OptEx: Expediting First-Order Optimization with Approximately Parallelized Iterations

First-order optimization (FOO) algorithms are pivotal in numerous computational domains such as machine learning and signal denoising. However, their application to complex tasks like neural network training often entails significant inefficiencies due to the need for many sequential iterations for convergence. In response, we introduce first-order optimization expedited with approximately parallelized iterations (OptEx), the first framework that enhances the efficiency of FOO by leveraging parallel computing to mitigate its iterative bottleneck. OptEx employs kernelized gradient estimation to make use of gradient history for future gradient prediction, enabling parallelization of iterations -- a strategy once considered impractical because of the inherent iterative dependency in FOO. We provide theoretical guarantees for the reliability of our kernelized gradient estimation and the iteration complexity of SGD-based OptEx, confirming that estimation errors diminish to zero as historical gradients accumulate and that SGD-based OptEx enjoys an effective acceleration rate of $\Omega(\sqrt{N})$ over standard SGD given parallelism of N. We also use extensive empirical studies, including synthetic functions, reinforcement learning tasks, and neural network training across various datasets, to underscore the substantial efficiency improvements achieved by OptEx.

Paper

References (53)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer YWBT7/10 · confidence 3/52024-07-10

Summary

The paper proposes a general framework for accelerating first-order optimization methods. The framework leverages parallel computing by estimating the gradients using kernel methods and breaking iterative dependencies. The paper establishes theoretical guarantees that the method gives an acceleration rate of $\sqrt{N}$ for $N$ parallelism. Then the paper does extensive empirical studies to show the improvements of this framework.

Strengths

1. Very clearly written. It's immediately clear from the abstract and introduction what kind of work the authors are doing and in the main body the details are given in a very pleasant manner with beautiful plots and formula. The whole paper is well organized with theory and practice covered in style. 2. Novelty. I'm quickly convinced that the paper's contribution is novel. The related work session well summarizes previous work and it's clear to me that the paper investigates a direction that is distinctively different and important. 3. A complete story. The paper studies both theoretically and empirically the aspect, with both results agreeing with each other to a certain extent, forming a complete story.

Weaknesses

The theoretical guarantees rely on assumptions that might not work for $N$ very large, thus limit the extent to which the framework can accelerate optimization. In the empirical part, $N$ are not very large, thus unable to show that the framework can accelerate things to great extent in general.

Questions

Suppose that the step size is very small, can the $N$ potentially be very large?

Rating

7

Confidence

3

Soundness

4

Presentation

4

Contribution

3

Limitations

The authors adequately addressed the limitations.

Authorsrebuttal2024-08-08

Dear Reviewer YWBT, Thank you so much for your prompt and positive response after our rebuttal! We are so happy to hear that your concerns have been well addressed. We will incorporate the discussions above into our revised version. Sincerely, Authors

Reviewer 1m6n6/10 · confidence 3/52024-07-10

Summary

The paper "OptEx: Expediting First-Order Optimization with Approximately Parallelized Iterations" introduces a novel framework, OptEx, designed to improve the efficiency of first-order optimization (FOO) algorithms by approximately parallelizing their iterations.

Strengths

1. The introduction of a general framework for parallelizing iterations in FOO is novel and addresses a significant inefficiency in traditional optimization methods. 2. The paper provides robust theoretical guarantees for the performance of the OptEx framework, including bounds on estimation error and iteration complexity. 3. The extensive experiments across various domains (synthetic functions, reinforcement learning, and neural network training) demonstrate the practical applicability and efficiency gains of OptEx.

Weaknesses

1. The kernelized gradient estimation and the associated computational techniques are complex, which might pose challenges for practical implementation and understanding by a broader audience. 2. The efficiency and scalability of OptEx in very large-scale optimization problems or in highly dynamic environments might be limited 3. The theoretical guarantees rely on certain assumptions (e.g., the Gaussian distribution of gradients), which may not hold in all practical scenarios, potentially limiting the generality of the results.

Questions

see weakness

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

NA

Reviewer 1m6n2024-08-10

response to author

Thanks for the detailed response from the author, my concern has already been addressed.

Authorsrebuttal2024-08-10

Dear Reviewer 1m6n, Thank you so much for your positive response after our rebuttal! We are so happy to hear that your concerns have been well addressed. We will incorporate all the discussions above into our revised version. Sincerely, Authors

Reviewer ieTR3/10 · confidence 4/52024-07-18

Summary

This paper presents OptEx, an approach to parallelize optimization methods by using gradients from previous iterations to predict gradient for subsequent iterations which in turn breaks the serial nature of standard stochastic optimization thereby enabling approximately parallel iterations. The gradient prediction is done using Kernel methods. The paper also presents experimental evidence towards the effectiveness of the proposed approach.

Strengths

This appears to be an interesting contribution, albeit lacking clarity in terms of writing.

Weaknesses

-- The question of gradient estimation converging to the true gradient seems fairly far-fetched, and there is every reason to believe modeling gradients would essentially carry a constant bias without very strong regularity and realizability conditions. -- Furthermore, when applied to problems in RL, this becomes even more confounding because of the necessity to explore and using data collected with exploration to in turn construct gradient estimates. I am not sure if I follow how this can be achieved within this framework. -- The paper's experimental comparison doesn't offer any sources of comparisons against the many different approaches to parallelization that have already been studied in the literature, e.g. mini-batching, model averaging etc.

Questions

Can the authors clarify the speedup offered by utilizing their framework while using acceleration/momentum based methods? Standard mini-batch methods offer different trade-offs particularly wrt batch sizes used while still achieving linear parallelization speedups [1]. [1] Cotter et al: Better Mini-Batch Algorithms via Accelerated Gradient Methods.

Rating

3

Confidence

4

Soundness

2

Presentation

2

Contribution

2

Limitations

N/A.

Reviewer RYzW7/10 · confidence 3/52024-07-24

Summary

This paper introduces a new approach for parallelizing stochastic gradient descent for unconstrained, nonconvex, smooth stochastic optimization. The approach is based on building a Gaussian Process surrogate for the true gradient (based on a history of observed stochastic gradients), and uses this surrogate to suggest potential 'future' iterates, by iterating gradient descent using the surrogate. Iterates are then adjusted based on new stochastic gradient estimates, computed in parallel. They present worst-case complexity theory that says that being able to compute $N$ stochastic gradients in parallel improves the iteration complexity by a factor of $\sqrt{N}$, with numerical results on a selection of synthetic problems, reinforcement learning problems, and neural network training problems.

Strengths

1. I think the overall approach is an interesting contribution. I particularly like the discussion about the quality of the GP surrogate for the true function and think these results are interesting (Section 5.1). I think this could have useful links with optimization of genuinely complex functions, similar to those typically considered in Bayesian and zeroth order optimization. 2. The numerical results show some promise for the method, and do indeed show promise for a framework for using parallel stochastic gradient evaluations, where this is practical. 3. The paper is clearly written and easy to follow.

Weaknesses

1. There are gaps in the literature survey of parallelized SGD methods. For example, there is no mention of the highly cited and relevant papers: - Recht et al. Hogwild!: A Lock-Free Approach to Parallelizing Stochastic Gradient Descent. NeurIPS 2011 - Mnih et al. Asynchronous Methods for Deep Reinforcement Learning. ICML 2016 - Yu et al. Parallel Restarted SGD with Faster Convergence and Less Communication. AAAI, 2019 2. It is unclear to me whether the convergence theory (Theorem 2) actually supports the assertion of a $\sqrt{N}$ speedup from $N$ processes. The RHS gradient bound indeed is $O(1/\sqrt{N})$, but you are measuring the smallest gradient over $O(N)$ outer iterations. So, doesn't this say that you decrease your gradient by a factor of $O(1/\sqrt{N})$ after doing $N$ times extra work, which is exactly the convergence rate of SGD under your sub-Gaussian stochastic gradient assumption [33]? The work would also benefit from a direct comparison with just sample averaging of stochastic gradients (see question below). 3. It appears to me that Theorem 3 appears to contradict standard theory (e.g. Bottou, Curtis & Nocedal. Optimization Methods for Large-Scale Machine Learning. SIAM Review, 2018). You claim that the complexity bound for (nonconvex) SGD is tight when applied to the spherical function eq (47), which is strongly convex. However (stochastic) gradient descent with properly chosen stepsizes should converge linearly for this problem, not sublinearly as claimed. Can you explain this potential mismatch in results? 4. For very large problems, it is unrealistic that parallelism will be available, since storing the model and training data may already require parallelism (i.e. a single stochastic gradient evaluation is already computed in parallel). It would be helpful to include a comment about the specific regimes (dimension, dataset size, level of parallelism) where this method is intended to be used.

Questions

How is the next iterate of OptEx defined? Line 10 of Algorithm 1 says it is the final iterate that is chosen, but Figure 1 suggests that the next iterate is somehow taken from a combination of the $\theta_t^1, \ldots, \theta_t^N$ values. Can you please clarify? The latter interpretation would make more sense, otherwise all but one of the parallel evaluations are discarded immediately. Is the $O(1/\sqrt{N})$ speedup exactly the same as what you would get by just doing basic sample averaging? That is, at each iteration, just get $N$ different stochastic gradient estimates at the same point and average them (reducing the variance of the estimates). How does your bound compare with this simple use of parallelism? You say that "$\nabla F$ is assumed to be sampled from the Gaussian Process, $\nabla F \sim GP(0, K(\cdot,\cdot))$". Can you please clarify? I read this to mean that this is the assumed prior of your GP surrogate for grad F, but as written it suggests that your actual objective function has - on average - zero gradient, meaning we would expect the 'average point' to be almost stationary? You say the "Target baseline is equivalent to Algo. 1 with [GP mean] being replaced with [stochastic gradient], indicating the desired parallelized iteration we aim to approximate". Does the 'target' implementation also have the parallel step? If so, we would expect that to give an unfair advantage to 'target', since it does $2N$ stochastic gradient evaluations per outer iteration rather than $N$ for OptEx.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Limitations appear to have been addressed. The theoretical assumptions (e.g. smoothness) are clearly articulated and some practical limitations are addressed in Section 7. No potential negative societal impact of this work.

Authorsrebuttal2024-08-08

Dear Reviewer RYzW, Thank you so much for your positive feedback and thoughtful suggestions! We sincerely appreciate your recognition and support. We are glad to hear that our clarifications are very helpful. We will add our clarifications above in the writing to make our paper clearer. We would like to address your remaining concerns below. --- ### Questions #2 We totally agree with you that sample averaging applied to standard SGD can achieve similar complexity as our OptEx, without the overhead of managing the GP. As we have justified above, these two speedup approaches stem from different principles: - OptEx Speedup: This comes from reduced sequential iterations (first term on the RHS in Equation 45). - Sample Averaging Speedup: This is derived from reduced gradient variance (second term on the RHS in Equation 45). So, when gradient noise $\sigma$ is already decreased to a small value through sample averaging with parallelism $N'$ (not $N$), increasing $N'$ further (i.e., $N'+N$) provides diminishing returns due to the sublinear rate of $O(1/\sqrt{N'+N})$. For example, the benefit of reducing from $1/\sqrt{10}$ to $1/\sqrt{15}$ is less significant compared to reducing from $1/\sqrt{1}$ to $1/\sqrt{6}$ with an additional parallelism $N=5$. However, in this scenario of a small gradient noise $\sigma$, our OptEx can still achieve noticeable speedup (e.g., from $1/\sqrt{1}$ to $1/\sqrt{6}$ with the same additional parallelism $N=5$) by leveraging additional parallel computing to reduce the sequential iterations. This reduction in sequential iterations is however unattainable by mini-batch SGD itself. In light of this, Equation 45, leading to our Theorem 2, in fact, will provide valuable insights into when OptEx is preferable to simple sample averaging. Overall, OptEx complements rather than replaces existing parallelization methods, including sample averaging, to accelerate first-order optimization. This is particularly effective when other methods are **inapplicable** or **underperforming**, i.e., further noticeable improvements cannot be achieved merely by increasing parallel computing resources in those methods, as we have justified above. ### Weaknesses #2 We apologize for the confusion caused by our notation $\tau$. Our intention was to use $\tau$ to denote all gradients evaluated during the optimization process more easily, including those in parallel processes, as mentioned in line 260. In Theorems 2 and 3, the expression $\min_{\tau \leq NT} \left|\nabla F(\theta_{\tau})\right|$ is exactly meant to represent $\min_{t \leq T, s \leq N} \left|\nabla F(\theta_{t,s})\right|$. Thank you for bringing this to our attention. We will correct this notation in our revised paper. --- We want to thank Reviewer RYzW for your positive feedback and recognition again! We sincerely hope our response has addressed your remaining concerns and can increase your opinions of our work. If you have any other questions, we would like to address them. Sincerely, Authors

Reviewer RYzW2024-08-09

Thanks for this information. Regarding question #2, I think this is a very important aspect and warrants a discussion in any new version of the paper. However, I don't think the benefit is as good as suggested: if you want to do $N_1$ parallel samples for each stochastic gradient (to reduce the variance to $\sigma^2/N_1$), then if you run OptEx with $N_2$ parallel runs, then you will need $N_1\times N_2$ stochastic gradients in parallel at each outer loop, not $N_1+N_2$, so your speedup is still $O(1/sqrt{N})$ where $N=N_1\times N_2$ is the number of cores needed. This has been a helpful discussion, and pending the clarifications in a revised version as agreed by the authors above, I am happy to upgrade my scores of the paper in the initial review above.

Authorsrebuttal2024-08-09

Dear Reviewer RYzW, Thank you so much for your prompt and thoughtful response. We are pleased to hear that our clarifications have improved your opinion of our work, and we will incorporate all the clarifications and discussions mentioned above into our revised manuscript, as per your suggestion. We also appreciate your valuable insight regarding the parallel computing required by combining sample averaging and OptEx, in relation to your question #2. We agree with you that in this scenario, we will need $N=N_1 \times N_2$ parallel computations, leading to a speedup of $O(1/\sqrt{N})$. The benefits of our OptEx over sample averaging may be more straightforward by looking into our Equation 45, which mirrors the convergence results of sample averaging. Particularly, when the gradient noise $\sigma$ is already decreased to a small value, i.e., the second term on the RHS in Equation 45 is small, the first term on the RHS in Equation 45 that is related to sequential iterations will then dominate the convergence. Consequently, simply applying sample averaging with increased parallel computational resources yields only marginal improvements in convergence. In contrast, OptEx can achieve a noticeable improvement by reducing the first term on the RHS in Equation 45 by $O(1/N)$. We will also include this discussion in our revised paper. Should you have any other questions, we would like to address them. Sincerely, Authors

Reviewer YWBT2024-08-07

Very Good Rebuttal

Thanks the authors for their well-written rebuttal. The response on my confusion about N is very clear. Thanks a lot.

Reviewer RYzW2024-08-08

Thank you for these clarifications, they are very helpful. Pending seeing a new version of the manuscript, I think the responses given for all points above would be very helpful. I have one remaining concern (question Q2) and one point for clarification (weakness Q2) that I still think warrant more consideration. **Questions #2** I understand that you would like to potentially use OptEx in the exact gradient case ($\sigma=0$, e.g. Section 6.1), and your results make sense in this context. However, in the SGD context that seems to be your main focus, if I had the ability to compute $N$ stochastic gradients in parallel (each with a given noise level $\sigma$ as you assume), I could either use OptEx or compute $N$ stochastic gradients at the current iterate and average these, giving an effective variance of $\sigma^2/N$. In Theorem 1, your complexity bound (when $\sigma>0$) is $O(\sigma/\sqrt{NT})$ for $T$ iterations, which suggests that sample averaging applied to standard SGD (where $\sigma \mapsto \sigma/\sqrt{N}$) would achieve the same complexity as OptEx with essentially no overhead from managing the GP, etc. It would be useful if you could either discuss why/when OptEx should be preferred to simple sample averaging, as I'm not sure if your theoretical results show this. See, for example, the discussion "Trade-Offs of (Mini-)Batching" in Section 4 of Bottou, Curtis & Nocedal [your ref 26]. **Weaknesses #2** I think my confusion arose because your statement of Theorems 2 and 3 are not quite correct. Specifically, the LHS term in eqs (8) and (9) should be $\min_{\tau\leq T, s\leq N} \|\nabla F(\theta_{\tau,s})\|^2$ rather than $\min_{\tau\leq NT} \|\nabla F(\theta_{\tau})\|^2$ as given (i.e. reflecting that you do $T$ outer iterations with $N$ parallel evaluations in each, rather than $NT$ outer iterations with $N$ parallel evaluations in each, as currently stated). This seems to align with what is shown in the proofs.

Authorsrebuttal2024-08-12

Dear Reviewer ieTR, Thank you so much for taking the time to review our paper and for your insightful questions. While we have thoroughly addressed the concerns and questions raised by all the other reviewers, we sincerely hope that our clarifications and discussions above have effectively addressed your concerns as well. If you have any more questions or need more details, we are happy to answer them promptly within the discussion period. Best, Authors

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC