phenomenon · claim-type · cross-domain-transfer · mechanism · claim-relation · replication-status · limitation — grounded in the full text| phenomenon | overfitting - a model learning a function of very high variance that models its training data perfectly— “Overfitting refers to the phenomenon when a network learns a function with very high variance such as to perfectly model the training data.” | |
| claim-type | method— “Data Augmentation encompasses a suite of techniques that enhance the size and quality of training datasets such that better Deep Learning models can be built using them.” | |
| method — evidence | — “Limited datasets is an especially prevalent challenge in medical image analysis.” | |
| object of study — evidence | — “Data Augmentation encompasses a suite of techniques that enhance the size and quality of training datasets such that better Deep Learning models can be built using them.” | |
| research question — evidence | — “In contrast to the techniques mentioned above, Data Augmentation approaches overfitting from the root of the problem, the training dataset.” | |
| contribution — evidence | — “Oversampling and Data Warping augmentations do not form a mutually exclusive dichotomy. For example, GAN samples can be stacked with random cropping to further inflate the dataset.” | |
| transfer | from tabular and vector data mining: SMOTE and Borderline-SMOTE were built to interpolate new points for imbalanced TABULAR data and are carried over to images— “SMOTE and the extension of Borderline-SMOTE [30] create new instances by interpolating new points from existing instances via k-Nearest Neighbors.” | transfers[0] |
| transfer | from human learning and cognition: transfer learning is explained by the anecdote that a pianist learns guitar faster, and augmentation by analogy with imagination and dreaming— “For example, a human intelligence anecdote of transfer learning is illustrated in learning music.” | transfers[1] |
| mechanism | AUGMENTATION IS AN ASSUMPTION, AND THE PAPER STATES IT: that more information can be extracted from the dataset already held. Everything else follows from whether that is true of a given dataset— “This is done under the assumption that more information can be extracted from the original dataset through augmentations.” | mechanisms[0] |
| mechanism | WHAT AUGMENTATION IS ACTUALLY TEACHING: the invariances a recogniser must survive - viewpoint, lighting, occlusion, background, scale - are baked into the DATA rather than into the architecture— “The task of Data Augmentation is to bake these translational invariances into the dataset such that the resulting models will perform well despite these challenges.” | mechanisms[1] |
| mechanism | SIMULATING BIG DATA RATHER THAN OBTAINING IT: overfitting is not much of a problem when data is abundant, so augmentation gives a limited dataset the characteristics of an abundant one— “Overfitting is generally not as much of an issue with access to big data. Data Augmentation prevents overfitting by modifying limited datasets to possess the characteristics of big data.” | mechanisms[2] |
| mechanism | an unsafe transformation could be made useful rather than harmful by softening the label, so the model learns uncertainty instead of a falsehood - but building those labels is too costly— “A non-label preserving transformation could potentially strengthen the model’s ability to output a response indicating that it is not confident about its prediction.” | mechanisms[3] |
| mechanism | oversampling works by re-balancing the class distribution so the model is not biased toward answering with the majority class— “Oversampling is a technique used to re-sample imbalanced class distributions such that the model is not overly biased towards labeling instances as the majority class type.” | mechanisms[4] |
| mechanism | medical imaging is short of data FOR STRUCTURAL REASONS, and the paper names rarity first: `It is especially difficult to build big medical image datasets due to the rarity of diseases`. The remaining three reasons it lists - patient privacy, the need for expert labellers, and the cost of the scans - are separated from this clause by a journal running header in the staged text and are not inside this span.— “It is especially difficult to build big medical image datasets due to the rarity of diseases” | mechanisms[5] |
| mechanism | rotation safety is set by one parameter: small rotations help digit recognition and large ones destroy the label, so the same transform is safe and unsafe at different magnitudes— “The safety of rotation augmentations is heavily determined by the rotation degree parameter.” | mechanisms[6] |
| mechanism | cropping and translation differ in what they preserve - cropping shrinks the input, translation keeps the spatial dimensions and pads the vacated space— “The contrast between random cropping and translations is that cropping will reduce the size of the input such as (256,256) → (224, 224), whereas translations preserve the spatial dimensions of the image.” | mechanisms[7] |
| mechanism | colour transformations trade colour bias for spatial characteristics, which is a gain only where colour is not the distinguishing feature— “In effect, color space transformations will eliminate color biases present in the dataset in favor of spatial characteristics.” | mechanisms[8] |
| mechanism | a kernel filter can be rebuilt as a convolution layer with the right padding and activation, so the augmentation duplicates what the network already does— “For example, something like PatchShuffle Regularization could be implemented with a convolution layer.” | mechanisms[9] |
| sense | Data Augmentation - A DATA-SPACE SOLUTION, defined by WHERE it intervenes rather than by what it does: every other remedy for overfitting changes the model, and this one changes the dataset— “In contrast to the techniques mentioned above, Data Augmentation approaches overfitting from the root of the problem, the training dataset.” | senses[0] |
| sense | data warping - the label is PRESERVED and an existing image is transformed— “Data warping augmentations transform existing images such that their label is preserved. This encompasses augmentations such as geometric and color transformations, random erasing, adversarial training, and neural style transfer. Oversampling augmentations create synthetic instances and add them to the training set. This includes mixing… | senses[1] |
| sense | oversampling - a NEW instance is synthesised and added. The paper draws the distinction and then says in the next breath that it does not partition: GAN samples can be stacked with random cropping— “Oversampling and Data Warping augmentations do not form a mutually exclusive dichotomy. For example, GAN samples can be stacked with random cropping to further inflate the dataset.” | senses[2] |
| audit | data warping —distinct-from→ oversampling — Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span says the two augmentations do NOT form a mutually exclusive dichotomy; distinct-from is hedged.— “Data warping augmentations transform existing images such that their label is preserved. This encompasses augmentations such as geometric and color transformations, random erasing, adversarial training, and neural style transfer. Oversampling augmentations create synthetic instances and add them to the training set. This includes mixing… | concept_edges[0] |
| holds when | data warping —distinct-from→ oversampling — the two are NOT mutually exclusive - GAN samples can be stacked with random cropping, so a single augmentation can be both— “Data warping augmentations transform existing images such that their label is preserved. This encompasses augmentations such as geometric and color transformations, random erasing, adversarial training, and neural style transfer. Oversampling augmentations create synthetic instances and add them to the training set. This includes mixing…— “Oversampling and Data Warping augmentations do not form a mutually exclusive dichotomy. For example, GAN samples can be stacked with random cropping to further inflate the dataset.” | concept_edges[0] |
| weight | data warping —distinct-from→ oversampling — 0.8— “Data warping augmentations transform existing images such that their label is preserved. This encompasses augmentations such as geometric and color transformations, random erasing, adversarial training, and neural style transfer. Oversampling augmentations create synthetic instances and add them to the training set. This includes mixing… | concept_edges[0] |
| audit | Transfer Learning —distinct-from→ Pretraining — Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but endpoint by anaphora: 'This differs from Transfer Learning' - 'This' = Pretraining (previous sentences).— “This differs from Transfer Learning because in Transfer Learning, the network architecture such as VGG-16 [2] or ResNet [3] must be transferred as well as the weights.” | concept_edges[1] |
| edge | Dropout —is-a→ regularization technique— “Dropout [7] is a regularization technique that zeros out the activation values of randomly chosen neurons during training. This constraint forces the network to learn more robust features rather than relying on the predictive capability of a small subset of neurons in the network.” | concept_edges[2] |
| edge | Batch normalization —is-a→ regularization technique— “Batch normalization [9] is another regularization technique that normalizes the set of activations in a layer. Normalization works by subtracting the batch mean from each activation and dividing by the batch standard deviation.” | concept_edges[3] |
| effect direction (stated by the tagger, no quotation) | Data Augmentation —causes→ overfitting — reduces | concept_edges[4] |
| relation | Data Augmentation —causes→ overfitting — augmentation gives a limited dataset the characteristics of big data, and overfitting is not much of an issue when data is abundant— “Overfitting is generally not as much of an issue with access to big data. Data Augmentation prevents overfitting by modifying limited datasets to possess the characteristics of big data.” | concept_edges[4] |
| limitation | AUGMENTATION CANNOT CREATE A CLASS THAT IS ABSENT: no method surveyed will produce a golden retriever from a dataset of bulldogs. It can lessen biases of lighting, occlusion, scale and background, and it cannot supply missing content— “no augmentation method discussed, from SamplePairing to AutoAugment to GANs, will create a golden retriever.” | limitations[0] |
| limitation | The stated bias limit in the paper's own words, which bounds every technique it surveys— “Data Augmentation cannot overcome all biases present in a small dataset.” | limitations[1] |
| limitation | The survey covers IMAGE data only, with transfer to other data domains asserted rather than shown— “This survey focuses on applications for image data, although many of these techniques and concepts can be expanded to other data domains.” | limitations[2] |
| limitation | The taxonomy the survey is organised by does not partition its own subject - warping and oversampling are stated not to be mutually exclusive— “Oversampling and Data Warping augmentations do not form a mutually exclusive dichotomy. For example, GAN samples can be stacked with random cropping to further inflate the dataset.” | limitations[3] |
| limitation | safety is domain-dependent, which is precisely what makes a generalisable augmentation policy hard— “This is somewhat domain dependent, providing a challenge for developing generalizable augmentation policies, (see Auto- Augment [38] for further exploration into finding generalizable augmentations).” | limitations[4] |
| limitation | there is NO image processing function that cannot flip a label at some distortion magnitude - the strongest form of the claim, stated by the authors— “There is no image processing function that cannot result in a label changing transformation at some distortion magnitude.” | limitations[5] |
| limitation | constructing refined labels for every unsafe augmentation is computationally expensive, which is why the cheap option is to avoid unsafe transforms— “However, constructing refined labels for every non-safe Data Augmentation is a computationally expensive process.” | limitations[6] |
| limitation | many experiments deliberately restrict themselves to a subset of a large dataset in order to SIMULATE a limited-data problem rather than face a real one— “Many experiments constrain themselves to a subset of the dataset to simulate limited data problems.” | limitations[7] |
| limitation | kernel filters are very close to the internal mechanism of a CNN, which is the argument for making them a layer rather than a dataset operation— “A disadvantage of this technique is that it is very similar to the internal mechanisms of CNNs.” | limitations[8] |
| limitation | in medical imaging the gap between training and test data is more complex than position and translation, so geometric transformations have limited scope there— “Finally, in many of the application domains covered such as medical image analysis, the biases distancing the training data from the testing data are more complex than positional and translational variances.” | limitations[9] |
| limitation | colour transformations can discard the information the task depends on - darkening may hide the object entirely— “Additionally, color transformations may discard important color information and thus are not always a label-preserving transformation.” | limitations[10] |
| limitation | transformations that may alter the label, such as translation or random cropping, must be checked by hand— “Some geometric transformations such as translation or random cropping must be manually observed to make sure they have not altered the label of the image.” | limitations[11] |
finding · quantity · variable · definition · key-term — grounded in the full text| finding | AlexNet's experiments used augmentation to enlarge the dataset by a factor of 2048.— “Data Augmentation is used in their experiments to increase the dataset size by a magnitude of 2048.” | content_tags.findings[0] |
| finding | GAN-based augmentation beat classic augmentation on liver lesion classification on both sensitivity and specificity— “This improved classification performance from 78.6% sensitivity and 88.4% specificity using classic augmentations to 85.7% sensitivity and 92.4% specificity using GAN-based Data Augmentation.” | content_tags.findings[1] |
| finding | Horizontal flipping is much more common than vertical flipping.— “Horizontal axis flipping is much more common than flipping the vertical axis.” | content_tags.findings[2] |
| finding | converting to grayscale costs about three points of classification accuracy against RGB on ImageNet and PASCAL VOC— “Chatifled et al. [59] found a ~ 3% classification accuracy drop between grayscale and RGB images with their experiments on ImageNet [12] and the PASCAL [60] VOC dataset.” | content_tags.findings[3] |
| finding | in a head-to-head comparison of geometric against photometric augmentation, cropping produced the most accurate classifier— “Their results find that the cropping geometric transformation results in the most accurate classifier The italic value denote high performance according to the comparative metrics Geometric versus photometric transformations Taylor and Nitschke [63] provide a comparative study on the effectiveness of geometric and photometric (color space)… | content_tags.findings[4] |
| finding | randomly swapping pixels inside a small window lowered CIFAR-10 error against the same architecture without it— “Experimenting across different filter sizes and probabilities of shuffling the pixels at each step, they demonstrate the effectiveness of this by achieving a 5.66% error rate on CIFAR-10 compared to an error rate of 6.33% achieved without the use of PatchShuffle Regularization.” | content_tags.findings[5] |
| finding | Blurring training images could make models more resistant to motion blur at test time.— “Intuitively, blurring images for Data Augmentation could lead to higher resistance to motion blur during testing.” | content_tags.findings[6] |
| quantity | the factor by which AlexNet's augmentation multiplied the dataset, and the error reduction claimed for it— “Data Augmentation is used in their experiments to increase the dataset size by a magnitude of 2048.” | content_tags.quantities[0] |
| quantity | sensitivity and specificity before and after GAN-based augmentation on liver lesion classification— “This improved classification performance from 78.6% sensitivity and 88.4% specificity using classic augmentations to 85.7% sensitivity and 92.4% specificity using GAN-based Data Augmentation.” | content_tags.quantities[1] |
| quantity | the CIFAR-10 error rate with and without PatchShuffle regularization, and the hyperparameters that achieved it— “Experimenting across different filter sizes and probabilities of shuffling the pixels at each step, they demonstrate the effectiveness of this by achieving a 5.66% error rate on CIFAR-10 compared to an error rate of 6.33% achieved without the use of PatchShuffle Regularization.” | content_tags.quantities[2] |
| quantity | the accuracy drop from grayscale against RGB— “Chatifled et al. [59] found a ~ 3% classification accuracy drop between grayscale and RGB images with their experiments on ImageNet [12] and the PASCAL [60] VOC dataset.” | content_tags.quantities[3] |
| variable | rotation degree— “Rotation augmentations are done by rotating the image right or left on an axis between 1° and 359°.” | content_tags.variables[0] |
| definition | safety— “The safety of a Data Augmentation method refers to its likelihood of preserving the label post-transformation.” | content_tags.definitions[0] |
| concept | Data Augmentation— “Data Augmentation encompasses a suite of techniques that enhance the size and quality of training datasets such that better Deep Learning models can be built using them.” | concepts[0] |
| concept | Generalizability— “Generalizability refers to the performance difference of a model when evaluated on previously seen data (training data) versus data it has never seen before (testing data).” | concepts[1] |
| concept | overfitting— “In contrast to the techniques mentioned above, Data Augmentation approaches overfitting from the root of the problem, the training dataset.” | concepts[2] |
| concept | data warping— “Data warping augmentations transform existing images such that their label is preserved. This encompasses augmentations such as geometric and color transformations, random erasing, adversarial training, and neural style transfer. Oversampling augmentations create synthetic instances and add them to the training set. This includes mixing… | concepts[3] |
| concept | oversampling— “Data warping augmentations transform existing images such that their label is preserved. This encompasses… | concepts[4] |
| concept | Dropout | concepts[5] |
| concept | Batch normalization | concepts[6] |
| concept | Transfer Learning | concepts[7] |
| concept | Pretraining | concepts[8] |
| concept | translational invariances Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[9] |
| concept | training dataset Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[10] |
| concept | regularization technique Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[11] |
| concept | big data Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[12] |
| concept | GAN samples Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[13] |
| concept | safety — not defined in this paper's own words; NOT harm-avoidance: the likelihood that a transformation PRESERVES THE LABEL. A flip is safe for cat-versus-dog and unsafe for 6-versus-9 Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[14] |
| concept | class imbalance — not defined in this paper's own words; a skewed ratio of majority to minority samples - the second problem augmentation is used against, distinct from limited data Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[15] |
| concept | Random Oversampling — not defined in this paper's own words; the naive baseline - duplicates images at random. The rest of the sentence, naming the MINORITY class and the target RATIO, is on the far side of a journal running header in the staged text and is not inside this span. Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[16] |
| concept | SMOTE — not defined in this paper's own words; interpolating new points between existing ones by k-nearest neighbours; built for TABULAR data and imported here Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[17] |
| concept | refined labels — not defined in this paper's own words; soft labels like [0.5 0.5] that would let an unsafe transformation teach calibrated uncertainty instead of a wrong answer - rejected here as too expensive Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[18] |
| concept | Neural Style Transfer — not defined in this paper's own words; one of the augmentation families surveyed, grouped with GANs as methods that can 'imagine' alterations rather than merely transform pixels Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[19] |
| concept | positional bias — not defined in this paper's own words; the defect geometric transformations exist to remove - e.g. every face centred in frame, which would force the test set to be centred too Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[20] |
| concept | kernel filters — not defined in this paper's own words; sliding an n x n matrix to blur or sharpen; argued here to belong INSIDE the network rather than in the dataset Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[21] |
| concept | PatchShuffle — not defined in this paper's own words; a kernel filter that randomly swaps pixels within a window - augmentation by local scrambling rather than by transformation Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[22] |
| concept | color space transformations — not defined in this paper's own words; also called photometric transformations; aimed at lighting bias, and named here as the counterpart to the geometric family Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[23] |
| concept | noise injection — not defined in this paper's own words; adding a matrix of Gaussian values, offered as a route to more robust features rather than to more data Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[24] |
When we have changed our mind. Some entries in the table begin with a word in capitals and the reason for it. EVIDENCE CORRECTED — the finding still stands, but we had quoted the wrong sentence from the paper, and the quotation has been replaced with the right one. RETRACTED — the entry itself was wrong; it stays on the page, marked, with the reason, rather than being deleted. WITHDRAWN AND REPLACED — the entry was wrong, and a corrected version of it appears in the entries below. Each mark carries its own reason, in brackets, because the reasons differ from one entry to the next. An entry we have WITHDRAWN is removed from this page and kept in the record, with its reason. A line reading Audit and a date is a later re-reading of the entry against its whole sentence: DOUBTFUL means it was kept, and the reason says what is loose about it. Quotation marks inside the tagger's sentence. Where a tag is the tagger's own sentence, a phrase in quotation marks inside it is borrowed from the paper — the tagger's wording around it is not. The paper's own sentence is the evidence quote for that row; on this page it is withheld with the article, and the article is at its publisher. Nothing inside quotation marks anywhere on this page is ours: if it is quoted, it is the paper's. Whose words a concept carries. A concept entry records a term the paper itself uses, so the term, the sense given for it, and its evidence are the paper's own words, quoted where the paper fixes them. This is the opposite of a sense entry in the tag table, which is the tagger's sentence about how the paper uses a term. Where a concept entry could not be grounded in the paper's words, the entry says so on its own row.
Tag layer © the Meta-Tagging Project, CC BY 4.0. The paper text is © its authors/publisher
and is linked, not reproduced. Excerpts shown are short quotations for identification and verification.
· back to index
· concepts · authors
· עברית
© 2026 Shir Sivroni · ORCID 0009-0000-2597-2824