Saddle-to-Saddle Dynamics in Diagonal Linear Networks

In this paper we fully describe the trajectory of gradient flow over two-layer diagonal linear networks for the regression setting in the limit of vanishing initialisation. We show that the limiting flow successively jumps from a saddle of the training loss to another until reaching the minimum ℓ1-norm solution. We explicitly characterise the visited saddles as well as the jump times through a recursive algorithm reminiscent of the LARS algorithm used for computing the Lasso path. Starting from the zero vector, coordinates are successively activated until the minimum ℓ1-norm solution is recovered, revealing an incremental learning. Our proof leverages a convenient arc-length time-reparametrisation which enables to keep track of the transitions between the jumps. Our analysis requires negligible assumptions on the data, applies to both under and overparametrised settings and covers complex cases where there is no monotonicity of the number of active coordinates. We provide numerical experiments to support our findings.

Paper

Similar papers

Peer review

Reviewer tpj27/10 · confidence 3/52023-06-22

Summary

This paper characterizes the trajectory of gradient flow on 2-layer diagonal linear networks for linear regression tasks. Specifically, the paper considers the model parameterized as $x \mapsto \langle u \odot w, x \rangle$ in the linear regression setting. By interpreting the gradient flow on the nonconvex loss as a mirror flow, the authors show, in the limit of vanishing initialization, that the gradient flow dynamics jump between various saddles before converging to the minimum $\ell_1$ norm solution. The paper characterizes the exact times of these jumps (after appropriate rescaling based on the initialization size $\alpha$ in the $\alpha \rightarrow 0$ limit), as well as the location of these saddles; these times and locations can be computed via algorithm 1. As a consequence, the paper shows that each saddle is a solution to a constrained minimization problem where some subset of coordinates are fixed 0. Under a RIP assumption on the data, the paper shows that new coordinates are added sequentially at each saddle, thus demonstrating incremental learning.

Strengths

- The paper appears to be technically sound, and is well-written and easy to understand. - The quadratically parameterized regression problem has been well studied in prior works (Woodworth et al. 2020, Azulay et al. 2021, and others), which show that the gradient flow can be characterized by mirror descent and converges to the minimum $\ell_1$ solution in the $\alpha \rightarrow 0$ limit. However, the current paper exactly characterizes the limit of the gradient flow trajectory under minimal data assumptions, which is a novel analysis. The main technical novelty here is proving that as $\alpha \rightarrow 0$ the trajectory converges to a piecewise constant process. - While many prior works on implicit regularization only focus on properties of the solution at convergence, the current paper can describe the saddle points along the trajectory. I find it to be a significant contribution that the successive saddles which can be visited by gradient flow can be characterized (by Algorithm 1) explicitly. In particular, the observation that coordinates can be deactivated has not appeared in prior work and is quite interesting. - Finally, I find this paper to be of moderate significance, as the quadratically parameterized linear regression setting is a common toy problem to understand the implicit regularization effect of gradient descent more generally.

Weaknesses

There are a couple (minor) weaknesses of the current theory. - The analysis is limited to the quadratically parameterized regression setting, which while despite possessing rich implicit regularization behavior is still far from more practical settings in which implicit regularization occurs. - One weakness of the current paper is that the jump times and saddle locations are only implicitly defined via Algorithm 1. The paper does show that under an RIP assumption that the jump times are $1/\beta^*_s$ and the saddles correspond to incrementally learning more coordinates. However, it is difficult to interpret the intermediate iterates of Algorithm 1 more generally, which reduces the impact of this paper. - Another weakness is that the entirety of the analysis is done in the $\alpha \rightarrow 0$ limit, rather than at some small (but finite) initialization $\alpha$. Minor typos: - line 133 “diferred” → “deferred” - line 318 “independant” → independent”

Questions

- In general, what can one say about successive saddles? It would be interesting to understand more fine grained properties about the sequence of saddles, such as how many coordinates can change at each step and under what conditions coordinates can get deactivated. I find it unlikely that Algorithm 1 would loop through all possible subsets of activated coordinates, and instead find it more likely that coordinates would be deactivated less frequently. I thus think the paper would benefit from additional investigation (either theoretically or empirically) into how the set of activated coordinates changes between saddles. - What can be said in the case when $\alpha$ is small but isn’t taken to 0? Can one say something quantitatively about whether saddle-to-saddle dynamics occur, or the rate at which the trajectory limits to the piecewise constant trajectory $\tilde \beta^\circ$?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Limitations/Broader Impact are adequately addressed.

Reviewer P2pV8/10 · confidence 5/52023-07-04

Summary

This work concerns the behavior of gradient flow of 2-layer diagonal linear neural networks when the initialization scale goes to zero. The authors show that this limiting behavior is governed by a piecewise constant trajectory consisting of jumps from saddle to saddle.

Strengths

The incremental learning and saddle-to-saddle phenomenon is interesting. Although the model considered in this paper is equivalent to linear models, the dynamics that the authors uncover is nonetheless nontrivial and of broader interest. The writing quality is excellent and the technical material is presented in an intuitive to understand way.

Weaknesses

Is the final point $\beta_p$ in Theorem 2 the minimum $\ell_1$ solution to $L$? Theorem 2 does not explicitly state it. Much of the later analysis is regarding the $\beta$'s. But the model is stated in terms of the $u$'s and $v$'s. Can the authors comment on the behavior of the $u$'s and $v$'s? (Or at least point me to where this was discussed, in case I missed it). line 40 "diagonal linear networks which are simplified neural networks that have received significant attention lately...". Could the authors provide a slightly expanded discussion of what these prior work did? Since this paper is specifically about diagonal linear networks, this part of the related work I believe is especially important.

Questions

What is the significance of the connection to the Homotopy algorithm? I'm not familiar with it and would appreciate if the authors can explain a bit more why this connection matters. If the authors can address this and the weaknesses above, I'll be happy to raise the score to a 7.

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

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

4 excellent

Contribution

2 fair

Limitations

I do not believe the authors addressed the limitations, at least not in an explicit way. I think the paper would be greatly improved if it can touch upon the comment brought up on line 78 that the activation function is the identity. The authors make it very clear that the work is about linear networks, as many other prior works have done. I feel like if the authors can provide some form of experimental results/explanation of whether this phenomenon appear (or does not appear) where $\sigma$ is nonlinear would be greatly beneficial. That said, I think the paper's technical contribution outweighs this. If the authors can address this and all of the above, I'll be happy to raise the score to an 8.

Reviewer 8NQr6/10 · confidence 3/52023-07-05

Summary

This paper analyzes a two-layer diagonal linear network, which is a linear regression model where the linear weight is parameterized as the point-wise product of two weight vectors. It is shown that, with vanishing initialization, gradient flow will jump between saddle points of the training loss and eventually reach the $\ell_1$-minimum-norm solution. This paper further provides an algorithm to calculate all the saddle points.

Strengths

It is very important to understand the training dynamics of neural networks, and this paper analyzes a novel and interesting saddle-to-saddle phenomenon. Moreover, Algorithm 1 in this paper can compute all the jump times and saddle points, giving a complete characterization of the gradient flow trajectory. The writing is also clean, with illustrative examples and graphs.

Weaknesses

The main weakness is that the two-layer diagonal linear network model (specifically, the point-wise product of two weight vectors) is too simple. If there is some evidence that some of form of the saddle-to-saddle phenomenon also happens in practical networks, this paper will be more convincing.

Questions

N/A

Rating

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

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

N/A

Reviewer FL6c7/10 · confidence 4/52023-07-07

Summary

In this paper, the authors studied the training dynamics of gradient flow that minimizes mean-square loss with 2-layer diagonal linear networks and data in general position. Under the small initialization (initialization scale goes to 0), the authors showed that the limiting dynamics follows a saddle-to-saddle dynamics (jump from one saddle point to another saddle point) until reach the min-$\ell_1$-norm interpolator. They gave an algorithm that can compute the visited saddle points. The results generalized the previous works on incremental learning. Experiments are also provided to verify the results.

Strengths

1. The paper is clearly written and easy-to-follow. The proof sketch and example are given to help the readers to understand the proof easier. 2. Understanding the training dynamics of neural networks/nonlinear models is an interesting and important problem. 3. This paper gives a precise characterization of the training dynamics that jumps between saddle points. This recovers and goes beyond some of the previous works on the incremental learning for linear diagonal networks. 4. The technique of reparametrized time to “accelerate” time seems to be interesting and might be of independent interest.

Weaknesses

1. The current paper focuses on the linear diagonal linear networks. It would be interesting to see if such analysis could be generalized to other more complicated problems.

Questions

1. In Theorem 2, it seems that the final solution is min-$\ell_1$-norm interpolator, so I think that means we are implicitly assuming the input dimension $d$ is at least the number of samples $n$?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, 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

3 good

Presentation

3 good

Contribution

3 good

Limitations

The limitation is discussed in the paper. This is a theoretical work and therefore does not seem to have negative societal impact.

Reviewer BqoH5/10 · confidence 3/52023-07-10

Summary

This paper studies the saddle-to-saddle dynamics in Diagonal Linear Networks. The authors present solid theoretical understanding. They show that over 2-layer diagonal linear network, gradient flow starting with vanishing initialization visits then jump from saddles jumps from a saddle of the training loss to another until reaching the minimum $\ell_1$-norm solution.

Strengths

The paper is clear and well-written. Understanding the training dynamics of gradient descent over neural networks is a significant theoretical issue. Particularly, the phenomenon of saddle-to-saddle during neural network training remains mysterious, and this paper makes a valuable contribution by offering a solid theoretical analysis for 2-layer diagonal linear networks. The proof presented in the paper incorporates innovative techniques, including mirror flow and time-reparametrization.

Weaknesses

The article's significance may be constrained by the focus on 2-layer diagonal linear networks, which have limited representation abilities and are no better than linear models. However, this limitation is not a major concern since the problem itself is non-convex, even in this simplified scenario.

Questions

1. I would like to know how much the vanishing initialization would affect the saddle-to-saddle phenomenon. Will saddle-to-saddle dynamics also occur if a practical initialization is used? 2. As shown in the theory and Figure 2, the recovery of coordinates is sequential, which seems to be similar to the Coordinate Descent algorithm. In this setting of 2-layer diagonal linear networks, are GD and CD inherent related?

Rating

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

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

The analysis in this study focus on 2-layer diagonal linear networks, which have limited representation abilities and exhibit some special properties such as Prop 1.

Reviewer BqoH2023-08-11

Thanks for the reviewer's detailed response to my questions.

Reviewer tpj22023-08-11

Response to Authors

Thank you to the authors for responding to my questions. I agree with the authors that the current results are indeed novel and relevant, and that the additional results I asked about are beyond the scope of the current paper (though still interesting!). I have increased my score from a 6 to a 7.

Reviewer P2pV2023-08-11

I thank the authors for a thorough reply. I have raised my score to an 8.

Reviewer FL6c2023-08-11

Thanks for the response to address my question. I will keep my score.

Area Chair oVW32023-08-18

Thank you for the rebuttal

Dear authors, thank you for providing a rebuttal. Most of the reviewers have already replied, so this is just to let you know that I am in contact with the remaining one as well. Best, Your AC

Reviewer 8NQr2023-08-18

Thanks for the response! I will keep my score.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC