Does a sparse ReLU network training problem always admit an optimum?

Given a training set, a loss function, and a neural network architecture, it is often taken for granted that optimal network parameters exist, and a common practice is to apply available optimization algorithms to search for them. In this work, we show that the existence of an optimal solution is not always guaranteed, especially in the context of {\em sparse} ReLU neural networks. In particular, we first show that optimization problems involving deep networks with certain sparsity patterns do not always have optimal parameters, and that optimization algorithms may then diverge. Via a new topological relation between sparse ReLU neural networks and their linear counterparts, we derive -- using existing tools from real algebraic geometry -- an algorithm to verify that a given sparsity pattern suffers from this issue. Then, the existence of a global optimum is proved for every concrete optimization problem involving a shallow sparse ReLU neural network of output dimension one. Overall, the analysis is based on the investigation of two topological properties of the space of functions implementable as sparse ReLU neural networks: a best approximation property, and a closedness property, both in the uniform norm. This is studied both for (finite) domains corresponding to practical training on finite training sets, and for more general domains such as the unit cube. This allows us to provide conditions for the guaranteed existence of an optimum given a sparsity pattern. The results apply not only to several sparsity patterns proposed in recent works on network pruning/sparsification, but also to classical dense neural networks, including architectures not covered by existing results.

Paper

References (35)

Scroll for more · 23 remaining

Similar papers

Peer review

Reviewer ay347/10 · confidence 4/52023-06-22

Summary

The authors are interested in the following question : « Given a deep learning architecture possibly with sparsity constraints, does its corresponding optimisation problem actually admit an optimal $\theta^*$? » While it is mostly taken for granted that the answer is « yes », the author proves it to be true in some cases, and false in others. They derive an algorithm in order to verify a sufficient condition in order for there to exist. More then a single way the sparseness of the considered network in ensured is considered in the article.

Strengths

1. The relevance of the work is well defended in the introduction. 2. The results concern a large pool of predictors, considering a regression task. 3. The results concern more than one way of ensuring networks to be sparse. 4. The results seem to be a significant add-on to the literature concerning closeness and non-closeness of problems involving the training of (sometimes sparse) neural networks. 5. Example 3.1., though having some flaws (see Weaknesses - Major - 1.2.), is really important for illustrating what might otherwise be a bit hard to grasp for the reader, and proving the issues that might be encountered when not considering the initial question : « Given a deep learning architecture possibly with sparsity constraints, does its corresponding optimisation problem actually admit an optimal $\theta^*$? ».

Weaknesses

_Major_ 1.1. Regularisation is indeed (lines 41-42) a common way to bypass this question, but the notion of regularisation in itself is, especially when it comes to neural networks, mostly important for limiting over-fitting. I didn’t find the arguments regarding the discarding of regularisation for neural networks training really compelling (lines 44-47), especially since it wasn’t supported by any work from the literature. 1.2. Concerning Example 3.1 and the discussion that follows (lines 216-221), the discussion is based on what's displayed on Figure 1 only, forgetting about validation / test loss; saying that $L^2$ regularisation « might be detrimental » thus forget about validation / test loss. And if the goal is simply to over-fit a problem, then even smallest regularisation would make the parameter converge to certain values, practically not affecting the obtained error in train, which probably wasn't the case, since a single (and arbitrary) regularisation parameter was applied for each training. 2. The assumptions required for having Proposition 3.1. to hold (continuous, coercive) are limiting, making the use of 0-1 loss impossible (and cross-entropy as well, if I’m correct). The way the networks are defined too, since no soft-max activation function can be used on the output layer. Those loss functions (and the soft-max) being crucial when it comes to classifications, the results are mostly relevant only for regression, which is really not the most popular type of task when it comes to neural networks. Though it has been briefly mentioned on lines 176-177, I feel like the authors were not honest enough regarding how limiting the assumptions and the way the networks are defined are. For example, a sentence such as « Though our study theoretically applies to classification schemes, it more naturally suites regression schemes, considering the assumptions made in [...] » in the abstract, or at least in the Introduction section would be sufficient. 3. The motivations of the work are partially theoretical and mostly practical, but the results, though interesting, for it proves not every network has an optimal set of parameters in different settings, are mostly inapplicable in practice, limiting the impact of the work. 4. The terminology in Table 1 might be misleading, since the term « shallow » is vague and that when it comes to Theorem 3.4 and Corollaries 3.1, 4.1, 4.2 and 4.3, « shallow » stands for « 1-hidden-layer ». What are the limitations for the works from the literature with regards to the depth of the networks? If it is 1-hidden-layer, then the whole table should refer to « 1-hidden-layers architecture », and not « shallow architecture »; if it is deeper than 1-hidden-layer in at least one case, than the underlying maximum depth should be displayed in the table. _Typos / Minor_ 1. Line 126 : « agrees with $\mathbf{A}$ on rows in $S_{\underline{r}}$ (resp. columns in in $S_{\underline{c}}$) » 2. I expect Lemma 3.2. of the main paper to be coherent with what is stated on lines 505-506 for a camera-ready version of the article. 3. Lines 259 and 260 : The statement « This constraint on the sparsity level of each 260 layer is widely used in many works on sparse neural networks » is not supported by any citation. 4. Line 261 : « Consider scalar-val\underline{u}ed ». 5. Line 359 : « cf\underline{.} » 6. Using $\Omega$ for both the finite and the non-close cases is a bit confusing; it might be more convenient to differentiate these two cases with a separate notation.

Questions

1. If the practical interest in studying problems with finite $\Omega$ is clear, I have a hard time understanding the interest of studying non-closed $\Omega$ and how it applies to real-world situations.

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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

I feel like the authors downplayed the importance of some limitations of their work (see Weaknesses – Major).

Reviewer oxZZ6/10 · confidence 3/52023-07-01

Summary

This work studies the existence of global minimum in a sparse deep learning setting

Strengths

I think this work touches on a very important problem that is understudied in the field: the existence of the global minimum for deep nonlinear networks. It is easy to image toy examples for which the global minimum does not exist, but until this work, there does not exist a formal study of this problem The problem is important and novel. Linking this problem to training a sparse network is novel, I thus support its publication, though with some reservation

Weaknesses

To me, I think the main problem is that the results are quite weak, and that the main example is not quite convincing or relevant. The example of a LU network is quite artificial and it is hard to imagine that this situation arises in deep learning. I think this work can benefit greatly if it identifies a more relevant and convincing example (and it should not be difficult to achieve). This is the main reason I give this work a weak accept There is one minor problem that I do not find serious but is worth some attention. While the paper mostly motivates from the viewpoint of the popular pruning literature, this work can benefit from discussing more about regularization-based methods for compressing neural networks. For example, see the recent work in https://arxiv.org/abs/2210.01212, which also discusses the existence of the global minimum. At its face value, the results in this work seems to motivate and advocate the use of regularization-based compression methods in deep learning; is this interpretation correct? If not, why? I think discussing this point can better clarify the implication of this work

Questions

See weakness

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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

I think the work discussed the limitations well

Reviewer 7u6K7/10 · confidence 3/52023-07-01

Summary

This paper provides necessary and sufficient conditions for whether a neural network training problem admits an optimal solution, that is, whether weights and biases exist such that the infimum of the loss function is actually attained. This is done for the classical case of empirical risk minimization on a finite training set as well as for function approximation on a continuous input domain.

Strengths

- I think the investigated question is highly interesting already for purely theoretical reasons: the existence of an optimal solution is such a fundamental property of an optimization problem. We should be able to answer this question for our "favorite" optimization problem, namely training a neural network. This work seems to play a crucial role in (i) posing this question and (ii) providing results for some cases. - Apparently the non-existence of an optimal solution can also explain divergence in practical settings, which I find compelling. Still, I would like to emphasize that the paper is a theory paper and should be judged as such. - The paper involves quite some amount of non-trivial mathematics. Although the tight reviewing schedule does not allow me to verify all details, the parts I read seem to be mathematically sound.

Weaknesses

- Many cases for the posed question remain open. In particular, most of the results are concerned with 2-layer NNs only. - The presentation of the results in the intro could be improved. Some definitions could be made earlier (or at all). See my detailed comments below. Also I find it quite difficult to read the Tables 1 and 2. One basically has to read the rest of the paper in order to truely understand what these tables are about. Comments for the authors to improve the paper (not meant as true weaknesses): - title: remove the space between "optimum" and the question mark - line 54: "the best approximation property (BAP), which guarantees the existence of an optimal solution" -> Is this the definition of BAP or just a consequence? Please make this clear. You might want to define BAP in a proper definition environment. - lines 57-79: for readers unfamiliar with sparsity in neural networks, it is quite hard to understand your contributions without reading the "notations" part later. You should spend some efforts in explaining terms like "(structured) sparse networks", "fixed sparsity level" vs. "fixed sparsity pattern", etc. already in the introduction. I think "sparsity level" is only implicitly defined very late in the paper. - line 67: Here (and at some other places in the paper) you use the term "learning problem" as a synonym to "training problem". Some people understand "learning problem" as the problem to minimize the generalization error, as opposed to the "training problem", which only aims to minimize the training error. At least in the finite domain case you are definitely in the latter regime, so I suggest to use "training problem" consistently. - Tables 1 and 2: the relation to the sparsity constraints is not really clear in the tables. In particular, what does the "sparse" adjective in brackets in the architecture column mean? Why is it in brackets? - caption of table 2: there is an extra space after the opening bracket - line 124: it is a bit weird that the "notations" section is part of the "related work" section. - line 195: contains -> contain - line 261: valed -> valued - line 293: a bit redundant "many other domains such as [...] and much more". - line 301: "(in the whole paper we naturally assume B > 0)" -> state this earlier in the paper when you use B for the first time. - line 303: provide a reference to the proof of this theorem in the appendix. - line 312: what you call "plain" here is what I would call "fully connected". You should maybe use this term here and elsewhere in the paper.

Questions

- Is there any chance to obtain a hardness result for the computational problem considered in Section 3.3? NP-hardness? Hardness for the existential theory of the reals (which is always a good candidate if the best known solution uses quantifier elimination)? See related results for NN training: https://arxiv.org/abs/2102.09798; https://arxiv.org/abs/2204.01368. - Pushing the relation to the existential theory of the reals even further, I find it very interesting that you obtain such a sharp difference between the scalar-valued output case and the multi-dimensional output case (Section 3.4). The same seems to be true for the training complexity of such shallow networks: while scalar-output networks can be trained with a combinatorial search algorithm, certifying that the problem is in NP, it turns out that training a shallow network with multi-dimensional output is ER-complete and therefore much harder (https://arxiv.org/abs/2204.01368). Again I wonder: is there any relation to your results?

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

3 good

Contribution

3 good

Limitations

The paper properly states under which circumstances (mathematical assumptions) the results are valid.

Reviewer 5AbX7/10 · confidence 3/52023-07-07

Summary

The paper studies the existence of an optimal solution to the objective of training a ReLU neural network under certain sparsity patterns. They consider two topological properties of the input function space (the best approximation property, and closedness). They provided series of results on scalar-valued neural networks, shallow neural networks, neural network with one hidden layer. Specifically, they provide necessary and sufficient conditions on the sparsity (e.g., fixed sparsity level or pattern) of the neural network to guarantee existence of an optimal solution to the training loss function.

Strengths

The paper studies an interesting problem which relates to the stability of the neural network training due to the non-existence of an optimal solution given a sparsity pattern. The paper covers a good literature review and stand themselves very well compared to prior works. Their contribution is clear. The title of the paper is very related to the goal of the paper. The paper is coherent in studying the problem of interest. The paper is written clearly. Each theorem/proposition is followed by lemma/corollary and informal proof sketch which makes it very easy to follow. For examples, the paper contains a numerical example very early on in the paper to further give intuition on the importance of the problem and a plausible scenario where an optimal solution does not exist. The results provide new insights on if a neural network will have an optimal solution. Table 1 and 2 clearly state how their work differs from prior work.

Weaknesses

It would be nice in the main paper to provide more tangible examples on what sparsity examples on what architectures may/may not result in existence of an optimal solution. NeurIPS has a broad range of audiences from theory and application (addition of this can attract more application-based deep learning researchers). - minor: corollary 3.1, valed --> valued

Questions

- For corollary 3.1, what is the dimension of the hidden layer? - Could you clarify in simple words the term number 2 on line 305 (addition of this will improve clarity).

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

Addressed properly.

Reviewer 7u6K2023-08-13

I thank the authors for sharing their thoughts in the rebuttal and answering my questions. I remain curious about the connections to results in training complexity and hope future work will generate more insights on this. I continue to vote for acceptance of this paper.

Reviewer 5AbX2023-08-13

Reviewer after Rebuttal

I thank the authors for additional clarifications on my questions. I have read their response and recommend acceptance (keep my score).

Reviewer ay342023-08-14

I thank the authors for their insightful remarks. I'm glad to see a few changes will be made in the final version of the manuscript, reflecting the comments of the various reviewers. I stand by my score and recommend the acceptance of the paper.

Reviewer oxZZ2023-08-16

reply

Thanks for the response. As there is no significant concern, I keep my original score of weak acceptance.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC