Deep neural networks (DNN) are complex learning systems, which have been used in a variety of tasks with great success in recent times. In some fields, like visual recognition or playing games, DNNs can compete with or even outperform their human counterparts [10, 22], showcasing their utility and effectiveness.
In real-world applications, however, there are a number of quality criteria that go beyond single scalar performance metrics such as classification accuracy that are typically considered when comparing different classification algorithms. These quality criteria include interpretability , the quantification of uncertainty  but also robustness in a general sense. The aspect of robustness comprises both robustness against label noise, see  for a review, and robustness against any kind of input noise. In this work we are only concerned with robustness in the latter sense, which we further sub-categorize based on the kind of input perturbations under consideration. We distinguish on the one hand noise distorsions that include for example Gaussian noise, JPEG compression artifacts but also adversarial examples  and on the other hand transformative distorsions comprising image transformations such as rotations and crops. A particular challenge arises from the fact that the test data might exhibit distortions that have not been encountered by the network during training both in terms of distorsion strength as well as in terms of distorsion types. Therefore we strive to develop methods that ideally lead to generalizable robustness beyond distorsions seen during training.
One way of increasing the robustness against input perturbations is to use data augmentation (DA) . In fact, DA by adding perturbed copies of existing data samples is by now an established method and has been shown to greatly increase the generalizability of a given model [29, 26]. DA has two aspects: On the one hand, it enlarges the available training data to achieve a better generalization performance and on the other hand it increases the model’s robustness against transformations used for DA. As we will see, this robustness, however, is highly specific to the kind and the particular characteristics of the perturbations used for DA. In particular, in the worst case DA can degrade the model performance for unperturbed inputs compared to the baseline model performance.
An alternative approach, which will be explored further in this work, was put forward under the name of stability training (ST) by Zheng et al . Instead of adding distortion instances to the training corpus, ST feeds the perturbed and the reference sample to the network simultaneously and introduces a consistency constraint as additional optimization objective that tries to align the network’s outputs of the perturbed image and the reference sample. By extending the original work beyond Gaussian noise perturbation considered in the original work , we propose ST as a general-purpose alternative to DA that offers similar advantages with limited negative side effects. We also introduce modifications to the method, which mitigate weaknesses and extend its applicability. In particular, our contributions can be summarized as follows:
We establish stability training as a competitive alternative to data augmentation, which produces comparable or superior robustness improvements across a wide range of distortion types, while exhibiting significantly lower risk to deteriorate performance compared to an unstabilized baseline model. To this end, we present a detailed analysis of the robustness of both data augmentation and stability training when trained/tested on specific distorsion types.
We propose a symmetrical stability objective that increases the method’s effectiveness in learning from data transformations, like rotations, that do not distinguish an unperturbed reference sample. The modified objective offers superior performance to data augmentation in scenarios that are likely to be encountered in real-world applications.
We evaluate stability training as an alternative to adversarial training to increase robustness against adversarial examples generated via the fast gradient sign method .
As an outlook we demonstrate the prospects of using multiple distorsion types simultaneously to further improve the robustness properties.
2 Stability Training
Stability training aims to stabilize predictions of a deep neural network in response to small input distortions. The idea behind this approach is that an input that is similar to and semantically equivalent ought to produce similar outputs , where
denotes the trainable parameters of the neural network. The full optimization objective is then defined as a composite loss function of the original task, for example cross-entropy between the network’s prediction
and the (one-hot encoded) ground truth label, and a separate stability objective which enforces the consistency constraint. More explicitly, given an original image and a perturbed version of it, the combined training objective is then given by
where the stability loss is defined via
and the hyperparameter adjusts the relative importance of the two loss components.
The choice of a distance function is task-specific. Whereas for regression tasks the -distance is a straightforward choice, in a classification setting with classes, where
the Kullback-Leibler (KL) divergence as considered in  represents a canonical way of comparing likelihoods of original and distorted inputs
The stability loss function is then given by
As the KL-divergence is not symmetric with respect to its arguments, using it as distance measure is most appropriate in situations, where the reference sample can be clearly distinguished as undistorted from the modified copy, as it is the case for distortions we categorize as coming from the noise category. For transformative distorsions such as rotations it turns out to be beneficial to consider a symmetrized stability term, i.e.
At this point it is worthwhile to point out a crucial difference between stability training and DA: In the DA-setting the neural network is trained on perturbed input samples while the loss is the original loss evaluated using the label of the reference image. On the contrary, stability training decouples solving the original task from stabilizing class predictions. This is achieved by evaluating the original loss only for the unperturbed samples while the perturbed samples only enter the consistency-enforcing stability loss term (2). This construction reduces the potential negative side-effects of DA, where for certain hyperparameter choices DA can worsen performance on the original unperturbed dataset compared to a baseline model. Note that stability training leads to increased memory requirements as the results for passing the original batch and the perturbed batch have to be held in memory simultaneously.
3 Related Work
This work builds on the original article on ST  that stands out from the literature as one of the few works that explicitly addresses increasing the robustness a neural network against generic input perturbations. We extend their work by considering train distorsions beyond Gaussian noise and a symmetric stability objective to increase the method’s performance for transformative distorsions.
There is a large body of works on related methods in the context of semi-supervised learning. The most relevant works for the present context can be subsumed as consistency/smoothness enforcing methods. The common idea in all cases is to impose a consistency constraint to enforce similar classification behavior for the original and perturbed input. On the labeled subset of the data both the original loss and the consistency loss can be evaluated, whereas on the unlabeled subset only the consistency loss is imposed. The focus in these works lies on incorporating information from the unlabeled subset to increase the model performance but none of them considered the aspect of robustness with respect to input distorsions. The main difference between the different methods lies in the way how the consistency constraint is implemented [1, 14, 21, 25, 15, 28, 2].
In the domain of DA, Rajput et al  recently presented a first theoretical investigation of the robustness properties of DA. On the practical side, there have been proposals for more elaborate implementations of DA that try to circumvent the need for dataset-specific hyperparameter searches by appropriate meta search algorithms [4, 30], which focusing, however, on model performance rather than robustness. In fact, it would represent an interesting line of research to investigate also these techniques from the robustness point of view to see in detail how a far a DA strategy tailored for robustness can get. Mixup  can be seen as an extension to DA in the sense that not only on perturbed input samples but rather on convex combinations of input samples and the corresponding labels are used during training. Recent extensions such as [9, 27, 2] incorporate also a consistency constraint for stabilization.
4 Experimental Setup
4.1 Tiny ImageNet Full-Sized (TIFS) Dataset
We base our experiments on a dataset inspired by the Tiny ImageNet dataset . The latter represents a reduced version of the original ImageNet dataset with only 200 instead of the original 1000 ImageNet classes and 500 samples per class downsized to resolution . As we wanted to keep the advantage of the reduced computational demands of Tiny ImageNet while working on full ImageNet resolutions, we decided to design our custom TIFS dataset along the lines of Tiny ImageNet but keeping the original samples from the ILSVRC2012 ImageNet dataset111A corresponding preprocessing script will be made available after publication.. As no labels are provided for the official ImageNet test set, we used the images from the original validation set as the basis for the TIFS test set and split the original training set in a cross-validation fashion, such that for each class we randomly assigned 450 samples to the training set and 50 to the validation set.
Images contained in the ImageNet dataset and thus in the TIFS dataset come at varying sizes. We preprocess the dataset such that every image in the dataset is first resized, such that the shortest side is 256 pixels in length. Next, we crop out a 224 pixels wide quadratic area, such that the center points of the crop and the image coincide. The image is then converted from RGB integer color values ranging 0 to 255, to floating-point values ranging 0 to 1. Distortions are applied at this point. Finally, we normalize and whiten the (undistorted) images such that the channel-wise mean and variance across the entire dataset are 0 and 1, respectively.
4.2 Model Architecture and Optimization
In our experiments, we use a deep convolutional residual network (ResNet18) 
as prototype for a state-of-the-art convolutional neural network that is presently used predominantly in computer vision applications. To optimize our model, we use mini-batch stochastic gradient descent with Nesterov accelerated momentum12]. We use the default torchvision
implementation of ResNet18 as supplied with Pytorch and use randomly initialized weights rather than pretrained weights that already include DA during pretraining.
We find that introducing the stability objective late during the training phase produces similar results to applying it from the beginning and therefore decided to use the following experimental procedure: We initially train a model for 30 epochs on the original training set with no distortions added to the preprocessing procedure. We save the model at each epoch and select the model with the highest performance on the undistorted validation set. This model serves as baseline model in the following experiments. All subsequent models are initialized with the weights of the baseline model. In an individual run using either ST or DA, the model is trained in the same fashion as the baseline model, except for a limited number of 10 epochs. For a selected distorsion we train a number of models according by varying distorsion hyperparameters. Similarly to the training of the baseline model we use the model performance on the undistorted validation set to perform early stopping, i.e. for a specific hyperparameter choice we select the model with the highest score on the undistorted validation set. We keep a fixed learning rate ofand a batch size of 128 for all experiments.
|Gaussian noise (N)||standard deviation||0.05|
|JPEG compression (N)||quality||30|
|Thumbnail resizing (N)||pixel size||150|
|FGSM Adversarial examples (N)||strength||0.001|
|Rotation (T)||max. rotation||30|
|Random cropping (T)||pixel offset||3|
We consider a range of different distortion types that are summarized in Tab. 1. As already discussed above, we broadly categorize distortions as undirected noise distorsions or as transformative distortions arising from geometric transformations of the input image. For later reference we subjectively identified a parameter value for practical distorsions, that reflects a distorsion strength likely to be encountered in real world applications or in certain settings, like FGSM, a distorsion strength where the distortion is on the verge of being detectable by a human, but has not yet surpassed that threshold.
We consider the following distorsions: Gaussian noise adds pixel-wise independent Gaussian noise to the image such that , where indicates the index of the pixel in and the standard deviation serves as hyperparameter. JPEG compression is an image compression algorithm that aims to minimize file size while maximizing retained semantic meaning. It offers a quality level indicating how much image quality is favored over file size. Thumbnail resizing crops the image down to a quadratic area with side-length
and afterwards resizes the image to its original dimensions. This introduces interpolation artifacts that increase in severity with lower values of. We choose FGSM  as an example method to produce adversarial examples for its simplicity. An example is generated via , where is a strength parameter and is the full loss function. The rotation distortion represents a random rotation up to degrees. Random cropping, unlike the other distortions considered here, changes the common preprocessing procedure. Usually the center point of the final crop and of the image coincide. This distortion offers a parameter such that these points are displaced by up to pixels in all four directions.
For DA we introduce the probabilitythat the selected augmentation will be applied as an additional hyperparameter. This is the default setup in which DA is applied in practical applications and allows to mitigate effects of catastrophic forgetting , i.e. a performance deterioration on the original undistorted dataset that is observed for certain hyperparameter choices. The corresponding hyperparameter for ST is the coefficient , see Eq. 1, that sets the relative scale of the consistency loss term compared to the original loss.
We conclude this section with a remark on the intricacies of comparing ST and DA. If one uses the same batch size in both cases and fixes the number of epochs to the same value in both settings, one the one hand ST is fed twice the amount of raw data. On the other hand, if DA is trained using the same number of examples i.e. by doubling the number of epochs compared to ST, it allows DA to make the double number of label uses compared to ST. Whereas the ST performance typically stabilizes rather quickly during training, increasing the number of training epochs in the DA setting has implications on the robustness properties of the model that will be discussed qualitatively below.
|ST: relative weight||[0.01, 10.0, 3]||logarithmic|
|DA: transformation probability||[0.5, 1.0, 2]||linear|
|Gaussian noise||[0.01, 1.0, 4]||logarithmic|
|JPEG compression||[90, 10, 3]||linear|
|thumbnail resizing||[20, 200, 3]||linear|
|FGSM Adversarial examples||[0.001, 1.0, 7]||logarithmic|
|rotation||[0, 180, 3]||linear|
|random cropping||[0, 15, 3]||linear|
In the following, we refer to the type of distortion to be used for regularization in conjunction with ST or DA during training(testing) as training (test) distortion. In our experiments we train models with a variety of training distortions from both distorsion types and evaluate their performance on various test distortions, see Tab. 2 for the hyperparameter ranges considered during training. The validation set is only used for early stopping based on the model performance on the undistorted validation set as described above. With the scenario of unknown test perturbation in mind, we present the corresponding test set performances for different hyperparameter choices. To guide the eye we typically report both for ST and for DA the performance of the model that performed best/worst at the specified practical distorsion level but we do not perform any form of model selection with respect to the distorsion hyperparameters. Finally, for ST we also report the performance range upper and lower bounds of any training setting at any test setting as a shaded area. This is used as an unbiased measure for the robustness of ST with respect to hyperparameter choices. In our results, we distinguish between those experiments where training and test distortion are of the same type and those where they differ. On the one hand, this allows us to investigate how well a method can increase robustness against a given distortion type and, on the other hand, it shows what side effects on other distortions this might induce.
In Fig. 1 we compare the robustness properties of ST and DA for three different setups within the category of noise distorsions. Fig. 0(a) shows the robustness using Gaussian noise as training distorsion, which corresponds to the setting considered in the original ST paper. We observe in Fig. 0(a) that DA without reintroducing the reference image () outperforms ST only with its peak performance at the chosen practical distortion level but shows a worse performance compared to ST elsewhere. Even more importantly, an unfavorable hyperparameter choice for DA can lead to catastrophic forgetting, which can be seen quite drastically for the worst-performing DA models. This issue can be mitigated to some degree by the reintroduction of the reference image during training, here for , which, however results in worse performance (compared to DA with ) at the practical distortion level. We also use this setting to illustrate the impact of an increased number of training epochs. Here we show results for DA trained for 20 epochs, which corresponds to the same number of batches seen during ST but double the number of labeled examples. Interestingly, increasing the number of training epochs does not result in a model with better performance or robustness. It only marginally improves the best-performing result found at 10 epochs. Even though the worst-performing model is improved by a sizable amount from 10 to 20 epochs, the ST range as illustrated by the area shaded in gray in Fig. 0(a) is still considerably smaller than the range of results obtained from models trained with DA. This leaves ST as the most favorable choice for stabilizing against Gaussian noise consistent with the claims in the original paper. The qualitative findings concerning the impact of an increased number of epochs and the comparison of and represent a general pattern observed during all of our experiments. In the following plots we therefore conventionally show only the worst/best DA performance treating as an additional hyperparameter and restrict ourselves to the setting of 10 training epochs for DA.
Fig. 0(b) illustrates the cross-distorsion performance of both DA and ST. In this particular example, we trained with thumbnail rescaling and evaluated performance on JPEG compression. This setting resembles real-world applications where the model might encounter distorsion types unseen during training. As training and test distortion do not have many characteristics in common, we do not expect to see noticeable performance improvements via either method. Importantly, however, we can observe, that DA can severely harm the model’s robustness against distortion types that were not present during training. Even the best possible model trained via data augmentation performed with a deficit of nearly 10% compared to baseline. ST, on the other hand, did not show any worsening impact. This example should not convey the impression that ST cannot acquire robustness from cross-category distorsions. In fact, for any test distorsion in the noise category, ST performs best using Gaussian noise as training distorsion, while DA always favors coinciding training and test distorsions.
We also investigate the prospects of using ST to increase adversarial robustness. To this end, we dynamically generate adversarial examples via FGSM and feed them as the perturbed image via either ST, DA or adversarial training (AT) with , where the latter corresponds to the optimization objective
where , denotes the reference sample and the adversarial example and is the ground truth label. In Fig. 0(c), we observe that DA is unable to generalize at all, performing similar to baseline. ST offers a significant improvement compared to baseline performance across the entire intensity spectrum. Standard adversarial training offers no to marginal improvements at low intensity levels, but excels in mid-ranges, where it outperforms ST. To investigate this further, we also plot the results if we had selected the best-performing model for an extreme distorsion scenario () instead of the regular practical scenario (). Here we see, that adversarial training sacrifices performance in the low-intensity domain while gaining in the mid- to high-intensity domains. The ST performance remains virtually unchanged comparing the best-performing models at extreme distorsion to that a practical distorsion level. As the gray ST band indicates, ST performance never drops below baseline performance. Interestingly, even ST with Gaussian noise leads to an increased adversarial robustness compared to the baseline performance. We are very well aware of the limitations of robustness against FGSM as indicator for general adversarial robustness  and merely see our findings as an indicator the general robustness properties of ST and a potential direction of future research.
Now we turn to the results on transformative distorsions as presented in Fig. 2. In particular, we compare the performance of the model trained using the symmetric stability objective from Eq. 6 to the performance of those trained with the standard stability objective. In fact, we observe in Fig. 1(a) that the symmetrized stability objective does remedy the shortcoming of the original stability objective when training with rotation. Across all intensity levels , the symmetrical objective can significantly increase the model’s robustness towards rotation compared to regular ST. Even the increased performance of the original ST objective at is consistent with expectations as the original ST loss distinguishes the reference image. We also show performance for DA () with and without () reintroducing the reference image. While the reintroduction of the reference image showed improvements in Fig. 0(a), we observe the opposite effect in Fig. 1(a). It is important to note, that the tested rotation range beyond the practical scenario is unrealistically large for real-world applications. In the range up to rotations in the practical scenario, the symmetrical stability loss offers the best performance of all configurations. However, it is interesting to note that the performance of the worst symmetrical ST model unlike that of the worst original ST model drops below the baseline performance in the undistorted case. Similarly to the noise category experiments, we evaluate the performance across distortion instances of the different type. This is shown exemplarily in Fig. 1(b), where we trained models with varying crops of the input image and evaluated the resulting performance on rotated images. Again, we observe that symmetrical ST offers the best performance of all configurations across the entire intensity spectrum. Also, the gray area shows that no ST model drops below baseline performance, while DA does for unfavorable hyperparameter choices.
At this point we summarize the results of our single-distorsion experiments: Even though we decided to present the results in the form of examples to illustrate the performance characteristics through the full range of distorsion strengths, we want to stress that these examples just examplify the more general picture underlying our investigations. For all considered distorsions in the setting of identical train and test distorsions ST outperformed DA in the practical distorsion range in terms of best-model performance with a considerably smaller hyperparameter dependence. In particular, DA tends to optimize robustness by increasing the peak performance at the particular distorsion characteristics used during training whereas ST typically achieves a stable performance throughout the whole practical distorsion range, see e.g. Fig. 0(a). ST also shows general robustness i.e. generalizes to unseen distorsions within the same noise/transformative distorsion category. We also investigated the use of cross-category training, i.e. training using a distorsion from the noise category while evaluating on the transformative category, but observed no noticeable improvements compared to baseline performance.
The more favorable setting and an interesting direction for future research turned out to be the combination of distorsions from both categories. As ST performs best for test distorsions of the noise categoy using Gaussian noise during training, we investigate the impact of using both Gaussian noise and rotation during training and compare its performance against models trained solely with the same distortion as used for testing (rotation). As shown in Fig. 1(c), regular ST offers some improvements across the entire intensity spectrum and shows no difference when adding Gaussian noise in addition to rotation during training. Data augmentation performs well when trained solely with rotation across the entire spectrum and can improve on its performance on mid to high intensity levels by adding Gaussian noise as an additional regularization. This comes, however, at the cost of a reduced performance in the practically relevant intensity range. The symmetrical stability loss offers the best performance across the entire spectrum of intensities, showcasing again how stability training is capable to utilize Gaussian noise a universal distortion approximator to generally improve robustness. Note that the ST band is considerably larger than the ST band for stabilizing just using rotations, see Fig. 1(a), which is due to the fact that two training distortions are considered simultaneously. Even in this setting, the ST band is still considerably smaller than the performance range for DA applied using the two distorsions simultaneously.
In this work we investigated methods to increase the robustness of deep neural networks against various kinds of input distortions. To this end we thoroughly analyzed the prospects of using stability training  for this purpose, extending it beyond its original working domain of stabilization on Gaussian noise to arbitrary input distortions. We evaluated the proposed method on an ImageNet-scale image classification task and compared the robustness of models stabilized by stability training to those stabilized by data augmentation as predominantly used approach in practical applications to increase both generalization performance and robustness.
In our experiments we demonstrated that stability training performs on a par or even outperforms data augmentation in all investigated distortion settings ranging from noise distortions, including FGSM adversarial examples, to image transformations. Most importantly, stability training is considerably less sensitive to hyperparameter choices than data augmentation, whose performance on the original undistorted dataset may even deteriorate significantly compared to the baseline model for unfavorable hyperparameter values. Its generalizable robustness property makes stability training a particularly good choice for applications where the specific characteristics of the distortion to be encountered is unknown and general robustness is desired. The exploration of using multiple distortion types jointly for stabilization showed first promising results and represents an interesting direction for future research.
This work was supported by the Bundesministerium für Bildung und Forschung (BMBF) through the Berlin Big Data Center under Grant 01IS14013A and the Berlin Center for Machine Learning under Grant 01IS180371.
-  Bachman, P., Alsharif, O., Precup, D.: Learning with pseudo-ensembles. In: Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N.D., Weinberger, K.Q. (eds.) NIPS. pp. 3365–3373. Curran Associates, Inc. (2014)
-  Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.: Mixmatch: A holistic approach to semi-supervised learning (2019)
-  Carlini, N., Athalye, A., Papernot, N., Brendel, W., Rauber, J., Tsipras, D., Goodfellow, I., Madry, A., Kurakin, A.: On evaluating adversarial robustness. arXiv preprint arXiv:1902.06705 (2019)
-  Cubuk, E.D., Zoph, B., Mane, D., Vasudevan, V., Le, Q.V.: Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501 (2018)
-  Frénay, B., Verleysen, M.: Classification in the presence of label noise: a survey. IEEE transactions on neural networks and learning systems 25(5), 845–869 (2014)
-  French, R.M.: Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences 3(4), 128–135 (1994)
Gal, Y.: Uncertainty in deep learning. Ph.D. thesis, PhD thesis, University of Cambridge (2016)
-  Goodfellow, I., Shlens, J., Szegedy, C.: Explaining and Harnessing Adversarial Examples. International Conference on Learning Representations (2015)
-  Hataya, R., Nakayama, H.: Unifying semi-supervised and robust learning by mixup. In: ICLR The 2nd Learning from Limited Labeled Data (LLD) Workshop (2019)
-  He, K., Zhang, X., Ren, S., Sun, J.: Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification. CVPR pp. 1026–1034 (2015)
-  He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition. CVPR pp. 770–778 (2016)
-  Ioffe, S., Szegedy, C.: Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. ICML pp. 448–456 (2015)
-  Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Pereira, F., Burges, C., Bottou, L., Weinberger, K. (eds.) NIPS. pp. 1097–1105. Curran Associates Inc., USA (2012)
-  Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. ICLR (2017)
-  Miyato, T., ichi Maeda, S., Koyama, M., Ishii, S.: Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence (2018)
-  Montavon, G., Samek, W., Müller, K.R.: Methods for interpreting and understanding deep neural networks. Digital Signal Processing 73, 1–15 (2018)
-  Nesterov, Y.: A method for unconstrained convex minimization problem with the rate of convergence O(1/k^2). In: Doklady AN USSR. vol. 269, pp. 543–547 (1983)
-  Oliver, A., Odena, A., Raffel, C., Cubuk, E.D., Goodfellow, I.J.: Realistic evaluation of deep semi-supervised learning algorithms. arXiv preprint arXiv:1804.09170 (2018)
-  Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch. In: NIPS Autodiff Workshop (2017)
-  Rajput, S., Feng, Z., Charles, Z., Loh, P.L., Papailiopoulos, D.: Does data augmentation lead to positive margin? arXiv preprint arXiv:1905.03177 (2019)
-  Sajjadi, M., Javanmardi, M., Tasdizen, T.: Regularization with stochastic transformations and perturbations for deep semi-supervised learning. NIPS pp. 1171–1179 (2016)
-  Silver, D., Huang, A., Maddison, C.J., Guez, A., Sifre, L., van den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe, D., Nham, J., Kalchbrenner, N., Sutskever, I., Lillicrap, T., Leach, M., Kavukcuoglu, K., Graepel, T., Hassabis, D.: Mastering the game of Go with deep neural networks and tree search. Nature 529(7587), 484–489 (2016)
-  Stanford University’s CS231 course: Tiny ImageNet. https://tiny-imagenet.herokuapp.com/, accessed: 2019-05-07
-  Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., Fergus, R.: Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199 (2013)
-  Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NIPS (2017)
-  Taylor, L., Nitschke, G.: Improving Deep Learning using Generic Data Augmentation. arXiv preprint arXiv:1708.06020 (2017)
-  Verma, V., Lamb, A., Kannala, J., Bengio, Y., Lopez-Paz, D.: Interpolation consistency training for semi-supervised learning. arXiv preprint arXiv:1903.03825 (2019)
-  Xie, Q., Dai, Z., Hovy, E., Luong, M.T., Le, Q.V.: Unsupervised data augmentation. arXiv preprint arXiv:1904.12848 (2019)
Yaeger, L.S., Lyon, R.F., Webb, B.J.: Effective Training of a Neural Network Character Classifier for Word Recognition. In: Mozer, M.C., Jordan, M.I., Petsche, T. (eds.) NIPS, pp. 807–816. MIT Press (1997)
-  Zhang, C., Cui, J., Yang, B.: Learning optimal data augmentation policies via bayesian optimization for image classification tasks (2019)
-  Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. ICLR (2018)
-  Zheng, S., Song, Y., Leung, T., Goodfellow, I.: Improving the Robustness of Deep Neural Networks via Stability Training. CVPR pp. 4480–4488 (2016)