Adaptive Variance Reduction for Stochastic Optimization under Weaker Assumptions

This paper explores adaptive variance reduction methods for stochastic optimization based on the STORM technique. Existing adaptive extensions of STORM rely on strong assumptions like bounded gradients and bounded function values, or suffer an additional $\mathcal{O}(\log T)$ term in the convergence rate. To address these limitations, we introduce a novel adaptive STORM method that achieves an optimal convergence rate of $\mathcal{O}(T^{-1/3})$ for non-convex functions with our newly designed learning rate strategy. Compared with existing approaches, our method requires weaker assumptions and attains the optimal convergence rate without the additional $\mathcal{O}(\log T)$ term. We also extend the proposed technique to stochastic compositional optimization, obtaining the same optimal rate of $\mathcal{O}(T^{-1/3})$. Furthermore, we investigate the non-convex finite-sum problem and develop another innovative adaptive variance reduction method that achieves an optimal convergence rate of $\mathcal{O}(n^{1/4} T^{-1/2} )$, where $n$ represents the number of component functions. Numerical experiments across various tasks validate the effectiveness of our method.

Paper

References (45)

Scroll for more · 33 remaining

Similar papers

Peer review

Reviewer 4Ggd7/10 · confidence 4/52024-07-10

Summary

This paper introduces novel adaptive variance reduction methods for stochastic optimization, building on the STORM technique. The proposed Ada-STORM method closes the $O(\log T)$ gap and achieves an optimal convergence rate of $O(T^{-1/3})$ for non-convex functions under assumptions weaker than previous approaches. Furthermore, the paper extends the method to stochastic compositional optimization, maintaining a similar $O(T^{-1/3})$ convergence rate. For non-convex finite-sum problems, the authors develop another innovative adaptive algorithm that attains the optimal $O(n^{1/4}T^{-1/2})$ rate. Numerical experiments effectively demonstrate the effectiveness of the proposed method.

Strengths

1. The proposed Ada-STORM method overcomes major limitations in existing adaptive STORM methods. It does not require bounded gradient and function value assumptions, and achieves the optimal convergence rates without the additional $O(\log T)$ term. The problem investigated in the paper is challenging and the results are substantial. 2. The theoretical analysis is novel and easy to follow. Extending the method to stochastic compositional and finite-sum optimization problems illustrates its flexibility and potential impact across a wide range of optimization tasks. 3. The numerical experiments on various tasks (e.g., image classification, language modeling) validate the theoretical results and highlight the method's superior performance compared to other algorithms.

Weaknesses

1. In Theorem 4, while the authors clearly state the convergence rate of the proposed method in terms of $T$, $\Delta_F$, and $L$, they only present the convergence of previous adaptive finite-sum methods in terms of $T$. A comparison of other important constants such as $\Delta_F$ and $L$ with existing methods could enhance the discussion. 2. The content in Lines 240 to 242, especially Algorithm 4, should be included in the main body of the paper rather than being delayed to the Appendix. 3. Some typos in the paper: - Line 145: $(t>2)$ should be $(t \geq 2)$. - Line 516, the first equality in Proof 7: $L^2$ should be $L$.

Questions

1. Regarding Line 401, could the authors explain why the first inequality holds? 2. In Theorem 4, is the dependency on $\Delta_F$ and $L$ better or worse than that of previous methods?

Rating

7

Confidence

4

Soundness

4

Presentation

4

Contribution

4

Limitations

This paper does not present any negative societal impacts.

Reviewer 22397/10 · confidence 2/52024-07-10

Summary

This paper proposes a novel adaptive STORM method that achieves an optimal convergence rate of $O(T^{-3})$ for nonconvex stochastic optimization, which requires weaker assumptions and attains the optimal convergence rate without the additional $O(\log T)$ term.

Strengths

For stochastic non-convex optimization, this paper achieves the optimal convergence rate under more relaxed assumptions, which does not require the bounded function values or the bounded gradients and does not include the additional $O(\log T)$ term in the convergence rate. The proposed technique has also been extended to stochastic compositional optimization with the same optimal rate. For non-convex finite-sum optimization, this papaer further improve the adaptive algorithm to attain an optimal convergence rate, which outperforms the previous result by $O(\log(nT))$ factor.

Weaknesses

Although this paper is theoretically sound in general, there are still some questions need to be discussed: 1. The experimental part is relatively limited. The proposed algorithm is parameter free, so it is best to provide some experimental results to demonstrate whether these compared algorithms are sensitive to parameter changes. 2. In theory, the output {x_ \tau} of the proposed algorithms is randomly selected from {1,..,T}. Is the output in the experiment randomly selected?

Questions

1. The experimental part is relatively limited. The proposed algorithm is parameter free, so it is best to provide some experimental results to demonstrate whether these compared algorithms are sensitive to parameter changes. 2. In theory, the output {x_ \tau} of the proposed algorithms is randomly selected from {1,..,T}. Is the output in the experiment randomly selected?

Rating

7

Confidence

2

Soundness

3

Presentation

3

Contribution

4

Limitations

The authors discussed the limitations and there is no negative societal impact about this work.

Reviewer qB9i7/10 · confidence 3/52024-07-10

Summary

This paper studies non-convex stochastic optimization under the assumption of mean-square smoothness. It introduces, Ada-STORM, a variant of the STORM algorithm, which achieves the optimal rate $O(T^{-1/3})$. Unlike vanilla STORM, Ada-STORM eliminates the $O(\log T)$ factor and does not require the Lipschitz assumption. The algorithm's success hinges on a carefully designed adaptive learning rate schedule. Instead of the standard AdaGrad learning rate $\eta_t \approx (\sum_{i=1}^t \\|g_i\\|^2)^{-1/3}$ used in STORM, Ada-STORM employs $\eta_t = \min\\{T^{-1/3}, T^{-(1-\alpha)/3}(\sum_{i=1}^t \\|v_i\\|^2)^{-\alpha}\\}$ where $\alpha < 1/3$ and $v_t$ is the STORM update.

Strengths

The proposed algorithm, Ada-STORM, achieves the optimal rate $O(T^{-1/3})$ without the log factor. Unlike STORM+ which also eliminates the log factor, this algorithm does not assume $f$ to be Lipschitz and bounded above. This improvement represents a significant contribution. Moreover, the required modification is straightforward: it involves merely adjusting the learning rate scheduler $\eta_t$ and the momentum factor $\beta_t$. I particularly enjoy the constant momentum constant $\beta_t\equiv T^{-2/3}$ rather than time-varying and dependent on $\eta_t$. In addition, the proposed learning rate scheduler is of interest in its own right. The paper extends the application of this learning rate technique to related problems, achieving optimal rates in composite optimization and finite-sum optimization (ERM). Finally, empirical experiments also validate Ada-STORM's performance, which matches or outperforms existing STORM variants and other state-of-the-art optimizers.

Weaknesses

One major concern pertains to the technical correctness of the proof of Theorem 1, in particular from line 440-444. I don't immediately see how $\mathbb{E}[\sum_{t=1}^{s-1} \\|v_t\\|^2] \le C_0T^{1/3}$ implies $\mathbb{E}[\frac{1}{T}\sum_{t=1}^{s-1}\\|v_t\\|] \le \sqrt{C_0}T^{-1/3}$, and similarly how the equation in line 441 implies line 442. My concern arise from the fact that $s$ is a random variable dependent on the entire history $\mathcal{H}_T$. As a result, applying Jensen's inequality like $\mathbb{E}[\frac{1}{T}\sum\_{t=1}^s\\|v_t\\|] \le \sqrt{\mathbb{E}[\frac{1}{T}\sum\_{t=1}^s\\|v_t\\|^2]}$ seems to be inappropriate when $s$ is a random variable dependent $v_t$'s. It's likely that I miss something, and I encourage the authors to elaborate this part of the proof to make it clear. Another concern is about the doubling trick. In each stage $k$, the guarantee is $$\frac{1}{2^{k-1}} \sum_{t=2^{k-1}}^{2^k} \mathbb{E}\\|\nabla f(x_t)\\| \lesssim \frac{\Delta_f^{3/4}+O(1)}{(2^{k-1})^{1/3}}.$$ Since the algorithm is not resetting $x_t=x_0$ in the new stage, the parameter $\Delta_f$ is not constant. In particular, in stage $k$, $\Delta_f = \mathbb{E}[f(x_{2^{k-1}}) - \inf f(x)]$. Consequently, the doubling trick implicitly assumes $f$ has bounded function value in order to bound $\Delta_f$ in all stages. If the authors can address my major concerns, I am inclined to revise my score upwards, given the otherwise strong results of the paper.

Questions

- Is there a specific reason why the authors prefer $\alpha$ as large as possible in the trade-off of $\Delta^{1/2(1-\alpha)} + \sigma^{1/(1-\alpha)} + L^{1/2\alpha}$. Is it due to the assumption (or some empirical observation) that the smoothness constant $L$ is usually the dominating parameter? - In the experiments, did the authors apply any learning rate scheduler (e.g., linear decay or cosine decay) on top of the adaptive learning rate? Although somewhat tangential, it would be interesting to compare the new adaptive scheduler with other popular schedulers.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A

Reviewer G4tN6/10 · confidence 3/52024-07-15

Summary

This paper studies adaptive variants of STORM, a variance reduction technique proposed by Cutkosky and Orabona (2019), for nonconvex stochastic minimization problems. Through introducing a novel adaptive parameter and step size tuning method, the authors aim to remove the bounded gradients and bounded function values assumptions in existing literatures while achieving the optimal convergence rate without an additional penalty of $O(\log T)$ term due to adaptiveness. Based on the same techniques, they further propose Compositional STORM for compositional functions and STORM for finite-sum problems. They support their theoretical claims with numerical experiments on image classification and small-scale language modelling tasks.

Strengths

1. Their proposed method can achieve optimal convergence rate of $O(T^{-1/3})$ under more relaxed assumptions, namely by removing the need for bounded function values and bounded gradients. Their method does not incur a $\log(T)$ term penalty like in other adaptive methods (Liu et al., 2022). 2. They show that their learning rate design and analysis technique can be extended to stochastic compositional problems under further assumptions, namely Lipschitz continuity. 3. Although rather unsurprising, they show that when the finite-sum structure is present, their method can be adapted to and showed improved convergence rates of a log factor over existing methods.

Weaknesses

1. My understanding of the original non-adaptive STORM proposed by Cutkosky and Orabona (2019) does not have a log factor in their convergence rate when noise is present. Specifically their convergence rates in expectation is $O(\log T / T^{1/2} + \sigma^{1/3} / T^{1/3})$, thus the $\log T$ term is dominated and should not appear in your big-O convergence rates in many places of the paper (e.g., Line 28, and Table 1). Can you please explain and check your comparisons? 2. I am not sure about the significance of removing the bounded gradients and bounded function values assumption in practice. Cutkosky and Orabona (2019) proposed an adaptive $G_t$ approach to remove the bounded gradients assumption in their original paper, while bounded function values seems to be a very mild assumption under smoothness conditions when the final bounds all depend on the initial function value gap. Thus, I do not agree with the authors claiming them to be "strong assumptions". It would be useful if the authors can provide examples to justify why these assumptions can be problematic in practice. 3. The authors should be clearer that their extension to stochastic compositional optimization is not exactly under weaker assumptions. They require standard Lipschitz continuity assumptions which have been standard in the literature, which in essence is same as bounded gradients. Overall I think this paper has potential but may benefit from addressing the above points.

Questions

See weaknesses.

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

2

Limitations

Yes.

Reviewer G4tN2024-08-11

Thank you for detailed explanation, and upon further checking, I agree with the authors' claims regarding my W1 and W2. As such, I am increasing my score from 5 to 6.

Reviewer qB9i2024-08-08

Thank the authors for the response. Regarding the technical details in Theorem 1, sorry for not making my point clear initially. I wasn't questioning the validity of Jensen's inequality. Rather, I was unclear about how to apply Jensen's inequality properly in this context. I appreciate that the authors have now clarified this. Regarding the doubling trick, the authors resolved my misunderstanding. Please clarify this in later drafts. Regarding the additional experiments, I thank the authors for the extra efforts. Although tested on the rather simple dataset CIFAR-10, the additional results that the new adaptive rate still outperforms Storm with other learning rate schedulers (e.g. linear decay, cosine decay) strongly supports the validity of Ada-Storm. If time is allowed, I'd suggest the authors to repeat the experiments on larger tasks such as LLMs in the future version, and it will be an encouraging result if Ada-Storm also outperforms other optimizers on more complicated tasks. Overall, the authors have addressed my major concerns. Given its concrete theoretical results and encouraging empirical results, I revised my score accordingly.

Reviewer 4Ggd2024-08-09

Thank the authors for the rely. It addresses all my questions and concerns. I would keep my score.

Reviewer 22392024-08-10

Thanks for the response and I have no further questions.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC