitted from the code supplement. Can the authors elaborate what these conversations look like?
2 comments

Thank you for reading both the paper and the supplement. We introduced several approximations before explaining what earlier work had shown. That made it hard to see what our results add.

What can independent weights learn? Coker et al. gave a ReLU example where independent Gaussian weights learn a nonzero mean prediction. In our shallow model, the best limiting mean is proportional to the input norm even for non-Gaussian independent weights (Theorem 1). It must be the same at +1+1 and −1-1, where the posterior means in Figure 1 have opposite signs.

We can keep different neurons independent and still allow each neuron's incoming and outgoing weights to depend on each other. In Figure 1 the best such approximation gets both means right. This lowers the best limiting value of the objective (KL to the prior plus expected squared error) from 3 to 2. How many independent-weight distributions must we mix to reach 2.252.25? With NN neurons and LL components, the minimum log⁡L\log L grows like N\sqrt N (Theorem 2).

Even at the whole-neuron optimum, each output variance is twice the posterior variance. To get the KL to the full weight posterior below 0.250.25 nats in the same example, the minimum log⁡L\log L grows like NN (Theorem 3). This count at a given accuracy in a shallow network is what we meant by "representation cost."

Mixture bounds. In Eq. (6), the normalized sum of outgoing weights has a standard normal prior. Giving it mean mm costs at least m2/2m^2/2 in KL. Shifting each outgoing mean by m/Nm/\sqrt N attains it. The first term in Eq. (7) gives the corresponding Gaussian KL for shifting the mean outputs at the training inputs while keeping their covariance fixed. Appendix A (pp. 15-16) shows that independent neurons can attain this cost and describes the output distributions they can approach at fixed inputs as width grows. We called those possible limits the "closure."

To attain the first mixture rate, the construction changes about N\sqrt N neurons. That can shift the output by a constant because the network divides their sum by N\sqrt N. Restricting each changed weight to its own interval gives a rectangular cell and preserves independence within a component. Channel resolvability bounds how many such patterns suffice (Appendix B.2-B.3).

Variance and prediction. Suppose training reduces variance on a set of inputs with positive test probability. That reduction is what we meant by "contraction." In noisy Gaussian regression, whole-neuron mixtures with bounded prior KL and log⁡L=o(N)\log L=o(N) then have higher average log loss than the exact posterior even as width grows. This is Proposition 10 (p. 27), which we should have cited at the start of Section 5.

For two responses we also need their covariance. Foong et al. showed how factorized weights restrict variances at individual inputs. Our Eq. (9) shows that a full Gaussian over ReLU weights with bounded prior KL can change the limiting output covariance in only one direction. In Figure 3 that direction moves both outputs together. We first give each noisy response its correct marginal distribution. We then condition on the first response to predict the second. If the first is +1+1, the exact forecast for the second is −0.217-0.217 and the optimal full Gaussian gives +0.122+0.122.

Hron et al.'s construction (our Eq. (10)) keeps the hidden weights at their prior and uses the posterior of the outgoing weights given the hidden weights and data. It recovers the exact limiting forecast here. Our Figure 3 comparison asks whether we can recover the sign at the same KL to the prior as the full Gaussian. At the training input, the hidden activations determine how much each outgoing weight contributes to the output. Reducing the variance of that weighted sum gives −0.038-0.038.

Experiments. You are right that fits with L=1L=1 cannot test either growth rate. The independent-weight fits let us compare the fitted mean with Theorem 1's prediction. In Figure 4b we held the mean function fixed. Restoring the outgoing weights' prior covariance raised estimated log loss at all five inputs at each of four widths on one dataset.

Section 7 asks whether the approximation can change which prior scale we choose. In a separate preference model with fixed features and no hidden layer, the evidence favored scale 1 and the restricted ELBO favored 0.3. At scale 1, mixing LL products can improve the best single-product ELBO by at most η1log⁡L\eta_1\log L. Here η1=M/(1+M)\eta_1=M/(1+M), where MM is one quarter of the largest squared column norm of the feature matrix (Eq. (70), p. 49).

I thank the authors for their response.

Unfortunately, the authors have addressed none of my points and I believe that this submission requires at least a major round of revision, so I keep my score.

Reviewer_QMAFOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies Bayesian Neural Networks (BNNs), in particular how different ways of approximating the posterior over weights affects what is learned. The authors focus on the asymptotic case of infinitely-wide single-hidden-layer networks.

Critical strengths

Assessing strengths and weaknesses of BNN approximate posterior factorizations without particular distributional commitments is a nice approach.

The paper builds directly on and addresses some prior concerns of recent work, namely Coker et al 2022.

Critical weaknesses

This was very difficult to read. This paper has very little preamble, it has no prior-work or related-work section, and it does little to define its terms but uses jargony terms like "converges to the prior in KL" or "covariance contraction" etc. While one can try to infer what these terms (and many others) mean from context, it makes for an unpleasant reading experience.

Far too much material is defined in or otherwise deferred to the appendix.

Figures illustrate results but include no panels that would help setup up the problem or show context. A single "concept" or "experiment setup" panel per figure could greatly improve the readability.

The theorems contain conditional statements ("If N<<logLN<<N\sqrt N << log L_N << N ..." and "If logLN=o(N)log L_N = o(\sqrt N)...") but as far as I could tell the preconditions for these statements are not assessed.

Figure 2 shows that different factorizations are preferred in different regimes, according to the ELBO objective. One wonders whether the same is true just considering held-out predictive performance (the likelihood); in other words, if this was not a Bayesian paper then I would want to see which model does better just on test error, not on test error + regularization, which is what the ELBO measures.

Questions for authors

The results of Farquhar et al (2020) seem to call the whole research program into question: Farquhar et al show that deep BNNs suffer far less from issues of posterior approximation than shallow BNNs. The present paper, meanwhile, studies infinite-width shallow BNNs. While Farquhar et al are cited, the more substantial implication that deep BNNs may not suffer the same issues is never discussed. To what extent can we expect the present analysis to be relevant for deep BNNs?

1 comment

Thank you for the careful reading. We should have put more of the setup in the main text and explained the related work earlier. For example, consider independent weights and a tanh⁡\tanh activation in Theorem 1 (p. 4). As the network gets wider, the best approximation over weights approaches the distribution we had before seeing the data. That is what we meant by "converges to the prior in KL." A panel showing the four inputs and their labels would also have helped in Figure 2.

Does this extend to deep networks? We do not know whether the same limitation holds when all layers of a deep Bayesian network are learned together. Farquhar et al. (2020, Proposition 4) show that two hidden layers with independent weights can approximate a wide range of predictive distributions at individual inputs. Their experiments also give cases where depth helps. As they note, the weight distribution in their construction can still have a poor ELBO. We ask which weight distribution has the best ELBO for a fixed prior and dataset, and whether it gets the relation between two predictions right. We answer this for one hidden layer. The result also applies if we fit that model to features from a frozen deep network.

Would the same model win on new test inputs? Figure 2 compares a full Gaussian over all weights with a whole-neuron product, which keeps neurons independent but allows dependence within each neuron. The four training inputs lie on the unit circle. The two on one axis have label +t+t and those on the other have label −t-t, where t≥0t\geq 0 sets the label size.

To answer your question, we calculated expected test loss in the same infinite-width model. We chose the test input uniformly on the circle and averaged over the possible noisy responses there using the exact posterior. At t=0.6t=0.6 the ELBO favors the whole-neuron product, but the full Gaussian has lower test log loss by just 0.000640.00064 nats per response.

CriterionWhen does the whole-neuron product do better?
ELBO (submitted Figure 2, p. 6)For tt above about 0.4840.484
Expected test log lossFor tt above about 0.7190.719
Expected test squared errorFor every t>0t>0, because it gets the posterior mean right

Log loss also depends on the spread of the predictions. The full Gaussian can reduce uncertainty about the average of the four outputs. This is what we meant by "covariance contraction" in Figure 2.

When do the mixture bounds apply? Each component in Figure 1 keeps the individual weights independent. We choose how many components LL to mix at each width NN. With L=1L=1, log⁡L=0\log L=0 at every width. If log⁡L\log L grows like N3/4N^{3/4}, it grows faster than N\sqrt N and slower than NN. With the opposite labels in Figure 1, the best limiting objective (KL to the prior plus expected squared error) is 3 with one component and 2 with this larger mixture (Theorems 1 and 2, p. 4). If both labels are zero, the two limiting values agree. A gap to the full weight posterior remains (p. 5). Our finite-width fits used one component throughout, so they did not measure how the required mixture size grows.

We'd be glad to share the test-loss calculation if helpful.

Reviewer_bvPbOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates the quality of different approximate posteriors in neural networks, specifically, the full Gaussian, and mixtures of mean-field approximations, where the form is unconstrained. The paper claims a variety of results, although it is very difficult to discern them.

Through careful reading, and re-reading, I can discern a few:

  • The mixture-of-mean-field approximate posterior needs exp(N^0.5) to improve the KL bound over... I can't really tell. It refers to a "whole-neuron variational value". What variational posterior is this better than? It seems that this is implied by Theorems 1 and 2, but the theorem is stated in quantities that don't relate to what we actually want to prove.
  • ... and it needs exp(N) to get close to the full weight posterior.
  • It is really difficult to figure out what section 4 implies.
  • I understand the main claim of section 5.1, but how exactly Proposition 4 implies this, is unclear.
  • The claim in section 5.2 is unclear.

I have similar issues with sections 6 and 7.

Critical strengths

The paper addresses an important problem, of understanding what characteristics of approximate posteriors matter for getting a good variational approximation.

Critical weaknesses

Blundell et al 2015 is certainly not the first example of a Bayesian neural network, that uses a distribution the weights.

The introduction is already hard to follow, introducing much terminology without mathematical definition. What is a "whole-neuron product"? What is dependence within neurons? What are products over neurons? You need mathematical notation to make these things clear. Currently, it is not possible to understand what is meant in the introduction. This is not a good state of affairs, even if things later become clearer in the rest of the paper.

The notation in equation 2 is unclear. It can be easily written as a density, which would allow the disambiguation between measures / densities (which depend on the weights) and scalars (e.g. Z_N).

Line 45: The first obstruction to what? Hard to understand this writing. This is probably referring to something defined in the introduction?

In line 203, how can V be a function of the matrix Gamma_H, if it was defined as a likelihood with as input a vector of function values?

Weaknesses in writing are also discussed in the summary.

Questions for authors
  • Can you explain the claim in section 4?
  • Can you explain how Proposition 4 implies the claim of the section?
  • What is the main claim in section 5.2, and how can I understand this from what is written?
  • See other issues in "paper summary".

Overall, I believe that this paper has merit, but I cannot see how a reader of the paper is supposed to glean this. If authors can provide very succinct explanations of what each section claims, and evidence that it was written in the original version, I am inclined to update my score.

2 comments

Thank you for taking the time to reread the paper. We should have made the baseline clear. The page numbers below are from our original submission.

Baseline: Mean field makes all scalar weights independent. A whole-neuron product Q=⨂iQiQ=\bigotimes_i Q_i keeps neurons independent. Within neuron ii, QiQ_i is a joint distribution for its incoming and outgoing weights (p. 3). In Figure 1 we minimize KL to the prior plus expected squared error. The limiting minimum is JS=3J_S=3 for mean field and JB=2J_B=2 for the whole-neuron product. By Eq. (2), this also lowers KL to the weight posterior by one nat.

If we mix LL mean-field distributions at width NN and log⁡L=o(N)\log L=o(\sqrt N), the best limiting objective is still 3 (Theorem 1). At bounded prior KL, these mixtures must give +1+1 and −1-1 the same limiting mean, although the posterior means are opposite. This is the "first obstruction" (Eq. (5)).

  • Theorem 2. To reach any fixed objective between 2 and 3, the minimum log⁡L\log L grows like N\sqrt N.
  • Theorem 3. At the whole-neuron optimum the means are right, but the limiting output variance at each input is twice the posterior variance. The remaining KL to the full weight posterior is 1−log⁡21-\log 2 nats. To get within a fixed positive tolerance below that value, the minimum log⁡L\log L grows like NN (pp. 4-5).

Section 4. All four inputs in Figure 2 have unit length. Both ends of one axis have positive labels and both ends of the other have negative labels. For ReLU and bounded KL to the prior, the limiting mean under a full Gaussian is a linear function of the input plus a multiple of its norm. It misses this pattern. Its gain from reducing variance in the average output stays fixed as the labels grow, so the whole-neuron product eventually has the higher ELBO (Eq. (9), pp. 5-6).

Section 5.1. The exact forecast for the second noisy response is a decreasing line through zero as a function of the first. Proposition 4 gives the full Gaussian an increasing line through zero even after each response has the correct marginal distribution. Predicting zero is closer to the true conditional mean for either sign of the first response. Ignoring it therefore gives lower average squared error (p. 6).

Section 5.2. The training response measures a weighted sum of outgoing weights. The neurons' activations at that input determine the coefficients. If the first response is +1+1, reducing variance in that sum gives a forecast of −0.038-0.038 where the full Gaussian gives +0.122+0.122. Both weight distributions have the same KL to the prior (Figure 3). Hron et al.'s construction keeps the hidden weights at their prior and uses the posterior of the outgoing weights given the hidden weights and data. Its KL to the prior is larger in this example, and it recovers the exact limiting forecast (Eq. (10), pp. 6-7).

Predictor (Table 2, p. 34)Forecast for the second response if the first is +1Mean squared error over all pairs
Exact posterior-0.2170.345
Ignore the first response00.362
Optimal full Gaussian+0.1220.387
Reduce variance in the weighted sum-0.0380.357

In Figure 4b we constructed two approximations with the same mean function. Restoring prior covariance raised estimated log loss at all five inputs and four widths on one dataset (p. 8). In the separate preference model, evidence favored prior scale 1 and the restricted ELBO favored 0.3 (Figure 5, p. 9).

We agree Eq. (2) is clearer as a density. It gives pN(θ)=uN(θ)e−V(fθ)/ZNp_N(\theta)=u_N(\theta)e^{-V(f_\theta)}/Z_N, with prior density uNu_N, training outputs fθf_\theta, and scalar normalizer ZNZ_N. In Eq. (4), ΓH\Gamma_H is a Gaussian vector with covariance KHHK_{HH}, so VV takes a vector. Bayesian neural networks predate Blundell et al. We should have cited them specifically for Bayes by Backprop.

We'd be glad to expand on the calculations if helpful.

I maintain my concerns about the understandability of this paper. There certainly can be merit to these results, but even as someone with a long-running interest in Bayesian neural networks, the paper is not clear enough. The issues are too large for the rebuttal to fix.

Reviewer_V6CHOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies how dependence between weights affects variational approximations in wide, shallow Bayesian neural networks.

Critical strengths

None

Critical weaknesses

These are raw notes taken while reviewing your submission:

  • Overall, I find the paper poorly written and borderline unreadable. It lacks a clear objective, and its contributions and relation to the literature are unclear.

  • The notation is sloppy, with many terms used without being introduced. For example:

    • Line 111: the subscript ss is undefined.
    • Eq. (3): what is πj\pi_j? What are superscript aa and bb?
    • QQ is never defined, and the notation uses four different forms: what are QQ, QiQ_i, QijQ_{ij}, and QijkQ_{ijk}? What does each index refer to?
    • Eq. (5): where does this inequality come from?
    • Eq. (6): what is ψH\psi_H?
  • Some sentences are difficult to make sense of. And these problems already start in the abstract, which is unclear and uses terms without introducing them. For example: "In the width limit, the optimal full Gaussian can give a conditional forecast in the wrong direction even after both predictive marginals are made exact.". Forecase of what? What directions? Which "both marginals" are we talking about? Another example, lines 47–50: "A mixture of L products selects one component for the whole network, then draws individual weights independently in the original coordinates. A whole-neuron product instead permits dependence within neurons while keeping neurons independent." I do not understand what you are trying to say here.

  • The overall state of the presentation prevents me from further assessing the quality and mathematical correctness of the work.

Questions for authors
  • Can you summarize the contribution of your paper in one sentence? I expect a clear list of 2/3 contributions.
1 comment

Thank you for pointing out where the paper became hard to follow. Figure 1 looks at two inputs, +1+1 and −1-1, as the number of hidden neurons tends to infinity. We allow a neuron's incoming and outgoing weights to depend on each other. We keep the weights of different neurons independent. This is what we mean by a whole-neuron product.

At both inputs, the best approximation of this kind gives the correct mean output but twice the posterior variance.

  1. Independent weights. Coker et al. (2022) showed that a ReLU network with independent Gaussian weights can learn a nonzero mean prediction. In our shallow ReLU model, the optimal limiting mean is proportional to the input norm, even when the independent weights are non-Gaussian (Theorem 1). Since +1+1 and −1-1 have the same norm, this approximation must give them the same mean.

  2. Mixtures. In Figure 1 the best objective (KL to the prior plus expected squared error) is 3 with all weights independent and 2 when we allow dependence within each neuron. Suppose we mix LL independent-weight distributions at width NN. To reach any fixed target strictly between 2 and 3, the minimum log⁡L\log L grows like N\sqrt{N}. Approximating the full weight posterior to a fixed KL tolerance can require log⁡L\log L of order NN (Theorems 2 and 3).

  3. The forecast. Figure 3 gives a separate example in the same width limit. We observe one noisy response and predict a second. If the first response is +1+1, the correct forecast for the second falls from zero to −0.217-0.217. The optimal full Gaussian over the weights raises it to +0.122+0.122. This happens even after we rescale the approximate responses so that each has the correct distribution on its own. Those two individual distributions are the marginals in the abstract.

In Eq. (3), first choose one component jj for the whole network with probability πj\pi_j. That choice determines which distribution to use for each weight. Given jj, sample each weight independently from its distribution.

SymbolMeaning
QQDistribution of all the weights
QiQ_iJoint distribution of neuron ii's weights in a whole-neuron product
QijaQ^a_{ij}Distribution of neuron ii's outgoing weight in component jj
QijkbQ^b_{ijk}Distribution of neuron ii's kkth incoming weight in component jj
ss (line 111)Number of observed inputs in HH
ψH\psi_H (Eq. (6))Vector of mean activations at those inputs, averaging over the incoming weights under the prior

For Eq. (5), consider one neuron in one component. Its outgoing weight and its activation are independent, so its mean output is the product of their means. The extra shift in Eq. (5) is the mean outgoing weight times the change in mean activation from its prior value.

Gaussian concentration bounds each factor in magnitude by a constant times the square root of its KL cost. Using 2xy≤x+y2\sqrt{xy}\le x+y then bounds the product by a constant times the sum of the costs. We sum over neurons with the 1/N1/\sqrt{N} factor from Eq. (1) and average over components. The average component KL is the mixture KL plus the information the weights reveal about jj. That extra term is at most log⁡L\log L (Appendix B.1, p. 17).

We'd be glad to expand on the calculation if helpful.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.