- 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 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, 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.jsonfile 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
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.
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 and −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 2.25? With N neurons and L components, the minimum logL grows like 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 0.25 nats in the same example, the minimum logL grows like N (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 m costs at least m2/2 in KL. Shifting each outgoing mean by m/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 neurons. That can shift the output by a constant because the network divides their sum by 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 logL=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, the exact forecast for the second is −0.217 and the optimal full Gaussian gives +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.
Experiments. You are right that fits with L=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 L products can improve the best single-product ELBO by at most η1logL. Here η1=M/(1+M), where M is one quarter of the largest squared column norm of the feature matrix (Eq. (70), p. 49).