Leveraging the two timescale regime to demonstrate convergence of neural networks

We study the training dynamics of shallow neural networks, in a two-timescale regime in which the stepsizes for the inner layer are much smaller than those for the outer layer. In this regime, we prove convergence of the gradient flow to a global optimum of the non-convex optimization problem in a simple univariate setting. The number of neurons need not be asymptotically large for our result to hold, distinguishing our result from popular recent approaches such as the neural tangent kernel or mean-field regimes. Experimental illustration is provided, showing that the stochastic gradient descent behaves according to our description of the gradient flow and thus converges to a global optimum in the two-timescale regime, but can fail outside of this regime.

Paper

Similar papers

Peer review

Reviewer VU9p7/10 · confidence 3/52023-06-26

Summary

This paper studies the training of 2-layer neural networks and proposes a two-timescale limit/regime. In this limit/regime, the learning rate of the first layer is much smaller than the learning rate of the second layer. As a result, the training of the network can be viewed as training the first layer and performing linear regression over the first layer outputs, which is simpler/more structured than training both layers at the same rate. To demonstrate the usefulness of this strategy, this paper considers a toy model and shows that a 2-layer network can fit a certain family of 1D piecewise constant functions. The authors also empirically show that SGD can fail to fit this family of functions outside the two-timescale regime.

Strengths

* The presentation is clear. * This paper not only contains results in the $\varepsilon \to 0$ limit, but also non-asymptotic results for small but non-vanishing $\varepsilon$. The derivation in the $\eta \to 0, \varepsilon \to 0$ limit is clean (Sec. 4.2), and the asymptotic-to-non-asymptotic parts are also easy to follow. * This paper reminds me of a technique that is gaining popularity in the theory community: training the network for one (large) step, freezing the first layer, and then performing linear regression in the second layer using the features learned in the first layer (cf. [AAM22], [DLS22]). The limitation of this technique is that since the first layer is fixed after the first layer, it lacks the ability to refine the learned features. I feel the strategy introduced in this paper is a potential remedy to this problem, as here we also have the linear regression part of the argument, but the first layer can be trained for multiple steps. [AAM22] Abbe, Emmanuel, Enric Boix Adsera, and Theodor Misiakiewicz. “The Merged-Staircase Property: A Necessary and Nearly Sufficient Condition for SGD Learning of Sparse Functions on Two-Layer Neural Networks.” In Proceedings of Thirty Fifth Conference on Learning Theory, 4782–4887. PMLR, 2022. https://proceedings.mlr.press/v178/abbe22a.html. [DLS22] Damian, Alex, Jason D. Lee, and Mahdi Soltanolkotabi. “Neural Networks Can Learn Representations with Gradient Descent.” arXiv, June 30, 2022. http://arxiv.org/abs/2206.15144.

Weaknesses

* Almost all things in this paper are 1D and somewhat tailored to this specific piecewise constant function class. I wonder whether/how this can be generalized to higher dimensions, other network architectures, and more general function classes. * It seems that the dynamics are still relatively local. That is, we need some neurons in each target interval at initialization, which may not be reasonable when the dimension is high. * The authors should probably add some discussion on the training-for-one-large-step-type technique (see the Strengths part of the review).

Questions

* How general is this strategy? Is it possible to apply this strategy to problems (even toy ones) with dimension higher than $1$ and more standard 2-layer networks? * Can we relax the requirement on the initialization? For example, can we remove the requirement that each interval contains at least $6$ $u_j$? Intuitively, when some interval contains no neurons, it is still possible that some neurons from those fitted intervals can move to that interval and fit the target function on that interval. I know the current requirement is reasonable in your setting, I am asking this because I feel this type of global feature learning process is important when the dimension is high.

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

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

This is a theoretical work and, as far as I can see, has no potential negative societal impacts.

Reviewer 83X85/10 · confidence 4/52023-07-04

Summary

This paper studied the problem of fitting a piecewise constant univariate function with a shallow neural network. Specifically, the authors consider a gradient flow with a time-scale difference between the dynamics of the first and the second layer weights. It is shown that the trained shallow network can be arbitrarily close to the target function in $\mathcal{L}_2$, as long as the weight update on the first layer is much slower than one on the second layer.

Strengths

The two-time-scale regime seems novel in training neural networks. The results are well presented and their proofs are clearly explained.

Weaknesses

Only a very special problem is studied. 1. The target function is univariate, piece-wise constant. This is very restrictive. 2. The loss is a population loss, thus the case of finite samples (fixed design or random samples) is not considered. 3. Network is quite different from those studied in other works. The activation function is a smoothed version of a stair function; the first layer is only parametrized by the bias at every neuron. With this many assumptions/restrictions, even though the main results are well presented and explained, it is hard to see whether those observations can give insights into the usefulness of implementing two-time-scale training for practical networks. The author did not discuss the limitations of these assumptions, nor did they show how two-time-scale training can be used in practice, even empirically.

Questions

See "Weaknesses"

Rating

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

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

2 fair

Limitations

See "Weaknesses"

Reviewer JDX56/10 · confidence 4/52023-07-07

Summary

In this paper, the authors considered the problem of learning piece-wise linear function in 1d using two-layer neural network. They considered gradient flow on mean-square loss with different learning rates for 2 layers (two-timescale). Specifically, the outer layer weights are moving much faster than the inner layer weights. The activation used in 2-layer network is similar to a rescale version of sigmoid activation. This paper shows that under proper choice of the parameters, GF converges to small loss within polynomial time in relevant parameters. Experiments are provided to support the results.

Strengths

1. Understanding the training dynamics and convergence of neural networks is an important problem. 2. The paper is overall easy to follow and clearly written. The proof sketch is given so that the reader can understand the main proof idea. 3. The idea of this two-timescale/two different learning rates in analyzing the training dynamics for neural networks seems to be interesting.

Weaknesses

1. The problem considered is only in 1d and it would be interesting to see if the analysis could be generalized to multi-dimension.

Questions

1. In the experiments, I was wondering if one could elaborate in Figure 5 that why ‘the dynamics create a zone with no neuron’. It seems to me that there are neurons near every discontinuity of $f^*$. Also, it seems in Figure 4(c) and Figure 5(c), the number of steps for training are different, what is the reason for that?

Rating

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

Reviewer WNiU6/10 · confidence 4/52023-08-02

Summary

The paper studies the training dynamics of fitting a one hidden layer shallow network with heaviside activation to a piecewise ground truth function with one-dimensional input. It proves that gradient flow always recovers the ground truth in finite time with only mild over-parametrization.

Strengths

This paper is well-written, and intuitions behind technical proofs are well-presented, and theoretical results are well-supported by numerical experiments. The theoretical result itself is a nice observation, despite being in the simple one-dimensional input case.

Weaknesses

The paper only considers the one-dimensional case. Based on the proof techniques, it is not clear if it is extendable to high dimension, which is of ultimate interests in the deep learning theory community, since the derivation in section 4.2 would not hold any more. While it is understandable that such a result would be difficult to obtain in high dimension, the paper didn't present any experimental results in the high-dimensional case either.

Questions

The paper would greatly benefit from a few comments on whether the authors would expect the same phenomenon to be observed in high dimension and what the difficulty would be. Some numerical experiments in the high dimensional case would be a great addition as well.

Rating

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

3 good

Contribution

3 good

Limitations

N/A

Reviewer JDX52023-08-11

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

Reviewer 83X82023-08-14

Thank the authors for the response. I personally still think the assumptions (1. and 3. in my original review) are very strong. Nonetheless, the authors provide additional numerical experiments in the rebuttal showing these assumptions might be relaxed, thus I raised the score.

Authorsrebuttal2023-08-15

Thank you for your questions that contribute to improving the paper by adding a thorough discussion on possible generalizations, and for your willingness to raise the score.

Reviewer VU9p2023-08-14

Thanks for the response. I will keep my score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC