NeurIPS 2020

Replica-Exchange Nos\'e-Hoover Dynamics for Bayesian Learning on Large Datasets


Review 1

Summary and Contributions: This is an interesting, if complicated, paper on using a combination of Nose-Hoover dynamics with replica exchange for training Bayesian models. The paper draws on literature of molecular modelling and applies this to more general Bayesian inference tasks. The methodology is not entirely novel as similar schemes appear already in chemical physics, but the use is new.

Strengths: Even if these types of methods are in use in molecular dynamics, they are not explored in machine learning and may be quite useful in this context. The article represents quite substantial numerical work to implement the indicated scheme.

Weaknesses: Designing the temperature ladder is always a substantial challenge in molecular modelling and is, if anything, even harder in machine learning contexts since we have no physical guidepoints to choose the cutoff temperatures. In practice the hyperparameters representing the large and small temperature will need to be tuned for efficiency and are model dependent. This question is not addressed at all in this artlcle. The numerical method used for NH dynamics with additive noise (adaptive Langevin) within the replica exchange framework of Algorithm 2 could be improved for little cost by better choice of splitting, see Leimkuhler and Shang, SIAM J. Sci Comput. 2016 https://epubs.siam.org/doi/pdf/10.1137/15M102318X

Correctness: It seems to be generally correct, although not all statements have been checked.

Clarity: Yes it is reasonably well written.

Relation to Prior Work: There is a genuine effort cite other works but I think some others could and should be mentioned: http://www2.stat.duke.edu/~scs/Papers/RemdJCP.pdf https://pubmed.ncbi.nlm.nih.gov/26203017/ They should also be carefully discussed in the text.

Reproducibility: Yes

Additional Feedback: Concerns have been raised regarding the ML examples considered here. Although a good submission and interesting, I am swayed that further work could benefit the article in demonstrating the practical value of the approach. Still I think this is a very good article.


Review 2

Summary and Contributions: This paper proposes a sampling methods based on multiple particles following Nosé-Hoover dynamics at different temperatures. The goal is to be able to 1) use the different temperatures to explore the whole state space and discover the modes 2) make it amenable to the large scale setting by mitigating the noise induced by the minibatch. They derive (asymptotic) acceptance tests to exchange the particles at different energy levels to ensure exploration while sampling only at T=1 (i.e. the right distribution). They provide synthetic experiments (well-separated mixture of gaussians) as well as sampling from the posterior of neural networks (on Fashion MNIST and CIFAR-10). ======= I have read the authors' response and lower my score to a 6. 150 epochs should be enough to reach good accuracy on CIFAR-10 and methods are not directly comparable if not properly tuned. In addition, the other reviewers raise good concerns, thus my lowering of the score.

Strengths: The paper does a good job at explaining the challenges of various methods and convincing us that their method makes sense. Furthermore, the method seems to be addressing many of the shortcomings of previous work. Section 2.2 is harder to follow, especially the derivation of the logistic test. The exposition seems correct and complete. They also explain clearly how to choose the hyperparameters of their method (e.g. test and bandwidth).

Weaknesses: One of the main weakness in my opinion is the experimental setup. First of all, on the synthetic example, it seems that a few particles of HMC initialized with a large isotropic gaussian should be able to capture all the modes---even though individual particle might not be able to traverse the low density region. Similarly, HMC with tempering should also be able to do the job. On CIFAR, the "optimization" baselines seems very weak in my opinion. SGD with a standard stepsize schedule and ResNet-18 should reach accuracy > at least 93-94% in less than 300 epochs. As such, it is hard to put the performance of the author's method in context. The fact that SGD/Adam degrades much more in the presence of noise is however encouraging. Another weakness is that, if I understood correctly, the test for the replica-exchange process only gives the right distribution asymptotically. While this can be acceptable in practice, I would have liked to see either under which conditions this is close to the true distribution or an empirical evaluation of the discrepancy.

Correctness: The claims and method seem correct albeit the baselines are a little weak.

Clarity: The paper is very well written and easy to follow. The formatting bothered me as it seems like a lot of white space was unnaturally eaten (e.g. margin between section titles, paragraphs, algorithms, footers etc...). Some equations also seem very tiny (e.g. Eq. (10')) as well as the table and algorithm.

Relation to Prior Work: The work is well put into context and it is clearly explained which variation address which challenges.

Reproducibility: Yes

Additional Feedback:


Review 3

Summary and Contributions: The paper considers the problem of sampling from the posterior distribution in Bayesian inference. To be more precise, the paper approaches the question of stochastic sampling that relies only on minibatches of data at each iteration. To achieve rapid mixing between isolated modes, the authors consider parallel tempered chains and introduce replica-exchange steps into the stochastic Nose-Hoover Dynamics. The crux of this approach is the stochastic test for the replica-exchange step. To develop such a test, the authors follow the paper [An efficient minibatch acceptance test for metropolis-hastings], which introduces the concept of correction distribution. The main contribution of the current paper is the derivation of the analytic formula for the correction distribution. ------ UPDATE ------ I've read the rebuttal and other reviews. Unfortunately, I'm pretty confused by the author's response. In what follows, I'm trying to describe my point of view. The authors propose a kind of parallel tempering algorithm, where they interleave stochastic Nose-Hoover dynamics with the replica-exchange step. The stochastic Nose-Hoover dynamics approximately samples from the target distribution without an acceptance test; hence, it can rely only on minibatches of data. To make this also possible for the replica-exchange, the authors employ Seita's paper [arxiv.org/pdf/1610.06848.pdf] that proposes a general-purpose minibatch modification of Barker's test. The only modification to the Seita's test is the deconvolution formula application to obtain the correction distribution, which is also not novel [https://projecteuclid.org/download/pdf_1/euclid.aoms/1177706375]. Thus, I see the proposed algorithm as a straightforward combination of several techniques: Nose-Hoover + parallel tempering + Seita's test. However, just to be clear, I don't want to claim that this is a weakness of the paper. I just think that the proposed technique cannot be seen as a stand-alone theoretical result and requires a thorough empirical study. The empirical study is the main weakness of the paper. There are two main issues: 1. Dealing with minibatch noise. Simply speaking, Seita's idea is to represent logistic random variable X_log from Barker's test as a sum of two other random variables: X_log = X_corr + Normal(0,\sigma^2), where X_corr adds up to the normal minibatch noise to obtain X_log. Note, that the variance of X_log is fixed, and Var(X_log) = Var(X_corr) + Var(noise), since the variables are independent. Therefore, if Var(noise) > Var(X_log) we cannot find suitable correction distribution. To overcome this issue, Seita et al. propose adaptively enlarge minibatches to reduce the noise's variance at the cost of the algorithm's data-efficiency. For instance, it could require thousands of objects on a simple task (classification between only two digits on MNIST) to evaluate the test. This is the central issue in Seita's paper, and I wonder why the authors of the current work completely excluded it from consideration. They also use minibatch enlargement but do not compare their approach to others in terms of data efficiency. Moreover, in their response, they claim that the correction distribution can be obtained with arbitrary precision for any variance of the noise. At this point, I'm pretty confused. Am I missing something? 2. Comparison with competitors. Besides clearly undertrained baselines on CIFAR-10, the paper lacks crucial points of comparison. As I mentioned in my review, the sampling from the posterior distribution of neural networks is tightly related to the ensembling methods [Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles; Snapshot Ensembles: Train 1, get M for free]. Moreover, there is a paper [Cyclical Stochastic Gradient MCMC for Bayesian Deep Learning] that also proposes a stochastic MCMC algorithm, has a similar motivation and has a more thorough empirical study. To sum up, I think although the algorithm sounds promising, the paper lacks proper empirical study.

Strengths: The paper considers the actual problem of sampling from the posterior distribution in the setting of large amounts of data. This problem can be considered as a fundamental approach to inference and is important in machine learning. - The paper's main strength is the analytical derivation of the correction distribution for the stochastic test proposed in [An efficient minibatch acceptance test for metropolis-hastings]. The proposed formula allows for a more fast and more stable evaluation of the correction distribution compared to the original procedure. - Another merit of the paper is the incorporation of the stochastic test into the replica-exchange method. Indeed, the replica-exchange method with tempering allows for fast exploration of the state-space, and, as a result, for more efficient sampling.

Weaknesses: - The lack of discussion on the choice of the bandwidth lambda. Indeed, the bandwidth seems like a crucial parameter of the proposed method. It controls the tradeoff between the accuracy and the computational cost. Some practical guidelines on the choice of lambda would greatly help the reader implement the algorithm and reproduce the results. - lines 242-244. Here the authors describe achieved acceleration compared to the original method by Seita et al. However, this comparison is entirely unclear. For instance, the acceleration is reported "in sampling", when the difference between approaches is in the correction distribution, which can be evaluated preliminary. Considering that this is the only comparison with the original method in the paper, I cannot find this discussion to be sufficient. - Unfortunately, the empirical comparison for the ResNet architecture on CIFAR-10 is not correct. The significant issue here is that the models are clearly undertrained since it is possible to achieve 92% test set accuracy for ResNet-20. Unfortunately, this single fact makes the primary experiment invalid. - There is no description of the inference procedure. MCMC methods are usually applied in deep learning to collect several samples in the weights space and then ensemble predictions over these models. Thus, the authors should compare all the methods for the same number of samples, and describe how they collect these samples (it is too expensive to collect all of them). Moreover, I would expect to see the comparison with competing ensembling approaches [Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles; Snapshot Ensembles: Train 1, get M for free; Cyclical Stochastic Gradient MCMC for Bayesian Deep Learning]. - The paper does not discuss the main weakness of the correction distribution approach: if the minibatch noise is too big, one cannot find a suitable correction distribution. This issue is directly related to the data-efficiency of the method. Since the method could require additional batches to reduce the variance of the energy estimate, at this point, I would expect a proper comparison with other methods. minor comments: - Throughout the paper, the authors refer to the spectrum of distributions and their characteristic functions (line 128, line 168). I think the text would benefit from the precise definition of this notion or informal description. - in the formulation of Theorem 2: what does the phrase "formally preserves" means? - line 207: typo "seems to complicated" -> "seems to be complicated"

Correctness: The derivations in Theorem 2 seem correct, although I have not checked all the formulas. Unfortunately, the empirical evaluation of the method is not correct, as I describe in the weaknesses.

Clarity: Overall, the paper is clearly written, though it lacks a description of some details.

Relation to Prior Work: The paper describes a large body of the prior work. However, I would suggest to add several ensembling techniques for comparison [Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles; Snapshot Ensembles: Train 1, get M for free; Cyclical Stochastic Gradient MCMC for Bayesian Deep Learning].

Reproducibility: No

Additional Feedback: I would advice the authors to check the concurrent work "Non-convex Learning via Replica Exchange Stochastic Gradient MCMC". This work approaches the same problem and propose similar empirical evaluation.


Review 4

Summary and Contributions: This work presents a replica-exchange sampler where replicas with different temperatures are run in parallel and exchanged periodically. The noise introduced by mini-batch convergence is prevented via Nose-Hoover dynamics. --UPDATE-- I have read the author's response and other reviews. I still think combination of several ideals namely, Nose-Hoover, parallel tempering and Seita's test (with deconvolution) makes this submission an interesting work however, I also think some concerns raised by other reviewers are also valid as such I increase my score to 7.

Strengths: Providing an easy-to-implement algorithm that can be used in arbitrary Bayesian inference settings with large data sets is the main novelty of this paper. The theory behind this approach seems sound and reasonably novel and its empirical evaluation shows a significant improvement over the baseline while proposing a general Bayesian framework that handles big data is indeed relevant to the NeurIPS community.

Weaknesses: -- UPDATE -- Following the reviewers discussion, it seems that the other methods are not fine-tuned as such the comparisons may not be fair and valid.

Correctness: Up to my understanding the presented approach is correct and sound.

Clarity: yes

Relation to Prior Work: Some related work to tackle the mentioned two problems of sampling-based Bayesian inference using mini-batches is discussed in the introduction but a more expanded discussion of prior work could add to the value of this submission.

Reproducibility: Yes

Additional Feedback:


Review 5

Summary and Contributions: When a target has multiple well-separated modes, MCMC chains suffer from poor mixing and typically get trapped exploring local modes. At higher temperatures, the MCMC chains can more easily traverse modes. Replica-exchange methods leverage this by running multiple copies of the MCMC chain at different temperatures that progressively swap states to improve mixing. However, these methods all require the evaluation of the likelihood function. In the Bayesian context, when working with large data sets, this becomes computationally infeasible, rendering traditional tempering methods, as well as gradient-based MCMC chains intractable. This paper builds upon Luo et al. (2019) which combined ideas from HMC and simulated tempering to create a continuous-time first-order MCMC algorithm that uses stochastic gradient with mini-batches and used Nose-Hoover dynamics to correct for the noise-induced from the mini-batches versus the full dataset. This paper proposes a replica-exchange version of Nose-Hoover dynamics which improves mixing and sampling efficiency similar to how classical replica exchange methods parallelize simulated tempering. The major contribution of this paper is how to propose a swap between replicas in a setting when only mini-batches are available. Update: After discussions with other reviewers and reading the author(s) feedback, I have decided to change my score to a 6. I think the theoretical contributions are interesting and novel for the machine learning literature. However, I feel that the methodology and experiments still need to be further developed as well as a careful study of its limitations. For it would be nice to know  (1) What is lost by using mini-batches versus full gradient in Nose-Hoover and deconvolution (2) What is the error incurred by approximating reversibility by using the barker versus the metropolis hasting acceptance (3) What are the class of problems this method is suited for and which ones is it not (4) a better benchmarking and comparison with other competing methods. I understand this is beyond the scope of this paper, but I think that is a good first step and demonstrates the potential of this work, however, I think there is still a fair bit of work that needs to be done before this is a practical method for general Bayesian inference in big-data context. 

Strengths: -  This is the first instance of replica-exchange method in the literature using the Barker acceptance probability instead of the Metropolis-Hastings for swapping between replicas     - The main contribution to the literature is Theorem 2. This gives a novel and mathematically interesting approach to proposing swaps between replicas using mini-batches while approximately preserving detailed balance. This is a very interesting addition to the replica-exchange literature.     - The fact that this can be implemented makes it the only replica-exchange method that can account for noise with mini-batches for posterior inference with well-separated modes. - This is a fairly theoretically well-grounded paper. It is clear the author(s) is/are mathematically strong. 

Weaknesses: - iid assumption and the reliance on CLT makes this restrictive for hard Bayesian problems in practice.  This seems like a strong assumption since the constant variance is critical to the building the algorithm. It should be noted the author(s) do identify empirically models built on Gaussian noise, where this assumption is unsurprisingly valid.     - The Barker probability is used to because it is analytic and its regularity properties, not for its Monte Carlo efficiency. It was shown that that the Metropolis-Hastings acceptance is optimal in terms of Peskun ordering (Peskun 1973, Tierney 1998). It is not clear what is lost compared to traditional replica exchange methods compared by using this versus MH or another local proposal satisfying g(t)=tg(1/t) as discussed in Zanella (2017).     - It seems that this algorithm is quite sensitive to hyper-parameters variance $\sigma^*$, $\lambda$, and batch size. Are their general guidelines on how to pick these? It should also be noted that replica exchange methods are known to be very sensitive to hyperparameters (temperature ladder, number of rungs, etc), the design choices for tuning rely heavily on the Metropolis-Hasting acceptance probability, eg. the geometric spacing for the Gaussian. To make this practical for real problems, there needs to be new guidelines for tuning that accounts for the Barker acceptance rule. 

Correctness: - Line 84: Not sure why we can assume the variance of $\tilde{f}$ is independent of the parameters $\theta$. It was not clear where in the reference on line 84 justifies this.     - Each step of the proof of theorem 2 seems correct, although admittedly I did not check every line/equation in detail.          - Was figure 1 generated by the author(s), it seems very similar to figure 8 in Falcioni & Deem (1999).     - Line 248-251: I note sure why adding more Gaussian noise and increasing the variance is done, and how it addresses the potential Ergodicity issues of Nose-Hoover dynamics as states by the author(s)?

Clarity: - I think it is overall well written. It clearly states its purpose and provides a good story. - The math is accessible and does not require anything too obscure to follow.     - My only criticism in terms of style is that the language in the paper is written for someone with intuition from the physics literature, which could be confusing for some readers.

Relation to Prior Work: - This paper does a good job of identifying how it differs from the current literature.     - It makes compelling arguments as to why Nose-hoover dynamics is a worthy competitor to stochastic gradient-based Langevin algorithms.     - The key contribution to this paper in my opinion is the novel approach to replica-exchange in the presence of mini-batch data and is unrelated to the nose-hoover dynamics. The Nose-Hoover dynamics is required as an ingredient for the replica-exchange algorithm, but those details were built in Luo et al. (2018).

Reproducibility: Yes

Additional Feedback: I think the general implementation/experiments can be improved, but in my opinion this paper is not about tuning replica-exchange methods, but rather showing it can be done in the large setting. I think it will take quite a bit of further research before it is practical for general purpose problems.