Addressing Spectral Bias of Deep Neural Networks by Multi-Grade Deep Learning

Deep neural networks (DNNs) suffer from the spectral bias, wherein DNNs typically exhibit a tendency to prioritize the learning of lower-frequency components of a function, struggling to capture its high-frequency features. This paper is to address this issue. Notice that a function having only low frequency components may be well-represented by a shallow neural network (SNN), a network having only a few layers. By observing that composition of low frequency functions can effectively approximate a high-frequency function, we propose to learn a function containing high-frequency components by composing several SNNs, each of which learns certain low-frequency information from the given data. We implement the proposed idea by exploiting the multi-grade deep learning (MGDL) model, a recently introduced model that trains a DNN incrementally, grade by grade, a current grade learning from the residue of the previous grade only an SNN composed with the SNNs trained in the preceding grades as features. We apply MGDL to synthetic, manifold, colored images, and MNIST datasets, all characterized by presence of high-frequency features. Our study reveals that MGDL excels at representing functions containing high-frequency information. Specifically, the neural networks learned in each grade adeptly capture some low-frequency information, allowing their compositions with SNNs learned in the previous grades effectively representing the high-frequency features. Our experimental results underscore the efficacy of MGDL in addressing the spectral bias inherent in DNNs. By leveraging MGDL, we offer insights into overcoming spectral bias limitation of DNNs, thereby enhancing the performance and applicability of deep learning models in tasks requiring the representation of high-frequency information. This study confirms that the proposed method offers a promising solution to address the spectral bias of DNNs.

Paper

Similar papers

Peer review

Reviewer otLk6/10 · confidence 3/52024-07-10

Summary

This paper addresses the issue of spectral bias in deep neural networks (DNNs), where DNNs tend to prioritize learning lower-frequency components of a function, struggling with high-frequency features. The proposed solution involves decomposing high-frequency functions into compositions of low-frequency functions using a novel approach termed Multi-Grade Deep Learning (MGDL). MGDL incrementally trains DNNs, each grade focusing on learning low-frequency information from the data. The efficacy of MGDL is demonstrated on synthetic, manifold, and MNIST datasets, showing improvements in capturing high-frequency components compared to traditional methods.

Strengths

Quality: The paper provides a thorough theoretical foundation and supports its claims with experimental results, although these are not entirely convincing. Clarity: The paper is generally well-organized, with clear explanations and effective use of figures and mathematical expressions.

Weaknesses

Originality: lacks new theoretical contributions for MGDL method. Performance and Stability: There are concerns about the performance and stability of MGDL, as evidenced by loss spikes and lack of significant loss reduction in real-world tasks (Figure 8). Minor Punctuation Errors: For example, On Table 1,"32,401" and "27, 817". Image Readability: The legends of figures in this paper are too small, making them difficult to read and interpret.

Questions

1. Is the MGDL method mentioned in the paper applicable to other model architectures such as CNNs and Transformers? 2. Does the training time of MGDL increase linearly with the number of grades? 3. Please add the following citation as it is one of the earlier methods proposed to alleviate slow convergence for high frequencies: Cai W, Xu Z Q J. Multi-scale deep neural networks for solving high dimensional PDEs. arXiv preprint arXiv:1910.11710, 2019.

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

See weeknesses.

Reviewer GtgE4/10 · confidence 4/52024-07-13

Summary

This paper tackles the spectral bias of deep neural networks by proposing a method to capture high-frequency data effectively. The authors assert that previous works have not adequately addressed spectral bias, emphasizing its significance. They propose that high-frequency functions can be decomposed into a sum-composition of low-frequency functions. To implement this, they employ a multi-grade deep learning (MGDL) model to achieve the sum-composition functions from the original function. The authors provide empirical evidence of the effectiveness of this approach in three distinct scenarios, including high-frequency noise-injected MNIST.

Strengths

- This paper is easy to follow and well-written. - The logical progression from the motivation that high-frequency components can be estimated by the sum of composites of low-frequency components is sound. - The proposed method would likely address scenarios involving high-frequency noises or data containing high-frequency components.

Weaknesses

- The paper fails to highlight why addressing spectral bias is significant. Recent network architectures perform well in various computer vision fields without specifically addressing this issue, where the datasets they use mostly contain high-frequency noises or noisy data themselves. From my perspective, the advantage of this paper in handling high-frequency components primarily benefits specific synthetic scenarios, such as those provided by the authors. I believe there would be more compelling examples to underscore the importance of the matter where the claim is significant. I recommend that the authors provide more practical scenarios where not addressing spectral bias leads to tangible failures, along with a rationale for why it is important to capture high-frequency elements. - While Figure 1 demonstrates the effectiveness of sum-composition in shallow deep neural networks, it does not provide a rationale for freezing components in each composition as done in MGDL. Based on this motivation alone, other learning methods beyond MGDL could potentially achieve the same or even better results. In this regard, why was MGDL exclusively used for the experiments? Considering that a residual network (ResNet) also employs a sum-composition approach, why wasn't it included in the experimental analysis? - I believe the major problem with this paper lies in the experiment section. It provides insufficient experimental evidence to substantiate the authors' claims. - First of all, as aforementioned, the scope of the experiments is very limited, involving only three synthetical scenarios and experiments with a fixed network architecture for each scenario. Additionally, comparisons are made solely with SGDL. - MGDL appears to capture diverse frequency successfully ranges through multiple grades (whereas SGDL does not). However, I suspect this outcome may stem from learning instabilities caused by a simplistic network architecture lacking 1) normalization layers; 2) skip connections (i.e., shortcuts); and 3) freezing some layers for MGDL. In this regard, a deep SGDL is highly likely to encounter training failures due to gradient issues, whereas MGDL freezing some layers may offer more training stability in reducing loss. Therefore, I believe that any methods or networks that train shallow networks sequentially could potentially benefit in a manner similar to [1]. All experiments should be reperformed to provide more convincing experimental support. - [1] Relay Backpropagation for Effective Learning of Deep Convolutional Neural Networks, ECCV 2016 - The experiments on high-dimensional data are limited to MNIST, which does not convincingly demonstrate the challenges of high-dimensional data as claimed. For example, the standard ImageNet training protocol uses 224x224 images, resulting in a dimension of 50,176 when vectorized, - Figure 8 contains very small fonts, which should be refined for better readability.

Questions

Please see the above weaknesses.

Rating

4

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

- Limitations are provided

Reviewer vDtJ6/10 · confidence 4/52024-07-29

Summary

DNNs suffer from spectral bias, which is the tendency to prioritize learning of low-frequency functions. The authors propose a method called MGDL, which tackles the spectral bias problem by changing the network parameterization and loss function such that each layer in the network learns the residual of the function from the previous layers. This framework forces higher layers in the network to learn high frequency residuals rather than an additive low frequency function. The authors justify their method using regression experiments on synthetic and MNIST data.

Strengths

- The idea presented is clear and intuitive. - I found section 2 to be easy to understand despite the many different notations used. - The experiments are organized in a logical order. Figure 4 clearly shows the advantage of MGDL in modeling functions with different frequency components. - Lastly the authors are honest in acknowledging the lack of theoretical justification of MGDL.

Weaknesses

(1) The main weakness (and probably the most obvious) is the question of scalability. There is a heavy focus in this paper on synthetic experiments. The only real world experiments are done on MNIST. I am uncertain as to whether this work has the potential to be adopted for any real applications (in computer vision or otherwise). In other words, I think the impact is limited due to the lack of more complex experiments.

Questions

(2) Is there a reason you are constrained to the MSE loss? Does this idea work on other loss functions? (3) I am confused by Figure 3. I see that this is a plot of the functions learned by each layer in the frequency domain, but what is the main take away? For example, Grade 4 does not seem to be learning anything in the first two settings, whereas according to my intuition it should be learning the highest frequencies.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Adequately addressed.

Reviewer vDtJ2024-08-10

Rebuttal response

I read the rebuttal and I don't have much to add to my original review. Thanks for the rebuttal and answering my questions!

Reviewer otLk2024-08-08

For the weakness: Our concern, "There are concerns about the performance and stability of MGDL, as evidenced by loss spikes and lack of significant loss reduction in real-world tasks (Figure 8)," was not clearly addressed, nor were additional numerical experiments provided to alleviate our doubts. Additionally, there is inconsistency in the use of commas in the numerical values in Table 1, such as "32,401" and "27, 817." For question2: Can I understand it this way: When we divide the network into multiple grades, the cost of forward propagation increases during training. For example, when training the n-th grade of the network, the forward propagation process requires additional computations from the previous n-1 grades, resulting in the computational cost of forward propagation growing quadratically.

Authorsrebuttal2024-08-12

Reply to Reviewer otLk's comments

For the weakness: Thanks the referee for additional comments. After reading the referee's additional comment, we conducted two more numerical experiments. (a) The first additional experiment serves as a supplement to Section 3.3. In the original experiment, MGDL was configured with three grades, each consisting of two hidden layers of width 128. In the new experiment, MGDL was configured with six grades, each consisting of a single hidden layer with width 128. New experiment results demonstrate improvement in MGDL's performance compared to the results in the originally submitted manuscript. (b) The second experiment is a totally new example on a color image. We compare performance of MGDL and SGDL when applying them to regress a two-dimensional color image. The input to the model is the pixel coordinates, and the output is the corresponding RGB values. The train dataset consists of a regularly spaced grid containing $1/4$ of the image pixels, while the test dataset contains the full image pixels. The network structure for SGDL has 10 hidden layers, each with width 256, and the corresponding structure for MGDL has 5 grades, each grade has two hidden layers of width 256. Both SGDL and MGDL use the ReLU function as the activation for hidden layers and the sigmoid function as activation function for the output layer. The peak signal-to-noise ratio (PSNR) values are computed over the full image pixels for both images obtained from MGDL and SGDL. The PSNR value for the image generated by MGDL and SGDL is 24.07 db and 22.41 db, respectively. This result indicates that MGDL outperforms SGDL in terms of PSNR value. Notably, MGDL demonstrates the ability to learn high-frequency details more effectively than SGDL. We hope these additional numerical experiments can alleviate the referee's doubts. Thanks for pointing out the inconsistency of the commas used in Table 1. We have corrected them throughout the entire paper. For question 2: When the network is divided into multiple grades, the cost of forward propagation does not increase during training. This is because it is not necessary to reevaluate the forward propagation of the previous $n-1$ grades during the training of grade $n$. In fact, the computation cost remains relatively consistent across all grades. Below, we elaborate this point. For $$ \mathbf{x}\_{\ell}^{n} :=\mathcal{H}\_{D\_{n-1}}(\Theta\_{n-1}^*;\cdot) \circ \mathcal{H}\_{D\_{n-2}}(\Theta\_{n-2}^*;\cdot) \circ \ldots \circ \mathcal{H}\_{D\_{1}}(\Theta\_{1}^*;\cdot)(\mathbf{x}\_{\ell}) $$ we recursively let $$ \mathbf{x}\_{\ell}^1:= \mathbf{x}\_{\ell}, \ \ \mathbf{x}\_{\ell}^{k} :=\mathcal{H}\_{D\_{k-1}}(\Theta\_{k-1} ^*;\cdot) \circ \mathbf{x}\_{\ell}^{k-1}, \ \ k=2, 3, \dots, n. $$ When training grade $n$, we use the output of grade $n-1$, denoted as $\mathbf{x}\_{\ell}^{n}$ along the residual $\mathbf{e}\_{\ell}^{n}$, which are already obtained. The training dataset in grade $n$ consists of $\left\\{(\mathbf{x}\_{\ell}^{n}, \mathbf{e}\_{\ell}^{n})\right\\}\_{\ell=1}^{N}$. This data is used to train a new shallow network, which is independent of the previous $n-1$ grades. Therefore, there is no need to evaluate the forward propagation of the previous $n-1$ grades during the training of grade $n$. Moreover, $\mathbf{x}\_{\ell}^{n}$ can be computed recursively (as shown in the above equation), ensuring that the computation cost for each grade remains relatively consistent. In the MNIST dataset example (Section 3.3), the network structures are $[784, 128, 128, 10]$ for grade 1 and $[128, 128, 128, 10]$ for both grades 2 and 3. With $\beta = 1$ and $\kappa = 1$ (as shown Table 3), we trained each grade for 2000 epochs. The total training time for MGDL was 3109 seconds, with grade 1 taking 1310 seconds, grade 2 taking 906 seconds, and grade 3 taking 893 seconds. While in the same setting, the training time for SGDL is 3298 seconds.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC