Summary and Contributions: The paper focuses on variational inference fo Bayesian neural networks with gaussian posteriors. The authors aim to improve the expressiveness of the approximate posterior and propose a memory-efficient low-rank approximation to the off-diagonal elements in the full covariance matrix. Furthermore, they propose a novel efficient local reparametrization trick to this more flexible posterior. Surprisingly, they find that augmenting a mean-field posterior does not seem to improve performance, but by taking a rank-0 approximation of the diagonal term, the authors do see improvement, validating the contribution.
Strengths: - Strongly positioned in related work, acknowledging both sides of the discussion for flexible or more restricted posteriors in variational Bayes. - To the best of my knowledge, the paper makes fundamental contributions in the derivation of the low-rank approximation of the off-diagonal term of the covariance matrix, as well as the local reparametrization trick. - Insightful experiments that acknowledge shortcomings and strengths.
Weaknesses: - The work would improve by providing better (visual) intuitions for what the proposed approximate posterior looks like and behaves like compared to a mean-field posterior. - The ELRG-VI posterior is in some sense more restricting than a mean-field posterior. This warrants empirical comparison to other papers that propose more-restricted approximate posteriors. In fact, the constant diagonal covariance can be considered the limiting case of the low-rank diagonal posterior by [Swiatkowski et al.] This suggests that such works should be incorporated in the empirical validations to explore how these two restricted posteriors compare.
Correctness: The proposed reparametrization trick and posterior appear correct.
Clarity: The paper is well written and is a joy to read.
Relation to Prior Work: The paper is well-positioned in both recent and older work in this area. The work clearly acknowledges ideas explored in related work and distinguish those from their contributions. Current areas of debate are highlighted appropriately.
Additional Feedback: - [Swiatkowski et al.] https://arxiv.org/abs/2002.02655 Edit: Although the rebuttal addressed the concern regarding the comparison with the K-tied normal, a closer look at the large scale experiments in the rebuttal paint the proposed posterior in a less flattering perspective than the introduction and conclusion portray. I have decided to lower the score (partly due to my own calibration on reviewer scores), but still vote in favour of accept.
Summary and Contributions: This paper proposes a reparametrization trick to speed up the computation for variational inference (VI) applied to neural networks, where the approximate posterior is a Gaussian family whose covariance is given by a sum of a low rank matrix and a diagonal matrix. By the reparametrization for the low rank structure in layer-wise forward pass, the computational cost is effectively reduced from cubically depending on the number of input and output neurons to linear dependence, which makes its application to large-scale neural networks tractable.
Strengths: The strengths of the paper are: 1. The proposed reparametrization can help to efficiently compute the update of the parameters per layer, which is scalable for large-scale deep neural networks. 2. The computational cost analysis is clear.
Weaknesses: The weaknesses are : 1. It is not clear how the reparametrization differ from the original parametrization with low rank covariance in terms of predictive performance. 2. It is not clear why the more expressive diagonal covariance is less predictive than the scaled identity covariance. 3. Why the latter covariance is more computationally efficient than the former for deep neural networks? 4. Given that the ELRG-VI has worse accuracy as in the Modern CNNs, what is the advantage of using this approximate posterior? 5. Evaluation of the quality of the posterior approximation by predictive performance is not appropriate as in many cases MAP can give more accurate predictions.
Correctness: The claims and method in terms of computational efficiency seems to be correct.
Clarity: The paper is a bit hard to follow with many notations not first defined and explained. For example, 1. what is p(y^*|x^*, \theta) in line 46? 2. what is the parametrization function g in line 59? 3. what is |B| in line 70?
Relation to Prior Work: The differences between this work and prior work are discussed. I am not sure if this work is significantly novel or original.
Additional Feedback: I did not check the details of the setup of the empirical study and cannot say for sure about the reproducibility.
Summary and Contributions: The authors explore the effects and quality of a variational approximation to the posterior of NNs based on low-rank Gaussian distributions per layer. In the course of this, they make the following contribunions: 1. exploration of a local reparametrization trick beyond diagonal Gaussians 2. the authors explore both low rank Gaussians as well as fixed diagonal and learned off-diagonal approximations. 3. the authors rigorously compare their two approximations, MFVI and MAP in a plethora of experiments and find that low rank MVN only performs well on small networks and their second approximation can give performance gains also on deeper architectures. I also want to point out that the paper is well-written, offers plenty of clear and appealing figures and tables and is overall well-executed.
Strengths: The paper lives on its empirical explorations. I think the strongest suit is the insights and comparisons of the two low rank factorizations and the discovery that small networks also work with LR-MVN, but deeper networks work better with the more heuristic fixed diagonal model. Another strong experimental insight is the discovery that per-sample variational samples improve performance vis-a-vis per batch variational samples. It is a nice add-on that the authors develop the local parametrization trick further.
Weaknesses: There are three crucial weaknesses for this paper. First, the paper is conceptually quite thin and does not have many fresh ideas to offer. This is also ok, as the paper is positioned to be very empirically focused. Second, given the empirical focus of the paper, I do not understand why the authors do not provide comparisons with the closely related 'k-tied Gaussians' by Swiatkowski et al paper that they cite in the introduction. How can an empirical analysis be complete for material focused on low-rank MVNs without that comparison? The first complete version of that paper was publicly presented at a Neurips workshop in 2019, so there has been enough time to be aware of its results. I understand the papers have slightly different focal points in that the k-tied paper focuses of covariance structures induced by weight columns, but for an empirical paper I would love to see that effect. Third, in the out-of-distribution experiments, it is unclear to me why the authors highlight their method as better 'due to the balance between uncertainty and confidence'. Is that what we'd expect here? My expectation of OOD data is that classes we have not trained on should yield close to uniform predictions, so I disagree with the author assessment here. Maybe they can provide a gold standard experiment, say via HMC, which can decide that? I am not clear this experiment here makes supporting statements for their method as claimed.
Correctness: Yes, the paper is overall of good quality and I did not detect jarring flaws.
Clarity: The paper is very well written, with clear notation, good overall structure, and great presentation.
Relation to Prior Work: I find the discussion and comparison to 'k-tied' lacking, given its relevance here. I am also perplexed how much this paper ignores Monte Carlo baselines for BNNs in order to evaluate its posteriors.
Additional Feedback: I would consider raising my score with a rigorous comparison to K-Tied. Unfortunately I feel this is warranted for this paper, as the focus is empirical and the paper needs more comparators from the literature. Other than that I found the ideas I highlighted around per-datapoint variational samples and the effects of model depth quite appealing and would be interested in seeing more papers throw in such insights. Edit: In light of the rebuttal, I am raising my score to 6 and thank the authors for incorporating my suggestions.
Summary and Contributions: This paper proposed adding a low-rank term to the local reparameterization trick (LRT) for improving the quality of mean-field approximation and adding additional regularization. Experimental results demonstrate that the proposed method may lead to improved test accuracy than the mean-field approach based on LRT with a small computational overhead. In addition, the uncertainty calibration results look better than MAP.
Strengths: The proposed approach modifies the LRT to facilitate low-rank Gaussian variational approximation for neural networks with a large number of parameters. The complexity of the proposed reparameterization is controlled by the matrix determinant lemma. Both isotropic and non-isotropic Gaussian variance are considered. The experiment evaluations are insightful. For example, the under-fitting issue of mean-field approach is highlighted, perhaps a little bit surprisingly, increasing the number of parameters in the proposed approach does not necessary improve the under-fitting. Moreover, the pitfalls of sharing parameters and uncertainty calibration results of baseline methods are presented.
Weaknesses: Although experimental results show that the proposed method can achieve a good balance between improved test accuracy, ELBO, and test uncertainty, I am still not sure about the takeaways from the paper, and how to use between MAP, MF-VI, and ELRG in practice. The improvement of quality in variational approximation is evidenced by the decrease of ELBO vs. K. It seems MAP can lead to better test accuracy but its uncertainty calibration is bad. Even for K=1, ELRG-VI has better test accuracy than MF-VI. But from Figure 3, it is unclear whether ELRG is better than MF-VI in characterizing uncertainty. Does the improvement of quality in variational approximation translate into better uncertainty quantification? In addition, I am not sure why low-rank plus isotropic variance (ELRG-VI) has better accuracy than the diagonal variant (ELRG-D-VI). Figure 2 (a) shows that for ELRG-D-VI, the negative log likelihoods are equivalent to approaches without low-rank terms. How to interpret the sharp drop and bump? Do the curves in Figure 3 (b) converge?
Correctness: The proposed approach seems correct except for the questions on the algorithm convergence.
Clarity: This paper is well-written and organized.
Relation to Prior Work: Yes.
Additional Feedback: ###### Thank you for the rebuttal! I think the empirical study in this paper is interesting and therefore I increased my score to 6. I hope the following points get addressed in the updated version of this paper. (1) The underfitting of MF-VI is evidenced in Figure 2(c) and Figure 2(a) suggests increasing K in ELRG-D-VI doesn't help. While in the paper it is claimed that "ELRG-VI improves predictive performance WITHOUT underfitting", in Figure 2(b) the likelihood still decreases after ~10 epochs, just now increasing K makes a difference. In the rebuttal, the authors explain that the isotropic variance can force the posterior variances NOT close to prior variances, therefore migrating this issue. I read A.7 but still confused why. More explanations are needed. (2) As indicated by predictive entropy (Figure 4), the uncertainty level of the proposed approach sits in between MAP and MF-VI. But it seems all the methods are more uncertain on OOD data than test data. Based on such results, do we have enough reason to believe the ELRG-VI has a better uncertainty estimate than MF-VI? Is there any other evidence? Does it seem the ECE score results of MF-VI in Table 5 (appendix) is actually better?