NeurIPS 2020

### Review 1

Summary and Contributions: This paper introduces a spike-and-slab prior to Bayesian neural networks using variational inference. The primary focus of the work is the considerable amount of theoretical work to show the consistency of variational inference in this framework. Additional experimental work shows the effectiveness, especially for recovering true variables in a sparse regression setting with low false positive rate and low false negative rate. Update: author feedback is appreciated, and based on reading other reviews, there were some minor issues I seem to have glossed over, and am slightly lowering my score (I'm not sure the author feedback was wholly convincing, but the points are sufficiently minor), but still vote for a clear accept.

Strengths: This work incorporates a natural way of modeling sparsity into Bayesian neural networks, which has the potential to be extremely valuable – allowing for a natural means of expressing model uncertainty while simultaneously providing a natural means of model compression. The theoretical grounding is considerable, and I am unaware of other works that incorporate the spike-and-slab prior/variational posterior family in BNNs, so it is a novel work as well.

Weaknesses: I think the only real limitation of this work is that additional outcomes for success and additional experiments could have been provided, but I think the simulation and real data examples are certainly adequate, especially for a work that has a primary focus on the theoretical nature of the problem.

Correctness: As far as I am aware there are no problems with the method or claims.

Clarity: The paper is indeed well-written, I noticed no issues. The notation is occasionally a bit difficult to follow, but that's more due to the complicated nature of the problem, so I would not recommend any changes.

Relation to Prior Work: I believe prior work was well-addressed as well and clearly shows the contributions made in this work.

Reproducibility: Yes

Additional Feedback: I see no obvious areas that need improvement. Maybe some of the equations (say, equations 6 and 7) could be compressed just to save space, but this is minor at best.

### Review 2

Summary and Contributions: The paper proposes to use a spike-and-slap prior to learn a sparse Bayesian Neural Network (BNN). The proposed variational posterior (again spike-and-slap) is then shown theoretically to be consistent together with a convergence rate result.

Strengths: The approach is well motivated with a theoretically well-founded evaluation.

Weaknesses: What is somewhat lacking is the empirical evaluation and a proper discussion with respect to existing approaches, how the proposed method fits in among them, and why it is needed. (See below for more detailed comments.)

Correctness: The claims and assumptions are well formulated with extensive proofs in the supplementary material and seem to be correct. The same holds for the empirical methodology.

Clarity: The paper is well written.

Relation to Prior Work: While the method itself is well defended, its relation to prior work is somewhat lacking. E.g. while Blundell et al. and Deng et al. are cited, their respective spike-and-slap variants (where the Dirac is replaced with a Gaussian with small variance in the case of Blundell and a Laplace in the case of Deng) are not really discussed and especially not evaluated against.

Reproducibility: Yes

Additional Feedback: #### Post Rebuttal Update I thank the authors for their feedback. While the two weaknesses (a proper discussion of similar approaches in the BNN literature, demonstration of the approach on more complex models) of the paper that I see still somewhat remain, I still recommend acceptance and change my score from 6 to 7. I encourage the authors to extend the theoretical discussion to related prior work in the final version as they already indicated in the rebuttal, and to pursue the question of how to extend the method to more complex models in future studies. ####################### ## Major questions/comments - As mentioned above a proper empirical evaluation and comparison with similar sparsity inducing approaches seems necessary to demonstrate that the strength of the model lies not only in the theoretical contribution but that it is also a valuable addition in practice. As mentioned above especially other spike-and-slap approaches such as Blundell et al. (sofar Blundell is only compared against in the plain Gaussian prior setting), and Deng et al.. But also other approaches such as horseshoe priors (e.g. (Louizos et al. 2017), (Ghosh et al, 2018) or the already cited (Ghosh et al.)). - Similarly, the experimental setting of the paper is constrained to rather small networks, where the sparsity motivation of the introduction is not really necessary. They serve as proper proof of concept that the approach works in principle, but do the authors have an intuition as to how well their approach scales to larger models? ## Minor questions/comments - Sparsity in itself won't help if it is not structured (e.g. dropping a whole convolutional filter helps in saving operations, while only dropping some weights in a filter while keeping the rest still keeps most operations in practice). Can the authors comment on how they assume their approach could be generalized? (E.g. in the direction of structured approaches such as (Neklyudov et al., 2017,...) - l112/113 claim Molchanov et al. to be a frequentist pruning approach. Can the authors give more details about why they do not consider the pruning approach of that paper to be Bayesian? - The regression results in the appendix are only on a subset of the commonly used UCI setup from Hernandez-Lobato and Adams. Is there a reason why those five are chosen over the others? - The authors provide all the necessary details for a reproduction of the results in the main paper & the supplementary together with code. A very minor improvement would be to mention in the respective places (Figure 1/Table 1/2) what the error bars are for completeness. (They are probably all standard deviations in the main paper? But e.g. in Table 2 in the Supplementary they are standard error instead.) ______ Ghosh et al., Structured Variational Learning of Bayesian Neural Networks with Horseshoe Priors, ICML 2018 Louizos et al., Bayesian Compression for Deep Learning, NeurIPS 2017 Neklyudov et al., Structured Bayesian Pruning via Log-Normal Multiplicative Noise, NeurIPS 2017

### Review 3

Summary and Contributions: The authors proposed a sparse Bayesian deep neural network model with spike-and-slab prior. This work is motivated to close the gap between theoretical studies and practical studies in deep Bayesian neural networks. From this perspective, they provided the theoretical guarantee for the variational consistency. The authors also showed that the corresponding convergence rate strikes the balance of statistical estimation error, variational error, and the approximation error when we chose the hyperparameter of Bernoulli distribution in spike-and-slab prior. Furthermore, through several experiments, they confirmed the validity of their theoretical results and demonstrated that the proposed method achieved the better performance of variable selection and uncertainty quantification than previous work.

Strengths: ・Although the idea itself is a simple (using sparse prior; spike-and-slab prior), the author provided a valid theoretical guarantee for consistency and showed the optimal hyperparameter value of Bernoulli distribution in spike-and-slab prior, that is helpful for users and guarantees the performances of the proposed method. ・There are enough experiments to confirm the theoretical results and predictive performance, uncertainty quantification, and variable selection for multi-layer networks. ・The paper is well-written and easy to understand what they want to do and what the contribution is.

Weaknesses: Limitation : (a) : Theoretical analysis only holds when we use the Gaussian variational distribution with the same family of the spike-and-slab prior. Therefore, the performance is not guaranteed when we set the other variational distributions. (b) : They conducted the experiments on a simple dataset, e.g., MNIST. Therefore, we cannot understand whether the performance holds when we want to apply the more complex BNN model for more complicated tasks. Furthermore, the performance is not guaranteed when using the more sophisticated variational inference framework for complex BNN models, e.g., black-box VI (Ranganath, 2014) or hierarchical VI (Ranganath, 2016). On more complex models and datasets (at least, Fashion-MNIST), how much better does the proposed method perform in terms of prediction accuracy or/and uncertainty quantification compared to comparative methods? (c) : There are no experimental results for convergence speed in terms of real-world time, which is one of the open problems in BNN. Certainly, the proposed method is effective for the memory storage and computational burden, but it is not clear how fast the inference is. How fast can the proposed method conduct inference with sparsification and variable selection?

Correctness: It seems that the claims from theoretical analysis and the empirical methodology are correct and reasonable.

Clarity: The structure and organization of the presentation are good. The explanation of contribution is clear and easy to follow the related work.

Relation to Prior Work: The related work section is well-written. This paper's contribution differs from the previous contributions in terms of the variable selection (for multi-layer networks) and providing the essential theoretical analysis for consistency in Bayesian deep learning.

Reproducibility: Yes

Additional Feedback: All of the feedbacks, comments, and suggestions are in the above sections. ========================= After reading the author response ========================= I would thank the author(s) for their feedback. I read it and the other reviewers’ comments. After reading this, my concerns almost has been addressed. I still think that it is important to compare the empirical results based on real-world time, even if the authors do not have the cutting-edge infrastructure; however, it does not largely reduce the contributions of this paper. Therefore, I decided not to change my score as 7. I recommend the authors to show the more experimental results on more complicated dataset in the future version of this paper, at least, on Fashion-MNIST, as the authors reported in their feedback. Furthermore, I encourage to extend the theoretical discussion to related work, as the authors claimed in their feedback. I’m looking forward to seeing their future work for the computational efficient algorithm of the proposed approach.

### Review 4

Summary and Contributions: The work investigates sparsity inducing spike-and-slap priors for training multi-layer neural networks. It derives theoretical requirements on hyperparameters of the prior distribution for posterior contraction. Finally, the work empirically shows that the devised approach leads to better recovery of ground truth sparsity in a teacher student experiments and sparse function regression problems. In both cases the effect of the prior hyperparameter setting is evaluated and it is shown that the theoretically motivated value leads to best performance.

Strengths: The paper applies the insights derived from theoretical considerations well to empirical evaluation scenarios. The experiments are detailed and the suggested approach is very well evaluated on multiple toy and real world datasets. The claims and theoretical results of the paper are thus well supported by empirical evidence. Overall, I would consider the work significant and of relevance to the NeurIPS community.

Weaknesses: The paper introduces a theoretically sound way for the selection of the hyperparameter of the prior $\lambda$. This hyperparameter can to my understanding be interpreted as the prior expected amount of sparsity in the weights and thus could also be utilized when benchmarking the other methods. This should for example easily be possible for the variational dropout baseline and would allow to better understand if the higher performance is due to the more theoretically motivated selection of $\lambda$ or the way the spike and slap prior is implemented. Further, I cannot find a description over which hyperparameter range the sparsity was evaluated for the approaches which require a prior definition of the degree of sparsity (AGP and LOT). Here it would be relevant to know if the exact degree of sparsity was in the hyperparameter grid. Also, similar to the previous point, how do these models perform, when the prior degree of sparsity is set to $\lambda$? Finally, I would be interested in seeing a comparison to methods with "adaptive sparsity" such as the horseshoe prior. Here a prior estimate of the sparsity degree as derived in the theoretical results can also be included in the prior definition.

Correctness: To my understanding the claims and the methodology of the paper are correct. Yet, I am not an expert in this particular field.

Clarity: The paper is well written and clear.

Relation to Prior Work: While the paper refers to previous work, it is in my opinion too coarse in it's description. In particular, the statement "XYZ propose Bayesian Methods, but the choice of prior specification still needs to be determined.", does not clearly describe how the previous work is related besides proposing Bayesian Methods. Here, a more nuanced description would be beneficial.

Reproducibility: Yes

Additional Feedback: Minor mistakes: Missing word: line 156: , then *the* existing result Some references are out of date and refer to publications on arXiv while the papers have actually been published in proceedings (Ghosh and Doshi-Velez 2017). There might be a mistake in the equations after line 18 of the appendix. Here the sum seems to be iterating over the wrong variable. In the proof of Lemma 4.1 in the appendix, condition 4.4 seems to be required. This is in contrast to the main text where Lemma 4.1 is assumed to be valid under conditions 4.1-4.3. Questions: Regarding Table 1: At which level was a weight considered off / on. Or is this the average number of active weights for draws from the approximate posterior? ----- Read other reviews and author response. The authors address some of my questions but not all. In particular, it seems like the methods which require the definition of a degree of sparsity were not tested with the degree of sparsity set to the true sparsity of the data. Thus, they would never be able to give optimal performance. This is sensible, as the degree of sparsity is usually not known ahead of time, but should be explicitly mentioned to support the interpretation of the results. I encourage the authors to reflect this point in the final version of the paper.