High-dimensional Asymptotics of Denoising Autoencoders

We address the problem of denoising data from a Gaussian mixture using a two-layer non-linear autoencoder with tied weights and a skip connection. We consider the high-dimensional limit where the number of training samples and the input dimension jointly tend to infinity while the number of hidden units remains bounded. We provide closed-form expressions for the denoising mean-squared test error. Building on this result, we quantitatively characterize the advantage of the considered architecture over the autoencoder without the skip connection that relates closely to principal component analysis. We further show that our results accurately capture the learning curves on a range of real data sets.

Paper

References (49)

Scroll for more · 37 remaining

Similar papers

Peer review

Reviewer pba77/10 · confidence 3/52023-07-04

Summary

This paper studies the performance of denoising autoencoders (DAEs) in high-dimensional settings. The DAEs are trained on data sampled from a Gaussian mixture with $K$ components perturbed by isotropic Gaussian noise. The DAEs are one-layer networks with arbitrary activation functions, tied weights, and a skip connection. They are trained with an $L_2$-regularized loss function, and the high-dimensional limit is considered, where the ratio of the number of training points to the input dimension converges to a constant α. The main results of the paper are formulas for the denoising test mean squared error (MSE) for the full DAE, as well as two simpler architectures: a "bottleneck network" in which the skip connection is removed, and a "scalar linear network" that simply rescales the input by a scalar. The final formulas involve complicated integrals and optimization problems, but they only depend on low-dimensional quantities, such as the number of neurons in the hidden layer and the number of components in the Gaussian mixture. The authors use the derived asymptotic test error to analyze the role and importance of each component in the performance of DAEs. They find that the skip connection is essential for good performance and that DAEs can outperform other denoising methods, such as principal component analysis (PCA). Taking the noiseless limit, where the variance of the noise converges to zero, they can also derive the test error of reconstruction autoencoders, which are trained with noiseless inputs.

Strengths

The authors managed to derive fairly complicated asymptotic formulas via the replica method from statistical physics. The empirical experiments seem to confirm the accuracy of their predictions. This provides yet another example of the successful application of the replica method and its surprisingly powerful nature. This paper fills a gap in the literature and complements previous results that mainly focused on RAEs and the infinite data regime (optimization of the population loss). The exact formulas derived in this paper are utilized to derive interesting non-trivial results, which can serve as stimuli for future research. Specifically: a) The MSE obtained by training a DAE with gradient descent matches the asymptotic predictions, suggesting that while non-convex, the optimization problem has a favorable landscape. b) The exact formulas obtained for the Gaussian mixture predictions align with those obtained for a DAE trained on MNIST datasets, indicating the presence of universality phenomena. c) As α increases, the denoising MSE of the DAE approaches the asymptotic performance of the oracle denoiser (derived in the appendix).

Weaknesses

The paper exhibits weaknesses similar to other papers in this field. The exact analytical formulas derived are somewhat obscure, and it seems that several of the observations made in the paper can be deduced directly from the results of the empirical experiments, without relying on the analytical formulas (e.g. the improved performance of the network with skip connections compared to the one without can be easily inferred). The authors should emphasize the results that their analytical formulas allow to obtain.

Questions

- It appears that all the experiments were conducted with a number of neurons $p = 1$. Could you please clarify if the empirical results align with the replica predictions when $p > 1$? Additionally, does the "universality phenomenon" manifest itself also in this case? - Based on Figure 3, it appears that the performance of the DAE approaches that of the oracle denoiser. Can this convergence be rigorously derived from equation (13)? - What happens when $K = 1$? Do the DAE, PCA, and bottleneck network exhibit similar performance in this scenario?

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

This theoretical paper does not possess any immediate potential negative societal impacts.

Reviewer ZGsS7/10 · confidence 4/52023-07-04

Summary

The authors consider a two layer weight-tied denoising autoencoder with a skip connection in the regime of vanishing rate and samples proportional to the dimension. They heuristically derive an exact characterization of the optimal network parameters and the corresponding network performance in the the high-dimensional limit using the replica method. Their experiments show that their results match on synthetic data and are very close to the performance on more practical data.

Strengths

- Exact characterization of all quantities of interest in the high-dimensional limit. - Strong experimental validation of theoretical claims

Weaknesses

- From a theoretical perspective there is a lack of rigor to this method

Questions

- Why do the limits in Line 152 exist? - In Figure 3 for high noise and low sample complexity the performance of the full DAE is worse than simply rescaling. Since the full network can also act as a rescaling network by setting $w=0$, this implies that during training the global optimum is not found. Can you elaborate on this?

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

4 excellent

Limitations

- It should be stated more clearly that the way the replica method is carried out only serves as a strong heuristic and not a rigorous proof.

Reviewer DrHV7/10 · confidence 4/52023-07-05

Summary

This paper presents theoretical results of the test error of 2-layer denoising auto-encoder. In a high-dimensional limit regime where data distribution follows from Mixture of Gaussian, closed-form expressions of the error are obtained. The results are further analyzed and supported by numerical experiments on real data sets.

Strengths

- The main result (result 3.3) is highly non-trivial and provides a tight formula for the test error of the 2-layer denoising auto-encoder (DAE). - It shows that DAE can perform much better than PCA in terms of MSE error. - The role and importance of the skip connection is also highlighted in both theory and in practice, as well as the non-linearity in DAE.

Weaknesses

- The optimal MSE error mse_o grows with the data dimension d (eq. 9), however the gap between mse_f and mse_o (eq 8) remains a constant, this suggests that the difference between DAE and PCA is not so significant in terms of the relative error of MSE, i.e. | mse_f - mse_o | / mse_o vs. | mse_PCA - mse_o | / mse_o. Thus it is still not very clear whether DAE brings something essentially different to PCA or not (I agree that at least they are different). - In the main result 3.3, it is not clear how the regularization term g is used, i.e. under Assumption 3.2, the parameter lambda does not appear in eq. (8), this is a bit strange to me .

Questions

- (line 33) Usually we say the sample complexity is related to n, it is here a bit strange to say it is alpha - (line 89) What does this mean sigma(.) stays order 1 ? - (eq 13) is there a unique solution in this system of 10 equations of 10 variables ? - (line 219) regarding the Gaussian universality, do you need to perform any normalization on the data for this to hold ? The dimension of MNIST seems quite small, what is it in your mind for the universality to hold ? - (line 255) to clarify a previous point regarding PCA, does mse_f grows linearly with d ? Do you have an idea of what is the relative error of MSE of DAE and PCA ?

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 main result is very specific, but this is quite normal in the literature .

Reviewer zMEV7/10 · confidence 3/52023-07-05

Summary

The authors set out to characterize the non-linear behavior of denoising auto-encoders (DAEs), for Gaussian mixtures, in the high dimensional limit with the number of hidden units being fixed. The authors particularly tease out the role of the skip-connection, compared to the reconstruction auto-encoder (RAE) which is known to essentially perform principal component analysis (PCA). Using the replica method, the authors obtain closed-form expressions for the mean-squared error (MSE), as well as the cosine-similarity (w.r.t cluster means). The obtained formulae are supported by experiments on both synthetic and real data sets, clearly highlighting the role of the sample complexity and the noise level.

Strengths

- Presents an effective characterization of (non-linear) DAE behavior on Gaussian mixtures, clearly explaining the role of the reconstruction and scaling components supported by compelling empirical evidence (also on real data). - Draws a number of important conclusions, pointing to fruitful directions of future work, e.g., L200, L218, L278.

Weaknesses

One would wish the long sequence of mathematical expressions could be made less opaque. - This is remedied by a seemingly complete and well-composed supplementary.

Questions

**Technical comments:** - L178: could you briefly explain the asymptotic dependence on $\alpha$? - L212: could you elaborate on the choice of closely related classes? - I wonder how the method fairs with a mixture of more than 2 classes. - Is it true that $K=2$ also for the MNIST experiments? **Presentation comments:** - L158: perhaps it helps to discuss this rationale before stating the assumptions. - Understandably, it is an essential part of the contribution to provide those closed-form expressions, but I wonder which subset of the readers would benefit from having Eqs.13 and 14 in the main paper.

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

4 excellent

Limitations

Dedicating more space to help make the theoretical concepts and techniques employed more accessible, i.e., to a broader set of readers, would have been greatly appreciated. It is a limitation of this reviewer that I'm unable to assess the full scale of the theoretical derivations.

Reviewer zMEV2023-08-11

Acknowledgement

Thank you for addressing my comments. I'm adjusting my score in light of the discussion with reviewer ZGsS.

Reviewer pba72023-08-13

I would like to thank the authors for their thorough answers to my questions and for running additional experiments. These have clarified my doubts and I have accordingly increased my score. I look forward to reading the final version of the paper!

Area Chair Ff8y2023-08-13

Dear reviewers and authors, Thank you very much for your work on this submission and its evaluation. Now that the authors have responded to the reviews, I strongly encourage the reviewers to acknowledge the review, to look at other reviews and rebuttals for this submission, and to adjust their scores if needed. Thanks to those that have already done so. Authors have the possibility to reply if further questions are needed, until the 16th. Thank you very much to all, Area Chair

Reviewer ZGsS2023-08-17

I would like to thank the authors for their clear answers and the added plot which has cleared up my doubts. In accordance I slightly increased the evaluation.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC