# how_do_vision_transformers_work__7d50acb2.pdf

Published as a conference paper at ICLR 2022

HOW DO VISION TRANSFORMERS WORK?

Namuk Park1,2, Songkuk Kim1

1Yonsei University, 2NAVER AI Lab {namuk.park,songkuk}@yonsei.ac.kr

The success of multi-head self-attentions (MSAs) for computer vision is now indisputable. However, little is known about how MSAs work. We present fundamental explanations to help better understand the nature of MSAs. In particular, we demonstrate the following properties of MSAs and Vision Transformers (Vi Ts):

1 MSAs improve not only accuracy but also generalization by flattening the loss landscapes. Such improvement is primarily attributable to their data specificity, not long-range dependency. On the other hand, Vi Ts suffer from non-convex losses. Large datasets and loss landscape smoothing methods alleviate this problem; 2 MSAs and Convs exhibit opposite behaviors. For example, MSAs are low-pass filters, but Convs are high-pass filters. Therefore, MSAs and Convs are complementary; 3 Multi-stage neural networks behave like a series connection of small individual models. In addition, MSAs at the end of a stage play a key role in prediction. Based on these insights, we propose Alter Net, a model in which Conv blocks at the end of a stage are replaced with MSA blocks. Alter Net outperforms CNNs not only in large data regimes but also in small data regimes.

1 INTRODUCTION

There is limited understanding of multi-head self-attentions (MSAs), although they are now ubiquitous in computer vision. The most widely accepted explanation for the success of MSAs is their weak inductive bias and capture of long-range dependencies (See, e.g., (Dosovitskiy et al., 2021; Naseer et al., 2021; Tuli et al., 2021; Yu et al., 2021a; Mao et al., 2021; Chu et al., 2021)). Yet because of their over-flexibility, Vision Transformers (Vi Ts) neural networks (NNs) consisting of MSAs have been known to have a tendency to overfit training datasets, consequently leading to poor predictive performance in small data regimes, e.g., image classification on CIFAR. However, we show that the explanation is poorly supported.

1.1 RELATED WORK

Self-attentions (Vaswani et al., 2017; Dosovitskiy et al., 2021) aggregate (spatial) tokens with normalized importances:

i Softmax QK

where Q, K, and V are query, key, and value, respectively. d is the dimension of query and key, and zj is the j-th output token. From the perspective of convolutional neural networks (CNNs), MSAs are a transformation of all feature map points with large-sized and data-specific kernels. Therefore, MSAs are at least as expressive as convolutional layers (Convs) (Cordonnier et al., 2020), although this does not guarantee that MSAs will behave like Convs.

Is the weak inductive bias of MSA, such as modeling long-range dependencies, beneficial for the predictive performance? To the contrary, appropriate constraints may actually help a model learn strong representations. For example, local MSAs (Yang et al., 2019; Liu et al., 2021; Chu et al., 2021), which calculate self-attention only within small windows, achieve better performance than global MSAs not only on small datasets but also on large datasets, e.g., Image Net-21K.

In addition, prior works observed that MSAs have the following intriguing properties: 1 MSAs improve the predictive performance of CNNs (Wang et al., 2018; Bello et al., 2019; Dai et al., 2021;

Published as a conference paper at ICLR 2022

(a) Loss landscape visualizations

Vi T Res Net

(b) Hessian max eigenvalue spectra

Figure 1: Two different aspects consistently show that MSAs flatten loss landscapes. Left: Loss landscape visualizations show that Vi T has a flatter loss (NLL + ℓ2 regularization) than Res Net. Right: Hessian max eigenvalue spectra show that the magnitude of the Hessian eigenvalues of Vi T is smaller than that of Res Net during training phases. We report the Hessian spectra at the end of the warmup phases, 100th, 200th, and 300th epochs. See Fig. 4 for a more detailed analysis.

Guo et al., 2021; Srinivas et al., 2021), and Vi Ts predict well-calibrated uncertainty (Minderer et al., 2021). 2 Vi Ts are robust against data corruptions, image occlusions (Naseer et al., 2021), and adversarial attacks (Shao et al., 2021; Bhojanapalli et al., 2021; Paul & Chen, 2022; Mao et al., 2021). They are particularly robust against high-frequency noises (Shao et al., 2021). 3 MSAs closer to the last layer significantly improve predictive performance (Graham et al., 2021; Dai et al., 2021).

These empirical observations raise immediate questions: 1 What properties of MSAs do we need to better optimize NNs? Do the long-range dependencies of MSAs help NNs learn? 2 Do MSAs act like Convs? If not, how are they different? 3 How can we harmonize MSAs with Convs? Can we just leverage their advantages?

We provide an explanation of how MSAs work by addressing them as a trainable spatial smoothing of feature maps, because Eq. (1) also suggests that MSAs average feature map values with the positive importance-weights. Even non-trainable spatial smoothings, such as a small 2 2 box blur, help CNNs see better (Zhang, 2019; Park & Kim, 2021). These simple spatial smoothings not only improve accuracy but also robustness by spatially ensembling feature map points and flattening the loss landscapes (Park & Kim, 2021). Remarkably, spatial smoothings have the properties of MSAs

1 3 . See Appendix B for detailed explanations of MSAs as a spatial smoothing.

1.2 CONTRIBUTION

We address the three key questions:

1 What properties of MSAs do we need to improve optimization? We present various evidences to support that MSA is generalized spatial smoothing. It means that MSAs improve performance because their formulation Eq. (1) is an appropriate inductive bias. Their weak inductive bias disrupts NN training. In particular, a key feature of MSAs is their data specificity, not long-range dependency. As an extreme example, local MSAs with a 3 3 receptive field outperforms global MSA because they reduce unnecessary degrees of freedom.

How do MSAs improve performance? MSAs have their advantages and disadvantages. On the one hand, they flatten loss landscapes as shown in Fig. 1. The flatter the loss landscape, the better the

Published as a conference paper at ICLR 2022

0.0 0.5 1.0 Frequency

Log amplitude

0.0 0.5 1.0 Frequency

Log amplitude

Normalized depth

(a) Relative log amplitudes of Fourier transformed feature maps.

0.1 0.3 0.5 0.7 0.9 Noise frequency

Accuracy (%)

Res Net Vi T

(b) Robustness for noise frequency

Figure 2: The Fourier analysis shows that MSAs do not act like Convs. Left: Relative log amplitudes of Fourier transformed feature map show that Vi T tends to reduce high-frequency signals, while Res Net amplifies them. Log amplitude is the difference between the log amplitude at normalized frequency 0.0π (center) and at 1.0π (boundary). See Fig. 8 for more detailed analysis. Right: We measure the decrease in accuracy against frequency-based random noise. Res Net is vulnerable to high-frequency noise, while Vi T is robust against them. We use frequency window size of 0.1π.

performance and generalization (Li et al., 2018; Keskar et al., 2017; Santurkar et al., 2018; Foret et al., 2021; Chen et al., 2022). Thus, they improve not only accuracy but also robustness in large data regimes. On the other hand, MSAs allow negative Hessian eigenvalues in small data regimes. This means that the loss landscapes of MSAs are non-convex, and this non-convexity disturbs NN optimization (Dauphin et al., 2014). Large amounts of training data suppress negative eigenvalues and convexify losses.

2 Do MSAs act like Convs? We show that MSAs and Convs exhibit opposite behaviors. MSAs aggregate feature maps, but Convs diversify them. Moreover, as shown in Fig. 2a, the Fourier analysis of feature maps shows that MSAs reduce high-frequency signals, while Convs, conversely, amplifies high-frequency components. In other words, MSAs are low-pass filters, but Convs are high-pass filters. In addition, Fig. 2b indicates that Convs are vulnerable to high-frequency noise but that MSAs are not. Therefore, MSAs and Convs are complementary.

3 How can we harmonize MSAs with Convs? We reveal that multi-stage NNs behave like a series connection of small individual models. Thus, applying spatial smoothing at the end of a stage improves accuracy by ensembling transformed feature map outputs from each stage (Park & Kim, 2021) as shown in Fig. 3a. Based on this finding, we propose an alternating pattern of Convs and MSAs. NN stages using this design pattern consists of a number of CNN blocks and one (or a few) MSA block as shown in Fig. 3c. The design pattern naturally derives the structure of canonical Transformer, which has one MSA block per MLP block as shown in Fig. 3b. It also provides an explanation of how adding Convs to Transformer s MLP block improves accuracy and robustness (Yuan et al., 2021; Guo et al., 2021; Mao et al., 2021).

Surprisingly, models using this alternating pattern of Convs and MSAs outperform CNNs not only on large datasets but also on small datasets, such as CIFAR. This contrasts with canonical Vi Ts, models that perform poorly on small amount of data. It implies that MSAs are generalized spatial smoothings that complement Convs, not simply generalized Convs.

2 WHAT PROPERTIES OF MSAS DO WE NEED TO IMPROVE OPTIMIZATION?

To understand the underlying nature of MSAs, we investigate the properties of the Vi T family: e.g., vanilla Vi T (Dosovitskiy et al., 2021); Pi T (Heo et al., 2021), which is Vi T + multi-stage ; and Swin (Liu et al., 2021), which is Vi T + multi-stage + local MSA . This section shows that these additional inductive biases enable Vi Ts to learn strong representations. We also use Res Net (He et al., 2016a) for comparison. NNs are trained from scratch with Dei T-style data augmentation (Touvron et al., 2021) for 300 epochs. The NN training begins with a gradual warmup (Goyal et al., 2017) for 5 epochs. Appendix A provides more detailed configurations and background information for experiments.

Published as a conference paper at ICLR 2022

Smoothing Block

2048, 1 1, 512

512, 3 3, 512

512, 3 3, 2048

Res Net Blocks L

Subsampling

(a) Spatial smoothing

512, 1 1, 2048

2048, 1 1, 512

512, MSA, D H

D H, 1 1, 512

MLP Block 1

A"ention Block

(b) Canonical Transformer

512, MSA, D H

D H, 1 1, 2048

2048, 1 1, 512

512, 3 3, 512

512, 1 1, 2048

2048, 1 1, 512

Res Net Blocks L

A"ention Block

Subsampling

(c) Alternating pattern (ours)

Figure 3: Comparison of three different repeating patterns. Left: Spatial smoothings are located at the end of CNN stages. Middle: The stages of Vi Ts consist of repetitions of canonical Transformers. D is the hidden dimension and H is the number of heads. Right: The stages using alternating pattern consists of a number of CNN blocks and an MSA block. For more details, see Fig. 11.

The stronger the inductive biases, the stronger the representations (not regularizations). Do models with weak inductive biases overfit training datasets? To address this question, we provide two criteria on CIFAR-100: the error of the test dataset and the cross-entropy, or the negative log-likelihood, of the training dataset (NLLtrain, the lower the better). See Fig. 5a for the results.

Contrary to our expectations, experimental results show that the stronger the inductive bias, the lower both the test error and the training NLL. This indicates that Vi T does not overfit training datasets. In addition, appropriate inductive biases, such as locality constraints for MSAs, helps NNs learn strong representations. We also observe these phenomena on CIFAR-10 and Image Net as shown in Fig. C.1. Figure C.2 also supports that weak inductive biases disrupt NN training. In this experiment, extremely small patch sizes for the embedding hurt the predictive performance of Vi T.

Vi T does not overfit small training datasets. We observe that Vi T does not overfit even on smaller datasets. Figure 5b shows the test error and the training NLL of Vi T on subsampled datasets. In this experiment, as the size of the dataset decreases, the error increases as expected, but surprisingly, NLLtrain also increases. Thanks to the strong data augmentation, Vi T does not overfit even on a dataset

Res Net Vi T Vi T (6%)

Figure 4: Hessian max eigenvalue spectra show that MSAs have their advantages and disadvantages. The dotted line is the spectrum of Vi T using 6% dataset for training. Left: Vi T has a number of negative Hessian eigenvalues, while Res Net only has a few. Right: The magnitude of Vi T s positive Hessian eigenvalues is small. See also Fig. 1b for more results.

size of 2%. This suggests that Vi T s poor performance in small data regimes is not due to overfitting.

Vi T s non-convex losses lead to poor performance. How do weak inductive biases of MSAs disturb the optimization? A loss landscape perspective provides an explanation: the loss function of Vi T is non-convex, while that of Res Net is strongly (near-)convex. This poor loss disrupts NN training (Dauphin et al., 2014), especially in the early phase of training (Jastrzebski et al., 2020; 2021). Figure 1b and Fig. 4 provide top-5 largest Hessian eigenvalue densities (Park & Kim, 2021) with a batch size of 16. The figures show that Vi T has a number of negative Hessian eigenvalues, while Res Net only has a few.

Figure 4 also shows that large datasets suppress negative Hessian eigenvalues in the early phase of training. Therefore, large datasets tend to help Vi T learn strong representations by convexifying the loss. Res Net enjoys little benefit from large datasets because its loss is convex even on small datasets.

Published as a conference paper at ICLR 2022

RX R Twins Swin Pi T Vi T Mixer

CNN Local MSA Global MSA

Error NLLtrain

(a) Error and NLLtrain for each model.

101 102 Dataset size (img/class)

(b) Performance of Vi T for dataset size.

Figure 5: Vi T does not overfit training datasets. R is Res Net and RX is Res Ne Xt. Left: Weak inductive bias disturbs NN optimization. The lower the NLLtrain, the lower the error. Right: The lack of dataset also disturbs NN optimization.

Vi T + CLS Vi T + GAP

Figure 6: GAP classifier suppresses negative Hessian max eigenvalues in an early phase of training. We present Hessian max eigenvalue spectrum of Vi T with GAP classifier instead of CLS token.

Loss landscape smoothing methods aids in Vi T training. Loss landscape smoothing methods can also help Vi T learn strong representations. In classification tasks, global average pooling (GAP) smoothens the loss landscape by strongly ensembling feature map points (Park & Kim, 2021). We demonstrate how the loss smoothing method can help Vi T improve performance by analyzing Vi T with GAP classifier instead of CLS token on CIFAR-100.

Figure 6 shows the Hessian max eigenvalue spectrum of the Vi T with GAP. As expected, the result shows that GAP classifier suppresses negative Hessian max eigenvalues, suggesting that GAP convexify the loss. Since negative eigenvalues disturb NN optimization, GAP classifier improve the accuracy by +2.7 percent point.

Likewise, Sharpness-Aware Minimization (SAM) (Foret et al., 2021), an optimizer that relies on the local smoothness of the loss function, also helps NNs seek out smooth minima. Chen et al. (2022) showed that SAM improves the predictive performance of Vi T.

MSAs flatten the loss landscape. Another property of MSAs is that they reduces the magnitude of Hessian eigenvalues. Figure 1b and Fig. 4 shows that the eigenvalues of Vi T are significantly smaller than that of CNNs. While large eigenvalues impede NN training (Ghorbani et al., 2019), MSAs can help NNs learn better representations by suppressing large Hessian eigenvalues. Figure 1a also support this claim. In Fig. 1a, we visualize the loss landscapes by using filter normalization (Li et al., 2018), and the loss landscape of Vi T is flatter than that of Res Net. In large data regimes, the negative Hessian eigenvalues the disadvantage of MSAs disappears, and only their advantages remain. As a result, Vi Ts outperform CNNs on large datasets, such as Image Net and JFT (Sun et al., 2017). Pi T and Swin also flatten the loss landscapes. For more details, see Fig. C.4.

A key feature of MSAs is data specificity (not long-range dependency). The two distinguishing features of MSAs are long-range dependency and data specificity, also known as data dependency, as discussed in Section 1.1. Contrary to popular belief, the long-range dependency hinders NN optimization. To demonstrate this, we analyze convolutional Vi T, which consists of two-dimensional convolutional MSAs (Yang et al., 2019) instead of global MSAs. Convolutional MSAs calculates self-attention only between feature map points in convolutional receptive fields after unfolding the feature maps in the same way as convolutions.

Figure 7a shows the error and NLLtrain of convolutional Vi Ts with kernel sizes of 3 3, 5 5, and 8 8 (global MSA) on CIFAR-100. In this experiment, 5 5 kernel outperforms 8 8 kernel on both the training and the test datasets. NLLtrain of 3 3 kernel is worse than that of 5 5 kernel, but better than that of global MSA. Although the test accuracies of 3 3 and 5 5 kernels are comparable, the robustness of 5 5 kernel is significantly better than that of 3 3 kernel on CIFAR-100-C (Hendrycks & Dietterich, 2019).

Published as a conference paper at ICLR 2022

3 3 5 5 8 8 Kernel size

Error NLLtrain

(a) Error and NLLtrain of Vi T with local MSA for kernel size

100 50 0 Max Eigenvalue

P = 3 3 P = 5 5 P = 8 8

0 100 Max Eigenvalue

(b) Hessian negative and positive max eigenvalue spectra in early phase of training

Figure 7: Locality constraint improves the performance of Vi T. We analyze the Vi T with convolutional MSAs. Convolutional MSA with 8 8 kernel is global MSA. Left: Local MSAs learn stronger representations than global MSA. Right: Locality inductive bias suppresses the negative Hessian eigenvalues, i.e., local MSAs have convex losses.

Figure 7b shows that the strong locality inductive bias not only reduce computational complexity as originally proposed (Liu et al., 2021), but also aid in optimization by convexifying the loss landscape. 5 5 kernel has fewer negative eigenvalues than global MSA because it restricts unnecessary degrees of freedom. 5 5 kernel also has fewer negative eigenvalues than 3 3 kernel because it ensembles a larger number of feature map points (See also Fig. 6). The amount of negative eigenvalues is minimized when these two effects are balanced.

It is clear that data specificity improves NNs. MLP-Mixer (Tolstikhin et al., 2021; Yu et al., 2021a), a model with an MLP kernel that does not depend on input data, underperforms compared to Vi Ts. Data specificity without self-attention (Bello, 2021) improves performance.

3 DO MSAS ACT LIKE CONVS?

Convs are data-agnostic and channel-specific. In contrast, MSAs are data-specific and channelagnostic. This section shows that these differences lead to large behavioral differences. It suggests that MSAs and Convs are complementary.

0.0 0.2 0.4 0.6 0.8 1.0 Normalized depth

Log amplitude

Figure 8: MSAs (gray area) generally reduce the high-frequency component of feature map, and MLPs (white area) amplify it. This figure provides log amplitude of Vi T at high-frequency (1.0π). See also Fig. 2a and Fig. D.2 for more results.

MSAs are low-pass filters, but Convs are highpass filters. As explained in Section 1.1, MSAs spatially smoothen feature maps with self-attention importances. Therefore, we expect that MSAs will tend to reduce high-frequency signals. See Appendix B for a more detailed discussion.

Figure 8 shows the relative log amplitude ( log amplitude) of Vi T s Fourier transformed feature map at high-frequency (1.0π) on Image Net. In this figure, MSAs almost always decrease the high-frequency amplitude, and MLPs corresponding to Convs increase it. The only exception is in the early stages of the model. In these stages, MSAs behave like Convs, i.e., they increase the amplitude. This could serve as an evidence for a hybrid model that uses Convs in early stages and MSAs in late stages (Guo et al., 2021; Graham et al., 2021; Dai et al., 2021; Xiao et al., 2021; Srinivas et al., 2021).

Based on this, we can infer that low-frequency signals and high-frequency signals are informative to MSAs and Convs, respectively. In support of this argument, we report the robustness of Vi T and Res Net against frequency-based random noise. Following Shao et al. (2021) and Park & Kim (2021), we measure the decrease in accuracy with respect to data with frequency-based random noise xnoise = x0 + F 1 (F(δ) Mf), where x0 is clean data, F( ) and F 1( ) are Fourier transform and inverse Fourier transform, δ is Gaussian random noise, and Mf is frequency mask.

Published as a conference paper at ICLR 2022

0.0 0.2 0.4 0.6 0.8 1.0 Normalized depth

Feature map variance

0.0 0.2 0.4 0.6 0.8 1.0 Normalized depth

Feature map variance

Figure 9: MSAs (gray area) reduce the variance of feature map points, but Convs (white area) increase the variance. The blue area is subsampling layer. This result implies that MSAs ensemble feature maps, but Convs do not.

As expected, the result in Fig. 2b reveals that Vi T and Res Net are vulnerable to low-frequency noise and high-frequency noise, respectively. Low-frequency signals and the high-frequency signals each correspond to the shape and the texture of images. The results thus suggests that MSAs are shape-biased (Naseer et al., 2021), whereas Convs are texture-biased (Geirhos et al., 2019).

MSAs aggregate feature maps, but Convs do not. Since MSAs average feature maps, they will reduce variance of feature map points. This suggests that MSAs ensemble feature maps (Park & Kim, 2021). To demonstrate this claim, we measure the variance of feature maps from NN layers.

Figure 9 shows the experimental results of Res Net and Vi T. This figure indicates that MSAs in Vi T tend to reduce the variance; conversely, Convs in Res Net and MLPs in Vi T increase it. In conclusion, MSAs ensemble feature map predictions, but Convs do not. As Park & Kim (2021) figured out, reducing the feature map uncertainty helps optimization by ensembling and stabilizing the transformed feature maps. See Fig. D.1 for more results on Pi T and Swin.

We observe two additional patterns for feature map variance. First, the variance accumulates in every NN layer and tends to increase as the depth increases. Second, the feature map variance in Res Net peaks at the ends of each stage. Therefore, we can improve the predictive performance of Res Net by inserting MSAs at the end of each stage. Furthermore, we also can improve the performance by using MSAs with a large number of heads in late stages.

4 HOW CAN WE HARMONIZE MSAS WITH CONVS?

Since MSAs and Convs are complementary, this section seeks to design a model that leverages only the advantages of the two modules. To this end, we propose the design rules described in Fig. 3c, and demonstrate that the models using these rules outperforms CNNs, not only in the large data regimes but also in the small data regimes, such as CIFAR.

4.1 DESIGNING ARCHITECTURE

We first investigate the properties of multi-stage NN architectures. Based on this investigation, we come to propose an alternating pattern, i.e., a principle for stacking MSAs based on CNNs.

Multi-stage NNs behave like individual models. In Fig. 9, we observe that the pattern of feature map variance repeats itself at every stages. This behavior is also observed in feature map similarities and lesion studies.

Figure 10a shows the representational similarities of Res Net and Swin on CIFAR-100. In this experiment, we use mini-batch CKA (Nguyen et al., 2021) to measure the similarities. As Nguyen et al. (2021) figured out, the feature map similarities of CNNs have a block structure. Likewise, we observe that the feature map similarities of multi-stage Vi Ts, such as Pi T and Swin, also have a block structure. Since vanilla Vi T does not have this structure (Bhojanapalli et al., 2021; Raghu et al., 2021), the structure is an intrinsic characteristic of multi-stage architectures. See Fig. D.3 for more detailed results of Vi T and Pi T.

Published as a conference paper at ICLR 2022

(a) Feature map similarity

0.00 0.25 0.50 0.75 1.00 Normalized depth

Accuracy (%)

0.00 0.25 0.50 0.75 1.00 Normalized depth

Accuracy (%)

(b) Accuracy of one-unit-removed model.

Figure 10: Multi-stage CNNs and Vi Ts behave like a series connection of small individual models. Left: The feature map similarities show the block structure of Res Net and Swin. E stands for stem/embedding and P for pooling (subsampling) layer. Right: We measure decrease in accuracy after removing one unit from the trained model. Accuracy changes periodically, and this period is one stage. White, gray, and blue areas are Conv/MLP, MSA, and subsampling layers, respectively.

Figure 10b shows the results of lesion study (Bhojanapalli et al., 2021), where one NN unit is removed from already trained Res Net and Swin during the testing phase. In this experiment, we remove one 3 3 Conv layer from the bottleneck block of Res Net, and one MSA or MLP block from Swin. In Res Net, removing an early stage layers hurts accuracy more than removing a late stage layers. More importantly, removing a layer at the beginning of a stage impairs accuracy more than removing a layer at the end of a stage. The case of Swin is even more interesting. At the beginning of a stage, removing an MLP hurts accuracy. At the end of a stage, removing an MSA seriously impairs the accuracy. These results are consistent with Fig. 8. See Fig. D.4 for the results on Vi T and Pi T.

Based on these findings, we expect MSAs closer to the end of a stage to significantly improve the predictive performance. This is contrary to the popular belief that MSAs closer to the end of a model improve the performance (Srinivas et al., 2021; d Ascoli et al., 2021; Graham et al., 2021; Dai et al., 2021).

Build-up rule. Considering all the insights, we propose the following design rules:

Alternately replace Conv blocks with MSA blocks from the end of a baseline CNN model.

If the added MSA block does not improve predictive performance, replace a Conv block located at the end of an earlier stage with an MSA block .

Use more heads and higher hidden dimensions for MSA blocks in late stages.

We call the model that follows these rules Alter Net. Alter Net unifies Vi Ts and CNNs by adjusting the ratio of MSAs and Convs as shown in Fig. 3. Figure 11 shows Alter Net based on pre-activation Res Net-50 (He et al., 2016b) for CIFAR-100 as an example. Figure D.5 shows Alter Net for Image Net.

Figure 12a reports the accuracy of Alter-Res Net-50, which replaces the Conv blocks in Res Net-50 with local MSAs (Liu et al., 2021) according to the aforementioned rules, on CIFAR-100. As expected, MSAs in the last stage (c4) significantly improve the accuracy. Surprisingly, an MSA in 2nd stage (c2) improves the accuracy, while two or more MSAs in the 3rd stage (c3) reduce it. In conclusion, MSAs at the end of a stage play an important role in prediction.

Figure 12c demonstrates that MSAs suppress large eigenvalues while allowing only a few negative eigenvalues. As explained in Fig. 4, large datasets compensate for the shortcomings of MSAs. Therefore, more data allows more MSAs for a models.

Published as a conference paper at ICLR 2022

Stage 1 (c1) Stage 2 (c2) Stage 3 (c3) Stage 4 (c4) Classifier

Figure 11: Detailed architecture of Alter-Res Net-50 for CIFAR-100. White, gray, and blue blocks mean Conv, MSA, and subsampling blocks. All stages (except stage 1) end with MSA blocks. This model is based on pre-activation Res Net-50. Following Swin, MSAs in stages 1 to 4 have 3, 6, 12, and 24 heads, respectively.

0 2 4 MSA number

Accuracy (%)

(a) Accuracy of Alter Net for MSA number

70.0 80.0 Accuracy (%)

Robustness (%)

Alter-Res Net

(b) Accuracy and robustness in a small data regime (CIFAR-100)

0 2000 Max eigenvalue

warmupth Res Net Alter Net

(c) Hessian max eigenvalue spectra in an early phase of training

Figure 12: Alter Net outperforms CNNs and Vi Ts. Left: MSAs in the late of the stages improve accuracy. We replace Convs of Res Net with MSAs one by one according to the build-up rules. c1 to c4 stands for the stages. Several MSAs in c3 harm the accuracy, but the MSA at the end of c2 improves it. Center: Alter Net outperforms CNNs even in a small data regime. Robustness is mean accuracy on CIFAR-100-C. RX is Res Ne Xt. Right: MSAs in Alter Net suppress the large eigenvalues; i.e., Alter Net has a flatter loss landscape than Res Net in the early phase of training.

4.2 PERFORMANCE

Figure 12b shows the accuracy and corruption robustness of Alter-Res Net-50 and other baselines on CIFAR-100 and CIFAR-100-C. Since CIFAR is a small dataset, CNNs outperforms canonical Vi Ts. Surprisingly, Alter-Res Net a model with MSAs following the appropriate build-up rule outperforms CNNs even in the small data regimes. This suggests that MSAs complement Convs. In the same manner, this simple modification shows competitive performance on larger datasets, such as Image Net. See Fig. E.1 for more details.

5 DISCUSSION

Our present work demonstrates that MSAs are not merely generalized Convs, but rather generalized spatial smoothings that complement Convs. MSAs help NNs learn strong representations by ensembling feature map points and flattening the loss landscape.

Since the main objective of this work is to investigate the nature of MSA for computer vision, we preserve the architectures of Conv and MSA blocks in Alter Net. Thus, Alter Net has a strong potential for future improvements. In addition, Alter Net can conveniently replace the backbone for other vision tasks such as dense prediction (Carion et al., 2020). As Park & Kim (2021) pointed out, global average pooling (GAP) for simple classification tasks has a strong tendency to ensemble feature maps, but NNs for dense prediction do not use GAP. Therefore, we believe that MSA to be able to significantly improve the results in dense prediction tasks by ensembling feature maps. Lastly, strong data augmentation for MSA training harms uncertainty calibration as shown in Fig. F.1a. We leave a detailed investigation for future work.

Published as a conference paper at ICLR 2022

ACKNOWLEDGEMENT

We thank the reviewers, Taeoh Kim, and Pilhyeon Lee for valuable feedback. This work was supported by the Samsung Science and Technology Foundation under Project Number SSTF-BA1501-52.

REPRODUCIBILITY STATEMENT

To ensure reproducibility, we provide comprehensive resources, such as code and experimental details. The code is available at https://github.com/xxxnell/how-do-vits-work. Appendix A.1 provides the specifications of all models used in this work. Detailed experimental setup including hyperparameters and the structure of Alter Net are also available in Appendix A.1 and Appendix E. De-facto image datasets are used for all experiments as described in Appendix A.1.

Irwan Bello. Lambdanetworks: Modeling long-range interactions without attention. In International Conference on Learning Representations, 2021.

Irwan Bello, Barret Zoph, Ashish Vaswani, Jonathon Shlens, and Quoc V Le. Attention augmented convolutional networks. In International Conference on Computer Vision, 2019.

Irwan Bello, William Fedus, Xianzhi Du, Ekin Dogus Cubuk, Aravind Srinivas, Tsung-Yi Lin, Jonathon Shlens, and Barret Zoph. Revisiting resnets: Improved training and scaling strategies. In Advances in Neural Information Processing Systems, 2021.

Srinadh Bhojanapalli, Ayan Chakrabarti, Daniel Glasner, Daliang Li, Thomas Unterthiner, and Andreas Veit. Understanding robustness of transformers for image classification. In International Conference on Computer Vision, 2021.

Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, 2020.

Xiangning Chen, Cho-Jui Hsieh, and Boqing Gong. When vision transformers outperform resnets without pretraining or strong data augmentations. In International Conference on Learning Representations, 2022.

Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In Advances in Neural Information Processing Systems, 2021.

Jean-Baptiste Cordonnier, Andreas Loukas, and Martin Jaggi. On the relationship between selfattention and convolutional layers. In International Conference on Learning Representations, 2020.

Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Advances in Neural Information Processing Systems, 2020.

Zihang Dai, Hanxiao Liu, Quoc V Le, and Mingxing Tan. Coatnet: Marrying convolution and attention for all data sizes. In Advances in Neural Information Processing Systems, 2021.

Stéphane d Ascoli, Hugo Touvron, Matthew Leavitt, Ari Morcos, Giulio Biroli, and Levent Sagun. Convit: Improving vision transformers with soft convolutional inductive biases. In International Conference on Machine Learning, 2021.

Yann N Dauphin, Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, Surya Ganguli, and Yoshua Bengio. Identifying and attacking the saddle point problem in high-dimensional non-convex optimization. In Advances in Neural Information Processing Systems, 2014.

Published as a conference paper at ICLR 2022

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021.

Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In International Conference on Learning Representations, 2021.

Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2019.

Behrooz Ghorbani, Shankar Krishnan, and Ying Xiao. An investigation into neural net optimization via hessian eigenvalue density. In International Conference on Machine Learning, 2019.

Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. ar Xiv preprint ar Xiv:1706.02677, 2017.

Ben Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet s clothing for faster inference. In International Conference on Computer Vision, 2021.

Jianyuan Guo, Kai Han, Han Wu, Chang Xu, Yehui Tang, Chunjing Xu, and Yunhe Wang. Cmt: Convolutional neural networks meet vision transformers. ar Xiv preprint ar Xiv:2107.06263, 2021.

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016a.

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European Conference on Computer Vision, 2016b.

Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019.

Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In International Conference on Computer Vision, 2021.

Jiri Hron, Yasaman Bahri, Jascha Sohl-Dickstein, and Roman Novak. Infinite attention: Nngp and ntk for deep attention networks. In International Conference on Machine Learning, 2020.

Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In European Conference on Computer Vision, 2016.

Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. In Advances in Neural Information Processing Systems, 2018.

Stanislaw Jastrzebski, Maciej Szymczak, Stanislav Fort, Devansh Arpit, Jacek Tabor, Kyunghyun Cho, and Krzysztof Geras. The break-even point on optimization trajectories of deep neural networks. In International Conference on Learning Representations, 2020.

Stanislaw Jastrzebski, Devansh Arpit, Oliver Astrand, Giancarlo B Kerg, Huan Wang, Caiming Xiong, Richard Socher, Kyunghyun Cho, and Krzysztof J Geras. Catastrophic fisher explosion: Early phase fisher matrix impacts generalization. In International Conference on Machine Learning, 2021.

Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. In International Conference on Learning Representations, 2017.

Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009.

Published as a conference paper at ICLR 2022

Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. In Advances in Neural Information Processing Systems, 2018.

Chaoyue Liu, Libin Zhu, and Mikhail Belkin. On the linearity of large non-linear models: when and why the tangent kernel is constant. In Advances in Neural Information Processing Systems, 2020.

Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In International Conference on Computer Vision, 2021.

Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations, 2017.

Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019.

Xiaofeng Mao, Gege Qi, Yuefeng Chen, Xiaodan Li, Ranjie Duan, Shaokai Ye, Yuan He, and Hui Xue. Towards robust vision transformer. ar Xiv preprint ar Xiv:2105.07926, 2021.

Paul Michel, Omer Levy, and Graham Neubig. Are sixteen heads really better than one? In Advances in Neural Information Processing Systems, 2019.

Matthias Minderer, Josip Djolonga, Rob Romijnders, Frances Hubis, Xiaohua Zhai, Neil Houlsby, Dustin Tran, and Mario Lucic. Revisiting the calibration of modern neural networks. In Advances in Neural Information Processing Systems, 2021.

Muzammal Naseer, Kanchana Ranasinghe, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In Advances in Neural Information Processing Systems, 2021.

Thao Nguyen, Maithra Raghu, and Simon Kornblith. Do wide and deep networks learn the same things? uncovering how neural network representations vary with width and depth. In International Conference on Learning Representations, 2021.

Namuk Park and Songkuk Kim. Blurs behave like ensembles: Spatial smoothings to improve accuracy, uncertainty, and robustness. ar Xiv preprint ar Xiv:2105.12639, 2021.

Namuk Park, Taekyu Lee, and Songkuk Kim. Vector quantized bayesian neural network inference for data streams. In Proceedings of the AAAI Conference on Artificial Intelligence, 2021.

Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, 2019.

Sayak Paul and Pin-Yu Chen. Vision transformers are robust learners. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022.

Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? In Advances in Neural Information Processing Systems, 2021.

Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Image Net Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 2015.

Shibani Santurkar, Dimitris Tsipras, Andrew Ilyas, and Aleksander M adry. How does batch normalization help optimization? In Advances in Neural Information Processing Systems, 2018.

Rulin Shao, Zhouxing Shi, Jinfeng Yi, Pin-Yu Chen, and Cho-Jui Hsieh. On the adversarial robustness of visual transformers. ar Xiv preprint ar Xiv:2103.15670, 2021.

Published as a conference paper at ICLR 2022

Aravind Srinivas, Tsung-Yi Lin, Niki Parmar, Jonathon Shlens, Pieter Abbeel, and Ashish Vaswani. Bottleneck transformers for visual recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021.

Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. ar Xiv preprint ar Xiv:2106.10270, 2021.

Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In International Conference on Computer Vision, 2017.

Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016.

Ilya Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, et al. Mlp-mixer: An all-mlp architecture for vision. In Advances in Neural Information Processing Systems, 2021.

Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, 2021.

Shikhar Tuli, Ishita Dasgupta, Erin Grant, and Thomas L Griffiths. Are convolutional neural networks or transformers more like human vision? In Proceedings of the Annual Meeting of the Cognitive Science Society, 2021.

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, 2017.

Peihao Wang, Wenqing Zheng, Tianlong Chen, and Zhangyang Wang. Scaling the depth of vision transformers via the fourier domain analysis. In International Conference on Learning Representations, 2022.

Phil Wang. Implementation of vision transformer. https://github.com/lucidrains/ vit-pytorch, 2021.

Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018.

Yeming Wen, Ghassen Jerfel, Rafael Muller, Michael W Dusenberry, Jasper Snoek, Balaji Lakshminarayanan, and Dustin Tran. Combining ensembles and data augmentation can harm your calibration. In International Conference on Learning Representations, 2021.

Ross Wightman. Pytorch image models. https://github.com/rwightman/ pytorch-image-models, 2019.

Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollár, and Ross Girshick. Early convolutions help transformers see better. In Advances in Neural Information Processing Systems, 2021.

Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017.

Weijian Xu, Yifan Xu, Tyler Chang, and Zhuowen Tu. Co-scale conv-attentional image transformers. In International Conference on Computer Vision, 2021.

Baosong Yang, Longyue Wang, Derek F Wong, Lidia S Chao, and Zhaopeng Tu. Convolutional self-attention networks. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), 2019.

Published as a conference paper at ICLR 2022

Zhewei Yao, Amir Gholami, Kurt Keutzer, and Michael W Mahoney. Pyhessian: Neural networks through the lens of the hessian. In 2020 IEEE International Conference on Big Data (Big Data), 2020.

Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A fourier perspective on model robustness in computer vision. In Advances in Neural Information Processing Systems, 2019.

Tan Yu, Xu Li, Yunfeng Cai, Mingming Sun, and Ping Li. Rethinking token-mixing mlp for mlp-based vision backbone. In BMVC, 2021a.

Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. ar Xiv preprint ar Xiv:2111.11418, 2021b.

Kun Yuan, Shaopeng Guo, Ziwei Liu, Aojun Zhou, Fengwei Yu, and Wei Wu. Incorporating convolution designs into visual transformers. In International Conference on Computer Vision, 2021.

Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In International Conference on Computer Vision, 2019.

Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018.

Richard Zhang. Making convolutional networks shift-invariant again. In International Conference on Machine Learning, 2019.

Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2020.

Published as a conference paper at ICLR 2022

A EXPERIMENTAL DETAILS

This section provides experimental details, e.g., setups and background information.

We obtain the main experimental results from two sets of machines for CIFAR (Krizhevsky et al., 2009). The first set consists of an Intel Xeon W-2123 Processor, 32GB memory, and a single Ge Force RTX 2080 Ti, and the other set of four Intel Intel Broadwell CPUs, 15GB memory, and a single NVIDIA T4. For Image Net (Russakovsky et al., 2015), we use AMD Ryzen Threadripper 3960X 24-Core Processor, 256GB memory, and four Ge Force RTX 2080 Ti. NN models are implemented in Py Torch (Paszke et al., 2019).

We train NNs using categorical cross-entropy (NLL) loss and Adam W optimizer (Loshchilov & Hutter, 2019) with initial learning rate of 1.25 10 4 and weight decay of 5 10 2. We also use cosine annealing scheduler (Loshchilov & Hutter, 2017). NNs are trained for 300 epochs with a batch size of 96 on CIFAR, and a batch size of 128 on Image Net. The learning rate is gradually increased (Goyal et al., 2017) for 5 epochs. Following Touvron et al. (2021), strong data augmentations such as Rand Augment (Cubuk et al., 2020), Random Erasing (Zhong et al., 2020), label smoothing (Szegedy et al., 2016), mixup (Zhang et al., 2018), and Cut Mix (Yun et al., 2019) are used for training. Stochastic depth (Huang et al., 2016) is also used to regularize NNs. This Dei T-style configuration, which significantly improves the performance (Steiner et al., 2021; Bello et al., 2021), is the de facto standard in Vi T training (See, e.g., (Heo et al., 2021; Liu et al., 2021)). Therefore, we believe the insights presented in this paper can be used widely. See source code (https://github.com/xxxnell/how-do-vits-work) for detailed configurations.

We mainly report the performances of Res Net-50, Vi T-Ti, Pi T-Ti, and Swin-Ti. Their training throughputs on CIFAR-100 are 320, 434, 364, and 469 image/sec, respectively, which are comparable to each other. Figures 5a and C.1a report the predictive performance of Res Ne Xt-50 (Xie et al., 2017), Twins-S (Chu et al., 2021), and MLP-Mixer-Ti (Tolstikhin et al., 2021). Figure E.1 additionally reports the performance of Con Vi T-Ti (d Ascoli et al., 2021), Le Vi T-128S (Graham et al., 2021), and Coa T-Lite-Ti (Xu et al., 2021). We use a patch size of 2 2 for Vi T and Pi T on CIFAR; for Swin, a patch size of 1 1 and a window size of 4 4. We use a patch size of 4 4 for Vi T only in Fig. 7. We halve the depth of the Vi T in Fig. C.5 and Fig. C.6 due to the memory limitation.

All models for CIFAR, and Res Net, Vi T, and Alter Net for Image Net are trained from scratch. We use pertained Pi T and Swin from Wightman (2019) for Image Net. The implementations of Vision Transformers are based on Wightman (2019) and Wang (2021).

For Hessian max eigenvalue spectrum (Park & Kim, 2021), 10% of the training dataset is used. We also use power iteration with a batch size of 16 to produce the top-5 largest eigenvalues. To this end, we use the implementation of Yao et al. (2020). We modify the algorithm to calculate the eigenvalues with respect to ℓ2 regularized NLL on augmented training datasets. In the strict sense, the weight decay is not ℓ2 regularization, but we neglect the difference.

For the Fourier analysis and the feature map variance experiment, the entire test dataset is used. We report the amplitudes and the variances averaged over the channels.

A.2 BACKGROUND INFORMATION

Below are the preliminaries and terms of our experiments.

Test error and training NLL. We report test errors on clean test datasets and training NLLs on augmented training datasets in experiments, e.g., Fig. 5 and Fig. C.1. NLL is an appropriate metric for evaluating convergence on a training dataset because an NN optimizes NLL. In addition, it is the most widely used as a proper scoring rule indicating both accuracy and uncertainty. To represent predictive performance on a test dataset, we use a well-known metric: error. Although NLL can also serve the same purpose, results are consistent even when NLL is employed.

If an additional inductive bias or a learning technique improves the performance of an NN, this is either a method to help the NNs learn strong representations , or a method to regularize it.

Published as a conference paper at ICLR 2022

An improved i.e., lower training NLL suggests that this bias or technique helps the NN learn strong representations. Conversely, a compromised training NLL indicates that the bias or technique regularizes the NN. Likewise, we say that an NN overfits a training dataset when a test error is compromised as the training NLL is improved.

Hessian max eigenvalue spectrum. Park & Kim (2021) proposed Hessian max eigenvalue spectra , a feasible method for visualizing Hessian eigenvalues of large-sized NNs for real-world problems. It calculates and gathers top-k Hessian eigenvalues by using power iteration mini-batch wisely. Ghorbani et al. (2019) visualized the Hessian eigenvalue spectrum by using the Lanczos quadrature algorithm for full batch. However, this is not feasible for practical NNs because the algorithm requires a lot of memory and computing resources.

A good loss landscape is a flat and convex loss landscape. Hessian eigenvalues indicate the flatness and convexity of losses. The magnitude of Hessian eigenvalues shows sharpness, and the presence of negative Hessian eigenvalues shows non-convexity. Based on these insights, we introduce a negative max eigenvalue proportion (NEP, the lower the better) and an average of positive max eigenvalues (APE, the lower the better) to quantitatively measure the non-convexity and the sharpness, respectively. For a Hessian max eigenvalue spectrum p(λ), NEP is the proportion of negative eigenvalues R 0 p(λ) dλ, and APE is the expected value of positive eigenvalues R 0 λ p(λ) dλ / R 0 p(λ) dλ. We use these metrics in Fig. C.5 and Fig. C.6.

Note that measuring loss landscapes and Hessian eigenvalues without considering a regularization on clean datasets would lead to incorrect results, since NN training optimizes ℓ2 regularized NLL on augmented training datasets not NLL on clean training datasets. We visualize loss landscapes and Hessian eigenvalues with respect to ℓ2 regularized NLL loss on augmented training datasets .

Fourier analysis of feature maps. We analyze feature maps in Fourier space to demonstrate that MSA is a low-pass filter as shown in Fig. 2, Fig. 8, and Fig. D.2. Fourier transform converts feature maps into frequency domain. We represent these converted feature maps on normalized frequency domain, so that the highest frequency components are at f = { π, +π}, and the lowest frequency components are at f = 0. We mainly report the amplitude ratio of high-frequency components and low-frequency components by using log amplitude, the difference in log amplitude at f = π and f = 0. Yin et al. (2019) also analyzed the robustness of NNs from a Fourier perspective, but their research focused on input images not feature maps in Fourier spaces.

B MSAS BEHAVE LIKE SPATIAL SMOOTHINGS

As mentioned in Section 1.1, spatial smoothings before subsampling layers help CNNs see better (Zhang, 2019; Park & Kim, 2021). Park & Kim (2021) showed that such improvement in performance is possible due to spatial ensembles of feature map points. To this end, they used the (Bayesian) ensemble average of predictions for proximate data points (Park et al., 2021), which exploits data uncertainty (i.e., a distribution of feature maps) as well as model uncertainty (i.e., a posterior probability distribution of NN weights):

p(zj|xj, D) X

i π(xi|xj) p(zj|xi, wi) (2)

where π(xi|xj) is the normalized importance weight of a feature map point xi with respect to another feature map point xj, i.e., P

i π(xi|xj) = 1. This importance is defined as the similarity between xi and xj. p(zj|xi, wi) and p(zj|xj, D) stand for NN prediction and output predictive distribution, respectively. wi is the NN weight sample from the posterior p(w|D) with respect to the training dataset D. Put shortly, Eq. (2) spatially complements a prediction with other predictions based on similarities between data points. For instance, a 2 2 box blur spatially ensembles four neighboring feature map points, each with 1 4 of the same importance.

We note that the formulations for self-attention and the ensemble averaging for proximate data points are identical. The Softmax term and V in Eq. (1) exactly correspond to π(xi|xj) and p(zj|xi, wi) in Eq. (2). The weight samples in Eq. (2) is correspond to the multi-heads of MSAs (See also (Hron et al., 2020)).

Published as a conference paper at ICLR 2022

RX R Twins Swin Pi T Vi T Mixer

CNN Local MSA Global MSA

Error NLLtrain

(a) CIFAR-10

R Twins Swin Pi T Vi T

CNN Local MSA Global MSA

(b) Image Net

Figure C.1: The lower the training NLL, the lower the test error. R is Res Net and RX is Res Ne Xt. Left: In small data regimes, such as CIFAR-10 and CIFAR-100 (Fig. 5a), the cons of MSAs outweigh their pros; i.e., the non-convex losses disturb Vi T optimization. Right: Large datasets convexify the loss functions. Therefore, the pros of MSAs outweigh their cons in large data regimes; i.e., MSAs help NNs learn strong representations by flattening the loss landscapes.

Likewise, the properties of spatial smoothing are the same as those of MSAs (Park & Kim, 2021):

1 Spatial smoothing improves the accuracy of CNNs. In addition, spatial smoothing predicts well-calibrated uncertainty. 2 Spatial smoothing is robust against MC dropout (which is equivalent to image occlusion), data corruption, and adversarial attacks, and particularly robust against highfrequency noise. 3 Spatial smoothing layers closer to the output layer significantly improves the predictive performance. In addition, concurrent works suggest that MSA blocks behave like a spatial smoothing. Wang et al. (2022) provided a proof that Softmax-normalized matrix is a low-pass filter, although this does not guarantee that MSA blocks will behave like low-pass filters. Yu et al. (2021b) demonstrated that the MSA layers of Vi T can be replaced with average pooling layers.

Taking all these observations together, we provide an explanation of how MSAs work by addressing themselves as a general form of spatial smoothing or an implementation of ensemble averaging for proximate data points. Spatial smoothing improves performance in the following ways (Park & Kim, 2021): 1 Spatial smoothing helps in NN optimization by flattening the loss landscapes. Even a small 2 2 box blur filter significantly improves performance. 2 Spatial smoothing is a low-pass filter. CNNs are vulnerable to high-frequency noises, but spatial smoothing improves the robustness against such noises by significantly reducing these noises. 3 Spatial smoothing is effective when applied at the end of a stage because it aggregates all transformed feature maps. This paper empirically shows that these mechanisms also apply to MSAs.

C VITS FROM A LOSS LANDSCAPE PERSPECTIVE

This section provides further explanations of the analysis in Section 2.

The lower the NLL on the training dataset, the lower the error on the test dataset. Figure 5a demonstrates that low training NLLs result in low test errors on CIFAR-100. The same pattern can be observed on CIFAR-10 and Image Net as shown in Fig. C.1.

In small data regimes, such as CIFAR-10 (Fig. C.1a) and CIFAR-100 (Fig. 5a), both the error and the NLLtrain of Vi Ts are inferior to those of CNNs. This suggests that the cons of MSAs outweigh their pros. As discussed in Fig. 4, Vi Ts suffers from the non-convex losses, and these non-convex losses disturb Vi T optimization.

In large data regimes, such as Image Net (Fig. C.1b), both the error and the NLLtrain of Vi Ts with local MSAs are superior to those of CNNs. Since large datasets convexify the loss functions as discussed in Fig. 4, the pros of MSAs outweigh their cons. Therefore, MSAs help NNs learn strong representations by flattening the loss landscapes.

Rigorous discussion on the regularization of CNN s inductive bias. In Fig. 5a, we compare models of similar sizes, such as Res Net-50 and Vi T-Ti. Through such comparison, we show that a weak inductive bias hinders NN training, and that inductive biases of CNNs inductive bias of Convs and multi-stage architecture help NNs learn strong representations. However, inductive biases of

Published as a conference paper at ICLR 2022

2 2 4 4 8 8 Patch size

Error NLLtrain

(a) Error and NLLtrain of Vi T for patch size

100 50 0 Max Eigenvalue

2 2 4 4 8 8

0 100 Max Eigenvalue

(b) Negative and positive Hessian max eigenvalue spectra in early phase of training

Figure C.2: A small patch size does not guarantee better performance. We analyze Vi Ts with three embedded patch sizes: 2 2, 4 4, and 8 8. Note that every MSA has a global receptive fields. Left: As expected, a large patch size harms the performance, but surprisingly, the same is observed from a small patch size. Right: A small patch size, or a weak inductive bias, produces negative eigenvalues. This is another evidence that a weak inductive bias hinders NN optimization. On the other hand, MSAs with a small patch size reduce the magnitude of eigenvalues because they ensemble a large number of feature map points. Performance is optimized when these two effects are balanced.

CNNs produce better test accuracy for the same training NLL, i.e., Convs somewhat regularize NNs. We analyze two comparable models in terms of NLLtrain on CIFAR-100. The NLLtrain of Res Net-18, a model smaller than Res Net-50, is 2.31 with an error of 22.0%. The NLLtrain of Vi T-S, a model larger than Vi T-Ti, is 2.17 with an error of 30.4%. In summary, the inductive biases of CNNs improve accuracy for similar training NLLs.

Most of the improvements come from the multi-stage architecture, not the inductive bias of Convs. The NLLtrain of the Pi T-Ti, a multi-stage Vi T-Ti, is 2.29 with an error of 24.1 %. The accuracy of Pi T is only 1.9 percent point lower than that of Res Net. In addition, the small receptive field also regularizes Vi T. See Fig. 7.

75 150 300 600 1200 Epochs

Error NLLtrain

Figure C.3: A large number of epochs does not make Vi T overfit the training dataset of CIFAR. Solid line is the predictive performance of Vi T and dashed line is that of Res Net.

Vi T does not overfit a small training dataset even with a large number of epochs. Figure 5b shows that Vi T does not overfit small training datasets, such as CIFAR. The same phenomenon can be observed in Vi T training with a large number of epochs.

In Fig. C.3, we train Vi T and Res Net for 75, 150, 300, 600, and 1200 epochs. Results show that both NLLtrain and error decrease as the number of epochs increases. The predictive performances of Vi T are inferior to those of Res Net across all ranges of epochs.

A smaller patch size does not always imply better results. Vi T splits image into multiple patches. The smaller the patch size, the greater the flexibility of expression and the weaker the inductive bias. By analyzing Vi T with three patch sizes 2 2, 4 4, and 8 8 we demonstrate once again that a weak inductive bias disturbs NN optimization.

Figure C.2a shows the error on the test dataset and the NLL on the training dataset of CIFAR-100. As expected, a large patch size harms the performance on both datasets. Surprisingly, however, a small patch size also shows the same result. As such, appropriate patch sizes help Vi T learn strong representations and do not regularize Vi T.

The Hessian max eigenvalue spectra in Fig. C.2b explain this observation. Results reveal that a small patch size reduces the magnitude of Hessian eigenvalues but produces negative Hessian eigenvalues. In other words, the weak inductive bias makes loss landscapes flat yet non-convex. A large patch size suppresses negative eigenvalues. On the other hand, it not only limits the model expression but also sharpens loss landscapes. Performance is optimized when these two effects are balanced.

Published as a conference paper at ICLR 2022

Res Net Vi T Pi T Swin

(a) Loss (NLL +ℓ2) landscape visualizations

750 500 250 0

Res Net Vi T Pi T Swin

750 500 250 0

Res Net Vi T Pi T Swin

(b) Negative and positive Hessian max eigenvalue spectra in early phase (left) and late phase (right) of training

Figure C.4: A multi-stage architecture (in Pi T) and a local MSA (in Swin) also flatten the loss landscapes. Top: Pi T has a flatter loss landscape than Vi T near the optimum. Swin has an almost perfectly smooth parabolic loss landscape, which leads to better NN optimization. Bottom: A multistage architecture in Pi T suppresses negative Hessian eigenvalues. A local MSA in Swin produces negative eigenvalues, but significantly reduces the magnitude of eigenvalues.

A multi-stage architecture in Pi T and a local MSA in Swin also flatten loss landscapes. As explained in Fig. 1, an MSA smoothens loss landscapes. Similarly, a multi-stage architecture in Pi T and local MSA in Swin also help NN learn strong representations by smoothing the loss landscapes.

Figure C.4 provides loss landscape visualizations and Hessian eigenvalue spectra of Res Net, Vi T, Pi T, and Swin. Figure C.4a visualizes the global geometry of the loss functions. The loss landscapes of Pi T is flatter than that of Vi T near the optimum. Since Swin has more parameters than Vi T and Pi T, ℓ2 regularization determines the loss landscapes. All the loss surfaces of Vi Ts are smoother than that of Res Net. Figure C.4b shows the local geometry of the loss functions by using Hessian eigenvalues. In the early phase of training, a multi-stage architecture in Pi T helps training by suppressing negative Hessian eigenvalues. A local MSA in Swin produces negative eigenvalues, but significantly reduces the magnitude of eigenvalues. Moreover, the magnitude of Swin s Hessian eigenvalue does not significantly increases in the late phase of learning.

A lack of heads may lead to non-convex losses. Neural tangent kernel (NTK) (Jacot et al., 2018) theoretically implies that the loss landscape of a Vi T is convex and flat when the number of heads or the number of embedding dimensions per head goes to infinity (Hron et al., 2020; Liu et al., 2020). In particular, Liu et al. (2020) suggests that ||H|| O(1/ m) where ||H|| is the Hessian spectral norm and m is the number of heads or the number of embedding dimensions per head. Therefore, in practical situations, insufficient heads may cause non-convex and sharp losses.

Fig. C.5 empirically show that a lot of heads in MSA convexify and flatten the loss landscapes (cf. Michel et al. (2019)). In this experiment, we use NEP and APE to measure the non-convexity and the sharpness as introduced in Appendix A.2. Results show that both NEP and APE decrease as the number of heads increases. Likewise, Fig. C.6 shows that high embedding dimensions per head also convexify and flatten losses. The exponents of APE are 0.562 for the number of heads and 0.796 for the number of embedding dimensions, which are in close agreement with the value predicted by the theory of 1/2.

Large models have a flat loss in the early phase of training. Figure C.7 analyzes the loss landscapes of large models, such as Res Net-101 and Vi T-S. As shown in Fig. C.7a, large models explore

Published as a conference paper at ICLR 2022

1 2 3 4 5 6 7 89 Head number

(a) NEP (non-convexity) and APE (sharpness) of head number

100 0 Max Eigenvalue

H = 1 H = 3 H = 9

0 100 200 Max Eigenvalue

(b) Hessian negative and positive max eigenvalue spectra in early phase of training

Figure C.5: Multi-heads convexify and flatten loss landscapes. Left: We use negative max eigenvalue proportion (NEP) and average of positive max eigenvalues (APE) to quantify, respectively, the non-convexity and sharpness of loss landscapes. As the number of heads increases, loss landscapes become more convex and flatter. Right: Hessian max eigenvalue spectra also show that multi-head suppress negative eigenvalues and reduce the magnitude of eigenvalues.

22 66 110 198 Embed dim / Head

(a) NEP (non-convexity) and APE (sharpness) of embedding dim

100 0 Max Eigenvalue

Embed = 22 Embed = 66 Embed = 198

0 100 200 Max Eigenvalue

(b) Hessian negative and positive max eigenvalue spectra in early phase of training

Figure C.6: High embedding dimensions per head convexify and flatten the loss landscape. Left: As the number of embedding dimensions per head increases, loss landscapes become more convex and flat. Right: Hessian max eigenvalue spectra also show that high embedding dimensions suppress negative eigenvalues and reduce the magnitude of eigenvalues as shown in Fig. C.5.

low NLLs. This can be a surprising because loss landscapes of large models are globally sharp as shown in Fig. C.7b.

The Hessian eigenvalue spectra in Fig. C.7c provide a solution to the problem: Hessian eigenvalues of large models are smaller than those of small models in the early phase of training. This indicates that large models have flat loss functions locally.

D VITS FROM A FEATURE MAP PERSPECTIVE

This section provides further explanations of the analysis in Section 3 and Section 4.1.

MSAs in Pi T and Swin also ensemble feature maps. In Fig. 9, we show that MSAs in Vi T reduce feature map variances. The same pattern can be observed in Pi T and Swin. Figure D.1 demonstrates that MSAs in Pi T and Swin also reduce the feature map variances, suggesting that they also ensemble feature maps. One exception is the 3rd stage of Swin. MSAs suppresses the increase in variance at the beginning of the stage, but not at the end of the stage.

MSAs in Pi T and Swin are also low-pass filters. As discussed in Fig. 8, MSAs in Vi Ts are lowpass filters, while MLPs in Vi T and Convs in Res Net are high-pass filters. Likewise, we demonstrate that MSAs in Pi T and Swin are also low-pass filters.

Published as a conference paper at ICLR 2022

Res Net-50 Res Net-101 Vi T-Ti Vi T-S

(a) NLL landscape visualizations

Res Net-50 Res Net-101 Vi T-Ti Vi T-S

(b) Loss (NLL +ℓ2) landscape visualizations

750 500 250 0

Res Net-50 Res Net-101 Vi T-Ti Vi T-S

750 500 250 0

Res Net-50 Res Net-101 Vi T-Ti Vi T-S

(c) Negative and positive Hessian max eigenvalue spectra in early phase (left) and late phase (right) of training

Figure C.7: Loss landscapes of large models. Res Net-50 and Res Net-101 are comparable to Vi T-Ti and Vi T-S, respectively. Top: Large models explore low NLLs. Middle: Loss landscape visualizations show that the global geometry of large models is sharp. Bottom: The Hessian eigenvalues of large models are smaller than those of small models. This suggests that large models have a flat local geometry in the early phase of training, and that this flat loss helps NNs learn strong representations. In the late phase of training, large Vi Ts have flat minima while large Res Net has a sharp minimum.

Published as a conference paper at ICLR 2022

0.0 0.2 0.4 0.6 0.8 1.0 Normalized depth

Feature map variance

0.0 0.2 0.4 0.6 0.8 1.0 Normalized depth

Feature map variance

0.0 0.2 0.4 0.6 0.8 1.0 Normalized depth

Feature map variance

0.0 0.2 0.4 0.6 0.8 1.0 Normalized depth

Feature map variance

Figure D.1: MSAs in Pi T and Swin also reduce feature map variance except in 3rd stage of Swin. White, gray, and blue areas are Conv/MLP, MSA, and subsampling layers, respectively.

Figure D.2 shows the relative log amplitude of Fourier transformed feature maps. As in the case of Vi T, MSAs in Pi T and Swin generally decrease the amplitude of high-frequency signals; in contrast, MLPs increases the amplitude.

Multi-stage Vi Ts have a block structures. Feature map similarities of CNNs shows a block structure (Nguyen et al., 2021). As Raghu et al. (2021) pointed out, Vi Ts have a uniform representations across all layers. By investigating multi-stage Vi Ts, we demonstrate that subsampling layers create a characteristic block structure of the representation. See Fig. D.3.

Convs at the beginning of a stage and MSAs at the end of a stage play an important role. Figure D.4 shows the results of a lesion study for Res Net and Vi Ts. In this experiment, we remove one 3 3 Conv layer from the bottleneck block of a Res Net, and one MSA or MLP block from Vi Ts. Consistent results can be observed for all models: Removing Convs at the beginning of a stage and MSAs at the end of a stage significantly harm accuracy. As a result, the accuracy varies periodically.

E EXTENDED INFORMATION OF ALTERNET

This section provides further informations on Alter Net.

Detailed architecture of Alter Net. Section 4 introduces Alter Net to harmonize Convs with MSAs. Since most MSAs take pre-activation arrangements, pre-activation Res Net is used as a baseline for consistency. We add one CNN block to the last stage of Res Net to make the number of blocks even. A local MSA with relative positional encoding from Swin is used for Alter Net. However, for simplicity of implementation, we do not implement detailed techniques, such as a cyclic shift and layer-specific initialization. For CIFAR, the patch size of the MSA is 1 1 and the window size is 4 4. If all Conv blocks are alternately replaced with MSA, Alter Net becomes a Swin-like model.

In order to achieve better performance, NNs should strongly aggregate feature maps at the end of models as discussed in Section 3 and Section 4. To this end, Alter Net use 3, 6, 12, 24 heads for MSAs in each stage.

Published as a conference paper at ICLR 2022

0.0 0.5 1.0 Frequency

Log amplitude

0.0 0.2 0.4 0.6 0.8 1.0 Normalized depth

Log amplitude

0.0 0.5 1.0 Frequency

Log amplitude

0.0 0.2 0.4 0.6 0.8 1.0 Normalized depth

Log amplitude

0.0 0.5 1.0 Frequency

Log amplitude

0.0 0.2 0.4 0.6 0.8 1.0 Normalized depth

Log amplitude

0.0 0.5 1.0 Frequency

Log amplitude

0.0 0.2 0.4 0.6 0.8 1.0 Normalized depth

Log amplitude

0.0 0.2 0.4 0.6 0.8 1.0 Normalized depth

Figure D.2: MSAs in Pi T and Swin also reduce high-frequency signals. Left: log amplitude of Fourier transformed feature map. We only provide the diagonal components. Right: The highfrequency (1.0π) log amplitude. White, gray, and blue areas are Conv/MLP, MSA, and subsampling layers, respectively.

Published as a conference paper at ICLR 2022

Figure D.3: Multi-stage Vi Ts have block structures in representational similarities. Block structures can be observed in all multi-stage NNs, namely, Res Net, Pi T, and Swin. E is the stem/embedding and P is the pooling (subsampling) layer.

0.00 0.25 0.50 0.75 1.00 Normalized depth

Accuracy (%)

0.00 0.25 0.50 0.75 1.00 Normalized depth

Accuracy (%)

0.00 0.25 0.50 0.75 1.00 Normalized depth

Accuracy (%)

0.00 0.25 0.50 0.75 1.00 Normalized depth

Accuracy (%)

Figure D.4: Lesion study shows that Convs at the beginning of a stage and MSAs at the end of a stage are important for prediction. We measure the decrease in accuracy after removing one unit from the trained model. In this experiment, we can observe that accuracy changes periodically. The white, gray, and blue areas are Convs/MLPs, MSAs, and subsampling layers, respectively.

Stage 1 (c1) Stage 2 (c2) Stage 3 (c3) Stage 4 (c4) Classifier

Figure D.5: Detailed architecture of Alter-Res Net-50 for Image Net-1K. The white, gray, and blue blocks each represent Convs, MSAs, and subsampling blocks. This model alternately replaces Conv blocks with MSA blocks from the end of a stage. Following Swin, MSAs in stages 1 to 4 use 3, 6, 12, and 24 heads, respectively. We use 6 MSA blocks for Image Net since large amounts of data alleviates the drawbacks of MSA. See Fig. 11 for comparison with the model for CIFAR-100, which uses 4 MSA blocks.

Published as a conference paper at ICLR 2022

20 40 60 80 Confidence (%)

Accuracy (%)

Res Net Res Net + Aug Vi T Vi T + Aug

(a) Reliability diagram

200 100 0 Max Eigenvalue

Res Net Res Net + Aug Vi T Vi T + Aug

0 1000 Max Eigenvalue

(b) Hessian max eigenvalue spectrum

Figure F.1: Distinctive properties of strong data augmentation. Aug stands for strong data augmentation. Left: Strong data augmentation makes predictions underconfident on CIFAR-100. The same phenomenon can be observed on Image Net-1K. Right: Strong data augmentation significantly reduces the magnitude of Hessian max eigenvalues. This means that the data augmentation helps NNs converge to better optima by flattening the loss landscapes. On the other hand, strong data augmentation produces a lot of negative Hessian eigenvalues, i.e., it makes the losses non-convex.

70.0 75.0 80.0 Accuracy (%)

Robustness (%)

Alter-Res Net

Con Vi T Le Vi T

Figure E.1: MSA with the appropriate build-up rules significantly improves Res Net on Image Net. Robustness is mean accuracy on Image Net-C. RX is Res Ne Xt.

The computational costs of Conv blocks and MSA blocks are almost identical. The training throughput of Alter-Res Net-50 is 473 image/sec on CIFAR-100, which is 20% faster than that of pre-activation Res Net-50.

The optimal number of MSAs depends on the model and dataset, so we empirically determine the number of MSAs as shown in Fig. 12a. A large dataset allows a large number of MSAs. For Image Net, we use 6 MSAs as shown in Fig. D.5, because a large datasets alleviates the shortcomings of MSAs.

MSAs improve the performance of CNNs on Image Net. Since MSAs complement Convs, MSAs improve the predictive performance of CNNs when appropriate build-up rules are applied as shown in Section 4.1. Figure E.1 illustrates the accuracy and robustness mean accuracy on Image Net-C of CNNs and Vi Ts on Image Net-1K. Since Image Net is a large dataset, a number of Vi Ts outperform CNNs. MSAs with the appropriate build-up rules significantly improves Res Net, and the predictive performance of Alter Net is on par with that of Swin in terms of accuracy without heavy modifications, e.g., the shifted windowing scheme (Liu et al., 2021). Alter Net is easy-to-implement and has a strong potential for future improvements. In addition, the build-up rules not only improve Res Net, but also other NNs, e.g., vanilla post-activation Res Net and Res Ne Xt; but we do not report this observation in order to keep the visualization simple.

F DISTINCTIVE PROPERTIES OF DATA AUGMENTATION

This section empirically demonstrates that NN training with data augmentation is different from training on large datasets. We compare Dei T-style strong data augmentation with weak data augmentation, i.e., resize and crop. In this section, a result without data augmentation stands for a result only with weak data augmentation .

F.1 DATA AUGMENTATION CAN HARM UNCERTAINTY CALIBRATION

Figure F.1a shows a reliability diagram of NNs with and without strong augmentation on CIFAR-100. Here, both Res Net and Vi T without data augmentation (i.e., only with weak data augmentation) predict overconfident results. We show that strong data augmentation makes the predictive results underconfident (cf. Wen et al. (2021)). These are unexpected results because the predictions without data augmentation on large datasets, such as Image Net, are not under-confident. A detailed investigation remains for future work.

Published as a conference paper at ICLR 2022

F.2 DATA AUGMENTATION REDUCES THE MAGNITUDE OF HESSIAN EIGENVALUES

How does data augmentation help an MSA avoid overfitting on a training dataset and achieve better accuracy on a test dataset? Figure F.1b shows the Hessian max eigenvalue spectrum of NNs with and without strong data augmentation. First of all, strong data augmentation reduces the magnitude of Hessian eigenvalues, i.e., data augmentation flattens the loss landscapes in the early phase of training. These flat losses leads to better generalization. On the other hand, strong data augmentation produces a lot of negative Hessian eigenvalues, i.e., data augmentation makes the losses non-convex. This prevents NNs from converging to low losses on training datasets. It is clearly different from the effects of large datasets discussed in Fig. 4 large datasets convexify the loss landscapes. A detailed investigation remains for future work.