Searching for Optimal Per-Coordinate Step-sizes with Multidimensional Backtracking

The backtracking line-search is an effective technique to automatically tune the step-size in smooth optimization. It guarantees similar performance to using the theoretically optimal step-size. Many approaches have been developed to instead tune per-coordinate step-sizes, also known as diagonal preconditioners, but none of the existing methods are provably competitive with the optimal per-coordinate stepsizes. We propose multidimensional backtracking, an extension of the backtracking line-search to find good diagonal preconditioners for smooth convex problems. Our key insight is that the gradient with respect to the step-sizes, also known as hypergradients, yields separating hyperplanes that let us search for good preconditioners using cutting-plane methods. As black-box cutting-plane approaches like the ellipsoid method are computationally prohibitive, we develop an efficient algorithm tailored to our setting. Multidimensional backtracking is provably competitive with the best diagonal preconditioner and requires no manual tuning.

Paper

Similar papers

Peer review

Reviewer eKZX5/10 · confidence 1/52023-07-06

Summary

The authors suggest incremental updates of $\mathbf{x}$ for finding the minimum of strongly-convex function $f$ that guarantee decreasing $f(\mathbf{x}_{t})-f(\mathbf{x}_\ast)$ based on only 1st-order gradient information. Their idea is in each step, - choose a candidate matrix $\mathbf{P}_t$ based on set $\mathcal{S}_t$, and - check the condition (4), that guarantee sufficient decrease of the $f$, and - if the condition is satisfied: - apply update $\mathbf{x}_{t+1} = \mathbf{x}_t - \mathbf{P}_t\nabla f(\mathbf{x}_t)$ - $\mathcal{S}_{t+1} = \mathcal{S}_t$ - else - $\mathbf{x}_{t+1} = \mathbf{x}_t$ - update $\mathcal{S}_{t+1} = \text{cut}(\mathcal{S}_t, \mathbf{x}_t, \mathbf{P}_t)$ They provide proofs for - approaching the optimal in Proposition 3.2 - how to choose candidate and cut algorithm in Theorem 5.3, and its maximum number of calls. They conduct some simple experiments in Section 6, and show the proposed algorithm's efficiency.

Strengths

originality - considering backtracking using preconditioned matrix $\mathbf{P}_t$ would be novel idea. But I'm not an expert of this field, and not so sure on the originality. quality - The proposed algorithm is supported by some proofs, and it shows good empirical results. clarity - Basically, the manuscript is readable. significance - The proposed algorithm seems be better than other baselines except for Diag. Hessian+LS, which uses information of 2nd order derivatives, i.e. Hessian, even the proposed algorithm uses 1st-order derivatives of $f$. It would be significant.

Weaknesses

- The manuscript contains some typos. - I have a concern on Figure 5. The horizontal axis shows number of f/grad evals, but I guess the number of CUT calls should be also taken into account.

Questions

- In line 129, the representation $\nabla^2 f$ appears. Does it mean Hessian? or Laplacian? - Between line 197 and 198, the final inequality may be not $\overset{(3)}{\leq}$ but $\overset{(4)}{\leq}$ ? - In the caption of Figure 3(b), $\mathcal{H}\_{>}(\mathbf{u})$ should be replaced by $\mathcal{H}_>(\mathbf{v})$ ? - In Figure 5, do the authors take the number of CUT into account?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

They address the limitation, their method is only supported for convex deterministic setting.

Reviewer hZtF8/10 · confidence 4/52023-07-09

Summary

This paper provides a backtracking approach for smooth convex optimization on a per-coordinate basis with a theoretical analysis that show the gain with respect to classical backtracking line-search and that compare to the optimal per-coordinate conditioners.

Strengths

This paper is super well written and organized. The contribution is also significant as it is a building block of many problems in machine learning. In general, further improving the "adaptivity" of optimization algorithms is essential to seamlessly apply theoretical results (i.e., optimal per coordinate step sizes) to operational purposes.

Weaknesses

The only drawback might be focusing on smooth and strongly convex problems, but it is still a significant first step.

Questions

None.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

None.

Reviewer 5i176/10 · confidence 1/52023-07-15

Summary

This paper extends backtracking to multi-dimension. The authors propose a cutting plane method to find optimal per-coordinate step-sizes (in other words, to find an optimal preconditioner) for smooth convex optimisation. Experiments on ill-conditioned logistic regression problems show that the proposed algorithm can find good preconditioner and improve over vanilla gradient descent.

Strengths

This paper fills a potential gap in the optimization literature by proposing multidimensional backtracking. The proposed method is technically sound and seems to work well in practice.

Weaknesses

I do not see any major issues with the paper, except maybe that it is a bit hard to follow and understand (even though the English is good). Maybe because I don't have enough background on the topic. I'm really sorry for the short review.

Questions

None.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

See weaknesses.

Reviewer JVBS8/10 · confidence 4/52023-07-21

Summary

This paper presents a generalized backtracking line-search method, which estimates coordinate-wise stepsizes referred to as 'preconditioner' of gradient descent. Stemmed from the observation that any existing methods do not exceed the performance of backtracking line-search method, this paper designs a generalized backtracking line-search technique which is realized as a cutting plane method, whose separating hyperplane comes from the hypergradient, i.e., gradient with respect to hyperparameter of the algorithm, which is a stepsize in this case. Followed by the worst-case convergence analysis for smooth strongly convex function, the writers also provide experimental results illustrating the competitiveness of this method for ill-conditioned problems and robustness among problem classes.

Strengths

Section 4 contains the key insight of this work: that a failed preconditioner (defined as one that violates a Armijo-type condition) provides a cutting plane on the set of valid preconditioners. This is a very nice idea that is, as far as I know, novel, and I expect this work to lead to a lot of follow-up work. This is a new type of result and I think it is valuable.

Weaknesses

.

Questions

The 'CUT' subroutine, which is a 'backtracking' phase of this algorithm, can be called up to the number of iterations linear in dimension $d$, which can be large in considering large-scale problems. It has been illustrated in the experimental result that for large problems it recovers preconditioner quite fast, but I'm also curious on how much total overhead is caused by the subroutine `CUT'. (p.3 line 121) It seems there is a typo on notation regarding $d, n, \\alpha$.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC