Gradient-Variation Online Learning under Generalized Smoothness

Gradient-variation online learning aims to achieve regret guarantees that scale with variations in the gradients of online functions, which has been shown to be crucial for attaining fast convergence in games and robustness in stochastic optimization, hence receiving increased attention. Existing results often require the smoothness condition by imposing a fixed bound on gradient Lipschitzness, which may be unrealistic in practice. Recent efforts in neural network optimization suggest a generalized smoothness condition, allowing smoothness to correlate with gradient norms. In this paper, we systematically study gradient-variation online learning under generalized smoothness. We extend the classic optimistic mirror descent algorithm to derive gradient-variation regret by analyzing stability over the optimization trajectory and exploiting smoothness locally. Then, we explore universal online learning, designing a single algorithm with the optimal gradient-variation regrets for convex and strongly convex functions simultaneously, without requiring prior knowledge of curvature. This algorithm adopts a two-layer structure with a meta-algorithm running over a group of base-learners. To ensure favorable guarantees, we design a new Lipschitz-adaptive meta-algorithm, capable of handling potentially unbounded gradients while ensuring a second-order bound to effectively ensemble the base-learners. Finally, we provide the applications for fast-rate convergence in games and stochastic extended adversarial optimization.

Paper

References (61)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer NEvk6/10 · confidence 3/52024-07-07

Summary

Problem: The paper studies the OCO problem under the generalized smoothness assumption, i.e. at each time $t$, $f_{t}$ satisfies $\|| \nabla ^ {2} f_{t} (x) \|| \le \ell_{t}(\|| \nabla f_{t}(x) \||)$ for all $x \in \mathcal{X}$, where $\ell_{t}$ is a positive non-decreasing function. In addition to the information model in OCO, the algorithm can query $\ell_{t}$ at any local point $x \in \mathcal{X}$. This condition subsumes global smoothness, i.e. $\ell \le L$ is upper bounded globally, and $(L_{0},L_{1})$ smoothness, i.e. $\ell(x) = L_0 + L_1 x$; which have been studied in the prior works. The authors obtain gradient-based variation bounds for several settings of interest (see contributions below). Contributions: 1) First, the authors extend the classic optimistic OMD algorithm to handle generalized smoothness and obtain an optimal $\mathcal{O}(\log V_{T})$ rate for strongly convex and $\mathcal{O}(\sqrt{V_{T}})$ static regret for convex functions (ref. Theorem 1, 2). However, these rates are first derived assuming that the algorithm knows the curvature of the $f_{t}$'s, i.e. whether all the $f_{t}$'s are convex or strongly convex. 2) To circumvent the curvature information, the authors devise an algorithm that is adaptive to the curvature of the functions $f_{t}$'s (referred to as universal online learning -- the goal is to obtain an algorithm with simultaneous guarantees for convex and strongly convex functions). The proposed algorithm obtains the optimal $\mathcal{O}(\sqrt{V_T}), \mathcal{O}(\log V_T)$ static regret bounds for convex, strongly convex functions respectively (ref. Theorem 4). 3) Finally, the authors consider stronger regret measures, i.e. interval regret and dynamic regret (referred to as non-stationary online learning), and propose an algorithm with the optimal gradient variation-based regret guarantees (ref. Theorem 5, 6).

Strengths

The paper is solid and shows that gradient variation-based bounds can be derived under a more general smoothness assumption. Prior works focused on global smoothness, or $(L_{0}, L_{1})$ smoothness, and did not consider some or the other regret metrics as considered in the paper, e.g. the authors mention that an adaptive regret guarantee was not obtained even in the context of global smoothness. While this might seem doable, the authors very well mention the key challenges in obtaining the desired algorithms, e.g. for contribution (2) above, section 3.2 does a great job of highlighting the key challenges. I liked this section and the authors spent considerable effort explaining why existing meta algorithms, e.g. McMwC-Master algorithm (Chen et al.) do not directly work since they cannot handle heterogeneous inputs. The function-to-gradient variation technique is simple and seems pretty useful since it allows us to decouple the meta-algorithm and the base learners and explicitly aim toward obtaining a meta-algorithm with the desired guarantee. It helps avoid the seemingly messier cancellation-based tricks as done by Yan et al.

Weaknesses

1) Section 4 is presented in a rush. Like section 3.1, it would be quite beneficial to have some more discussion on the algorithm and potential challenges in deriving non-stationary regret guarantees. The authors claim that their proposed algorithm is the first to obtain gradient variation-based non-stationary regret guarantees under the generalized smoothness assumption, whereas prior works did not obtain similar guarantees under the global smoothness assumption. The authors should have at least backed this with the potential difficulties faced by existing approaches. I recommend that the authors cut down some earlier discussions, such as the discussion about exp-concave functions, which in my opinion is not very relevant, given that the entire paper is about convex and strongly convex functions. 2) While the paper shows interesting results, I feel the key ideas are borrowed from existing works. While this is fine, in some places the authors failed to mention the potential technical challenges in the analysis. This is the most applicable to section 3. The authors mention that the proposed algorithm can be analyzed similarly to the prod-style algorithms. Are there technical difficulties in directly incorporating the analysis of prod-style algorithms into Algorithm 1? Some minor typos that I found: 1) Line 38: "stochatsic -> stochastic" 2) Lemma 1: Remove the $\ell_t$

Questions

1) The discussion towards the end of page 3 is confusing to me. The authors mention that assuming the Lipschitzness of $f_{t}$ is not reasonable since the bound on $\|| \nabla ^ {2} f(x) \||$ can be obtained from knowing $\ell$ directly, and the function is globally smooth. However, the bounds presented in the Theorems are a function of the Lipschitzness parameters. I would appreciate clarification from the authors. 2) In the Equation after Line 384, $\xi_{t, i}$ is not necessarily same for $f_t$ and $f_{t - 1}$ since the Mean Value Theorem only guarantees the existence of a $\xi$. With $\xi_t$ and $\xi_{t-1}$ being possibly different, I don't think the next inequality follows immediately. Is that a typo or am I missing something?

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

The authors adequately addressed the limitations.

Reviewer ian36/10 · confidence 2/52024-07-12

Summary

The paper aims to provide gradient-variation regret bound when only a generalized smoothness condition holds. It considers the optimistic mirror descent algorithm. It further designed a meta-algorithm, which can be Lipscthiz adaptive. Lastly, it considers adaptive regret and dynamic regret.

Strengths

(1) Compared with existing literature, the work relaxed the smoothness assumption. (2) The meta-algorithm to achieve universality seems novel to me.

Weaknesses

The clarity and effectiveness of the paper could be improved by condensing the extensive paragraphs that discuss "challenges." Instead, it would be beneficial to expand the remarks on how the findings compare with existing results following the theorems. This would provide readers with a clearer understanding of the paper's contributions to the current body of knowledge.

Questions

1. The paper addresses the generalized smoothness condition and frequently mentions the potential for unbounded gradients and smoothness. However, Assumption 2 posits a bounded domain. Does this imply that the gradients and smoothness are also bounded? In reference [7], such boundedness does not appear to be a requirement. 2. On page 5, line 203, the paper asserts that the complexity \(O(D\sqrt{V_T} + LD^2)\) aligns with the optimal constant dependency. It would be helpful if the authors could specify where exactly in reference [7] this lower bound is discussed.

Rating

6

Confidence

2

Soundness

3

Presentation

2

Contribution

3

Limitations

N/A

Reviewer VaQR7/10 · confidence 4/52024-07-15

Summary

This paper studies adaptive online learning under the generalized smoothness assumption. The authors proposed optimistic OMD based algorithms that achieves the first gradient variation bounds under this general setting. Under this assumption, they also provide uninveral algorithms which can adapt to different function types, and also algorithms with dynamic regret.

Strengths

1. This paper is the first to study adaptive online learning under generalized smoothness, which is a much milder condition compared to the standard global smoothness that considered by previous work. 2. In section 2, the authors provide optimistic OMD based algorithms with adaptive step size, and prove that it enjoys O(sqrt{V_T}) and O(log V_T/\lambda) regret bounds for convex and strongly convex functions. To my knowledge, these are the first adaptive bounds under the generalized smoothness condition, which is a novel and interesting contribution. 3. The authors also provide a universal algorithm for achieving gradient variation bounds under the generalized smoothness setting, which can adapt to convex and strongly convex functions, without knowing the type of loss functions in advance. Existing universal methods cannot be directly applied due to the unknown Lipschitzness parameters, and the authors address it by using novel techniques such as function-variation-to-gradient-variation conversion. 4. The authors also showed that the proposed methods can be applied to get adaptive bounds in stochastic extended adversarial online learning and the fast-rate games under more general assumptions.

Weaknesses

1. The base learner proposed in Section 2 needs to know the Smooth constant of the previous observed loss functions (at round t-1), which might be difficult to compute if the loss functions are complicated. 2. The proposed universal algorithm requires to query the gradient O(\log T) times per iterations, while most of the existing universal methods only need to query the gradient once. 3. Previous universal algorithms with gradient variation bound can deal with exp-concave functions, while the proposed algorithm can only deal with convex and strongly convex functions. 4. Previous work can also achieve small-loss bound in this setting. It would be great to understand if small-loss bound is achievable under this more general assumption. Minor comments: Line 56: sqrt{V}_T: T should be inside the square root; Line 140: “a finte bound but unknown L” is not assumed in this paper Line 197: it is more common to use equality for the big-O notation.

Questions

Please see weakness

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes

Reviewer YPdL2024-08-07

Thank you for the detailed response! My main concerns were addressed well so I have raised my score. **A1:** Makes sense; I think it would be helpful to the reader if you added more specifics as to which discussion you're referring to, since they might not be familiar with that paper

Authorsrebuttal2024-08-10

Thank you for your comment! We will improve the presentation to make this point clearer.

Reviewer VaQR2024-08-09

Response

Thank you for the detailed response. I have no further questions at this time. After reading the other reviews, I agree with Reviewer NEvk that Section 4 appears to have been written hastily. It is only 1/3 of a page but contains too much information/results. Additionally, the nonstationary part of the paper is not closely related to the universal results, which are the primary focus of the paper. It would be great if the authors could consider moving this section to the appendix or a journal extension.

Authorsrebuttal2024-08-10

Re: Response

Thank you for your comments! We will work on improving the clarity of the paper presentation to ensure that the key ideas and main results are conveyed more clearly.

Reviewer NEvk2024-08-10

Response to the Author Rebuttal

I thank the authors for their reply and appreciate their effort in the rebuttal. Response to A1: This makes sense to me now. I agree with the assumption that there exists a constant such that $||\nabla \ell_{t}(w)|| \le G$ for all $w \in \mathcal{W}$, is quite common in the online learning literature. Response to A2: Thanks for the clarification. Response to A3 + A4: I am convinced that the authors do propose some interesting techniques to get around the difficulties faced by existing works. Based on the author's responses, I have now increased my score to 6. However, please make sure to reorganize section 4 in the subsequent revisions of the paper.

Authorsrebuttal2024-08-10

Thanks for your helpful comments! We will carefully revise the paper to ensure that key ideas and main results are clearly and properly presented in the final version.

Reviewer ian32024-08-12

Thank you for the authors' reply. I do not have further questions.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC