데이터셋 탐색

The MNIST database (Modified National Institute of Standards and Technology database) is a large collection of handwritten digits. It has a training set of 60,000 examples, and a test set of 10,000 examples. It is a subset of a larger NIST Special Database 3 (digits written by employees of the United States Census Bureau) and Special Database 1 (digits written by high school students) which contain monochrome images of handwritten digits. The digits have been size-normalized and centered in a fixed-size image. The original black and white (bilevel) images from NIST were size normalized to fit in a 20x20 pixel box while preserving their aspect ratio. The resulting images contain grey levels as a result of the anti-aliasing technique used by the normalization algorithm. the images were centered in a 28x28 image by computing the center of mass of the pixels, and translating the image so as to position this point at the center of the 28x28 field. Source: http://yann.lecun.com/exdb/mnist/ Image Source: https://en.wikipedia.org/wiki/MNIST_database#/media/File:MnistExamples.png

Fashion-MNIST is a dataset comprising of 28×28 grayscale images of 70,000 fashion products from 10 categories, with 7,000 images per category. The training set has 60,000 images and the test set has 10,000 images. Fashion-MNIST shares the same image size, data format and the structure of training and testing splits with the original MNIST. Source: Generative Probabilistic Novelty Detection with Adversarial Autoencoders Image Source: https://github.com/zalandoresearch/fashion-mnist

EMNIST (extended MNIST) has 4 times more data than MNIST. It is a set of handwritten digits with a 28 x 28 format. Source: Domain Discrepancy Measure for Complex Models in Unsupervised Domain Adaptation Image Source: https://arxiv.org/pdf/1803.01900.pdf

Colored MNIST is a synthetic binary classification task derived from MNIST.

The Moving MNIST dataset contains 10,000 video sequences, each consisting of 20 frames. In each video sequence, two digits move independently around the frame, which has a spatial resolution of 64×64 pixels. The digits frequently intersect with each other and bounce off the edges of the frame Source: Mutual Suppression Network for Video Prediction using Disentangled Features Image Source: http://www.cs.toronto.edu/~nitish/unsupervised_video/

MNIST-M is created by combining MNIST digits with the patches randomly extracted from color photos of BSDS500 as their background. It contains 59,001 training and 90,001 test images. Source: A Review of Single-Source Deep Unsupervised Visual Domain Adaptation Image Source: https://arxiv.org/pdf/1505.07818v4.pdf

Permuted MNIST is an MNIST variant that consists of 70,000 images of handwritten digits from 0 to 9, where 60,000 images are used for training, and 10,000 images for test. The difference of this dataset from the original MNIST is that each of the ten tasks is the multi-class classification of a different random permutation of the input pixels. Source: Lifelong Learning with Dynamically Expandable Networks Image Source: https://arxiv.org/pdf/1810.12488.pdf

Kuzushiji-MNIST is a drop-in replacement for the MNIST dataset (28x28 grayscale, 70,000 images). Since MNIST restricts us to 10 classes, the authors chose one character to represent each of the 10 rows of Hiragana when creating Kuzushiji-MNIST. Kuzushiji is a Japanese cursive writing style. Source: Deep Learning for Classical Japanese Literature Image Source: https://github.com/rois-codh/kmnist

A collection of 10 pre-processed medical open datasets. MedMNIST is standardized to perform classification tasks on lightweight 28x28 images, which requires no background knowledge. Source: MedMNIST Classification Decathlon: A Lightweight AutoML Benchmark for Medical Image Analysis

Common corruptions dataset for MNIST.

The MultiMNIST dataset is generated from MNIST. The training and tests are generated by overlaying a digit on top of another digit from the same set (training or test) but different class. Each digit is shifted up to 4 pixels in each direction resulting in a 36×36 image. Considering a digit in a 28×28 image is bounded in a 20×20 box, two digits bounding boxes on average have 80% overlap. For each digit in the MNIST dataset 1,000 MultiMNIST examples are generated, so the training set size is 60M and the test set size is 10M. Source: https://arxiv.org/pdf/1710.09829.pdf Image Source: Sabour et al

The Stacked MNIST dataset is derived from the standard MNIST dataset with an increased number of discrete modes. 240,000 RGB images in the size of 32×32 are synthesized by stacking three random digit images from MNIST along the color channel, resulting in 1,000 explicit modes in a uniform distribution corresponding to the number of possible triples of digits. Source: Inclusive GAN: Improving Data and Minority Coverage in Generative Models Image Source: https://arxiv.org/abs/1705.07761

MedMNIST v2 is a large-scale MNIST-like collection of standardized biomedical images, including 12 datasets for 2D and 6 datasets for 3D. All images are pre-processed into 28 x 28 (2D) or 28 x 28 x 28 (3D) with the corresponding classification labels, so that no background knowledge is required for users. Covering primary data modalities in biomedical images, MedMNIST v2 is designed to perform classification on lightweight 2D and 3D images with various data scales (from 100 to 100,000) and diverse tasks (binary/multi-class, ordinal regression and multi-label). The resulting dataset, consisting of 708,069 2D images and 10,214 3D images in total, could support numerous research / educational purposes in biomedical image analysis, computer vision and machine learning. Description and image from: MedMNIST v2: A Large-Scale Lightweight Benchmark for 2D and 3D Biomedical Image Classification Each subset keeps the same license as that of the source dataset. Please also cite the corresponding paper of source data if you use any subset of MedMNIST.

MNIST8M is derived from the MNIST dataset by applying random deformations and translations to the dataset. Source: Scalable and Sustainable Deep Learningvia Randomized Hashing

The exact pre-processing steps used to construct the MNIST dataset have long been lost. This leaves us with no reliable way to associate its characters with the ID of the writer and little hope to recover the full MNIST testing set that had 60K images but was never released. The official MNIST testing set only contains 10K randomly sampled images and is often considered too small to provide meaningful confidence intervals. The QMNIST dataset was generated from the original data found in the NIST Special Database 19 with the goal to match the MNIST preprocessing as closely as possible. QMNIST is licensed under the BSD-style license. Source: https://github.com/facebookresearch/qmnist Image Source: https://github.com/facebookresearch/qmnist

See paper: Caldas, Sebastian, et al. "Leaf: A benchmark for federated settings." arXiv preprint arXiv:1812.01097 (2018).

Brief Description The Neuromorphic-MNIST (N-MNIST) dataset is a spiking version of the original frame-based MNIST dataset. It consists of the same 60 000 training and 10 000 testing samples as the original MNIST dataset, and is captured at the same visual scale as the original MNIST dataset (28x28 pixels). The N-MNIST dataset was captured by mounting the ATIS sensor on a motorized pan-tilt unit and having the sensor move while it views MNIST examples on an LCD monitor as shown in this video. A full description of the dataset and how it was created can be found in the paper below. Please cite this paper if you make use of the dataset. Orchard, G.; Cohen, G.; Jayawant, A.; and Thakor, N. “Converting Static Image Datasets to Spiking Neuromorphic Datasets Using Saccades", Frontiers in Neuroscience, vol.9, no.437, Oct. 2015

The dataset is based on the original MNIST dataset. Compared to the original dataset, the digits are scaled down by a factor of $0.75$ such that there is more space for the random translation.The PolyMNIST consists of 5 different modalities. The background of every modality $\mathbf{x}_m$ consists of random patches of size $28 \times 28$ from a large image. And the digit is placed at a random position of the patch. Using this setup, every modality has modality-specific information given by its background image and shared information given by the digit, which is shared between all modalities. An additional difficulty compared to the original PolyMNIST is the random translation of the digits

A binarized version of MNIST. Source: Binarized MNIST

The Kannada-MNIST dataset is a drop-in substitute for the standard MNIST dataset for the Kannada language. Source: https://github.com/vinayprabhu/Kannada_MNIST Image Source: https://github.com/vinayprabhu/Kannada_MNIST

This is a dataset with spurious correlations which can be used to evaluate machine learning methods for out-of-distribution generalization, causal inference, and related field.

A minimalist, low-memory, and low-compute alternative to classic deep learning benchmarks. The training examples are 20 times smaller than MNIST examples yet they differentiate more clearly between linear, nonlinear, and convolutional models which attain 32, 68, and 94% accuracy respectively (these models obtain 94, 99+, and 99+% on MNIST). Source: Scaling down Deep Learning

N-Digit MNIST is a multi-digit MNIST-like dataset.

CI-MNIST (Correlated and Imbalanced MNIST) is a variant of MNIST dataset with introduced different types of correlations between attributes, dataset features, and an artificial eligibility criterion. For an input image $x$, the label $y \in \\{1, 0\\}$ indicates eligibility or ineligibility, respectively, given that $x$ is even or odd. The dataset defines the background colors as the protected or sensitive attribute $s \in \\{0, 1\\}$, where blue denotes the unprivileged group and red denotes the privileged group. The dataset was designed in order to evaluate bias-mitigation approaches in challenging setups and be capable of controlling different dataset configurations.

The MNIST Large Scale dataset is based on the classic MNIST dataset, but contains large scale variations up to a factor of 16. The motivation behind creating this dataset was to enable testing the ability of different algorithms to learn in the presence of large scale variability and specifically the ability to generalise to new scales not present in the training set over wide scale ranges. The dataset contains training data for each one of the relative size factors 1, 2 and 4 relative to the original MNIST dataset and testing data for relative scaling factors between 1/2 and 8, with a ratio of $\sqrt[4]{2}$ between adjacent scales.

We introduce the Oracle-MNIST dataset, comprising of 2828 grayscale images of 30,222 ancient characters from 10 categories, for benchmarking pattern classification, with particular challenges on image noise and distortion. The training set totally consists of 27,222 images, and the test set contains 300 images per class. Oracle-MNIST shares the same data format with the original MNIST dataset, allowing for direct compatibility with all existing classifiers and systems, but it constitutes a more challenging classification task than MNIST. The images of ancient characters suffer from 1) extremely serious and unique noises caused by three-thousand years of burial and aging and 2) dramatically variant writing styles by ancient Chinese, which all make them realistic for machine learning research. The dataset is freely available at https://github.com/wm-bupt/oracle-mnist.

This is a simple audio-visual dataset artificially assembled from independent visual and audio datasets. The first modality corresponds to 28 × 28 MNIST images, with 75% of their energy removed by PCA. The audio modality is made of audio samples on which we have computed 112 × 112 spectrograms. The audio samples are 25,102 pronounced digits of the Tidigits database augmented by adding randomly chosen noise samples from the ESC-50 dataset. Contaminated audio samples are randomly paired, accordingly with labels, with MNIST digits in order to reach 55,000 pairs for training and 10,000 pairs for testing. For validation we take 5000 samples from the training set.

DeepFake MNIST+ is a deepfake facial animation dataset. The dataset is generated by a SOTA image animation generator. It includes 10,000 facial animation videos in ten different actions, which can spoof the recent liveness detectors.

DirtyMNIST is a concatenation of MNIST + AmbiguousMNIST, with 60k samples each in the training set. AmbiguousMNIST contains additional ambiguous digits with varying ambiguity. The AmbiguousMNIST test set contains 60k ambiguous samples as well. ## Additional Guidance 1. DirtyMNIST is a concatenation of MNIST + AmbiguousMNIST, with 60k samples each in the training set. 2. The current AmbiguousMNIST contains 6k unique samples with 10 labels each. This multi-label dataset gets flattened to 60k samples. The assumption is that ambiguous samples have multiple "valid" labels as they are ambiguous. MNIST samples are intentionally undersampled (in comparison), which benefits AL acquisition functions that can select unambiguous samples. 3. Pick your initial training samples (for warm starting Active Learning) from the MNIST half of DirtyMNIST to avoid starting training with potentially very ambiguous samples, which might add a lot of variance to your experiments. 4. Make sure to pick your validation set from the MNIST half as well, for the same reason as above. 5. Make sure that your batch acquisition size is >= 10 (probably) given that there are 10 multi-labels per samples in Ambiguous-MNIST. 6. By default, Gaussian noise with stddev 0.05 is added to each sample to prevent acquisition functions (in Active Learning) from cheating by disgarding "duplicates". 7. If you want to split Ambiguous-MNIST into subsets (or Dirty-MNIST within the second ambiguous half), make sure to split by multiples of 10 to avoid splits within a flattened multi-label sample.

A set of synthetic MNIST-style datasets for four orthographies used in Afro-Asiatic and Niger-Congo languages: Ge`ez (Ethiopic), Vai, Osmanya, and N'Ko. These datasets serve as "drop-in" replacements for MNIST. Source: Afro-MNIST: Synthetic generation of MNIST-style datasets for low-resource languages

We provide multiple human annotations for each test image in Fashion-MNIST. This can be used as soft labels or probabilistic labels instead of the usual hard (single) labels.

LymphoMNIST is a comprehensive dataset designed for the nuanced classification of lymphocyte images. It encompasses approximately 80,000 high-resolution 64x64 images, meticulously categorized into three primary classes: B cells, T4 cells, and T8 cells.​ Dataset Characteristics: Size: ~80,000 images​ Resolution: 64x64 pixels​ Classes: B cells, T4 cells, T8 cells​ Format: MNIST-like standardized biomedical imagery​ Modality: Microscopy-based high-resolution cell images​ Motivation and Summary: LymphoMNIST aims to bridge the gap in biomedical image analysis by providing a dataset that is vast in scale and rich in detail. It supports a wide array of research endeavors, from fundamental biological studies to advanced computational model development.​ Potential Use Cases: Medical Research: Studying lymphocyte morphology and characteristics​ Machine Learning & AI: Developing and evaluating image classification models​ AutoML & Benchmarking: Serving as a benchmark dataset for automated model training and performance evaluation​ Educational Purposes: Teaching deep learning concepts in biomedical imaging​ Data Collection Process: The dataset comprises high-resolution images of lymphocytes obtained through microscopy. Each image is standardized to a 64x64 pixel resolution to maintain consistency and facilitate analysis.​ Annotations and Labels: Each image is labeled as one of the three lymphocyte classes: B cells, T4 cells, or T8 cells. The labeling process was conducted by experts in the field to ensure accuracy.

Each dataset in the Mechanical MNIST collection contains the results of 70,000 (60,000 training examples + 10,000 test examples) finite element simulation of a heterogeneous material subject to large deformation. Mechanical MNIST is generated by first converting the MNIST bitmap images (http://www.pymvpa.org/datadb/mnist.html) to 2D heterogeneous blocks of material. Consistent with the MNIST bitmap ($28 \times 28$ pixels), the material domain is a $28 \times 28$ unit square. All simulations are conducted with the FEniCS computing platform (https://fenicsproject.org). The code to reproduce these simulations is hosted on GitHub (https://github.com/elejeune11/Mechanical-MNIST/tree/master/generate_dataset). The paper "Mechanical MNIST: A benchmark dataset for mechanical metamodels" can be found at https://doi.org/10.1016/j.eml.2020.100659. All code necessary to reproduce the metamodels demonstrated in the manuscript is available on GitHub (https://github.com/elejeune11/Mechanical-MNIST). For questions, please contact Emma Lejeune (elejeune@bu.edu).

The Mechanical MNIST Crack Path dataset contains Finite Element simulation results from phase-field models of quasi-static brittle fracture in heterogeneous material domains subjected to prescribed loading and boundary conditions. For all samples, the material domain is a square with a side length of $1$. There is an initial crack of fixed length ($0.25$) on the left edge of each domain. The bottom edge of the domain is fixed in $x$ (horizontal) and $y$ (vertical), the right edge of the domain is fixed in $x$ and free in $y$, and the left edge is free in both $x$ and $y$. The top edge is free in $x$, and in $y$ it is displaced such that, at each step, the displacement increases linearly from zero at the top right corner to the maximum displacement on the top left corner. Maximum displacement starts at $0.0$ and increases to $0.02$ by increments of $0.0001$ ($200$ simulation steps in total). The heterogeneous material distribution is obtained by adding rigid circular inclusions to the domain using the Fashion MNIST bitmaps as the reference location for the center of the inclusions. Specifically, each center point location is generated randomly inside a square region defined by the corresponding Fashion MNIST pixel when the pixel has an intensity value higher than $10$. In addition, a minimum center-to-center distance limit of $0.0525$ is applied while generating these center points for each sample. The values of Young’s Modulus $(E)$, Fracture Toughness $(G_f)$, and Failure Strength $(f_t)$ near each inclusion are increased with respect to the background domain by a variable rigidity ratio $r$. The background value for $E$ is $210000$, the background value for $G_f$ is $2.7$, and the background value for $f_t$ is $2445.42$. The rigidity ratio throughout the domain depends on position with respect to all inclusion centers such that the closer a point is to the inclusion center the higher the rigidity ratio will be. We note that the full algorithm for constructing the heterogeneous material property distribution is included in the simulations scripts shared on GitHub. The following information is included in our dataset: (1) A rigidity ratio array to capture heterogeneous material distribution reported over a uniform $64\times64$ grid, (2) the damage field at the final level of applied displacement reported over a uniform $256\times256$ grid, and (3) the force-displacement curves for each simulation. All simulations are conducted with the FEniCS computing platform (https://fenicsproject.org). The code to reproduce these simulations is hosted on GitHub (https://github.com/saeedmhz/phase-field).

MedMNIST-C is an open-source data set collection comprising algorithmically generated corruptions applied to the test sets of the MedMNIST collection following the concept of ImageNet-C. To maintain the integrity of the medical data, we have excluded any weather-dependent corruptions (“Snow”, “Frost”, “Fog”). Hence, each data set in the MedMNIST-C collection comprises 16 different corruptions (12 test corruptions and 4 validation corruptions) spanning 5 severity levels. For further information on the corruptions visit the original GitHub repository of ImageNet-C.

The Cifar10Mnist dataset is created using CIFAR-10 and MNIST data sources. Since the CIFAR-10 training set consists of 50000 images and the MNIST training set contains 60000 digits, the first 50000 digits from MNIST are padded on top of the CIFAR-10 images after making them slightly translucent. A first training dataset is then obtained (50000 images). Furthermore, the remaining 10000 MNIST digits are padded on top of 10000 random CIFAR10 images (with a fixed seed). This gives the possibility of having a second training dataset of 60000 images. For the test set, the 10000 CIFAR-10 images are padded over the 10000 MNIST digits.

Click to add a brief description of the dataset (Markdown and LaTeX enabled). Provide: * a high-level explanation of the dataset characteristics * explain motivations and summary of its content * potential use cases of the dataset

A simple dataset consisting of three geometric shapes (Triangle, Rectangle, Ellipsoid) of similar sizes but different orientations.

IllusionFashionMNIST_test ### Dataset Characteristics IllusionFashionMNIST_test is a generated dataset derived from the FashionMNIST dataset. It incorporates the concept of pareidolia—a phenomenon where patterns, often faces, are perceived in random or abstract stimuli. The dataset contains 11 classes: the original 10 classes from FashionMNIST, and an additional "No Illusion" class. It includes 1,267 samples, all synthetically created rather than real-world images. ### Motivations and Content Summary The dataset was created using ControlNet for image generation, with captions produced by four large language models (LLMs). The aim is to combine the complex visual patterns in fashion items with the phenomenon of pareidolia, encouraging models to reason about visual illusions in the context of non-digit, real-world-like data. This introduces an added layer of abstraction to the already challenging task of image classification. ### Potential Use Cases - Illusory VQA: Questioning models about the illusions present in the images. - Multimodal Model Evaluation: Benchmarking multimodal models' ability to interpret and reason about abstract and illusory patterns. - Perceptual Studies: Investigating how AI models perceive pareidolia in fashion-related visuals. - Synthetic Data Research: Exploring the use of generated datasets to present unconventional challenges to machine learning models.

IllusionMNIST_test ### Dataset Characteristics IllusionMNIST_test is a generated dataset derived from the MNIST dataset. It introduces a novel element of pareidolia—a phenomenon where patterns, often faces, are perceived in random or abstract stimuli. The dataset contains 11 classes: the original 10 digits from MNIST, and an additional "No Illusion" class. It includes 1,219 samples, all synthetically created rather than real-world images. ### Motivations and Content Summary The dataset was created using ControlNet for image generation, with captions produced by four large language models (LLMs). The goal is to introduce visual illusions and pareidolia into an otherwise familiar domain, encouraging models to handle the nuanced challenge of reasoning about visual illusions. This expands upon the MNIST dataset by blending the simplicity of handwritten digits with the complexity of human perception. ### Potential Use Cases - Illusory VQA: Questioning models about the illusions present in the images. - Multimodal Model Evaluation: Benchmarking multimodal models' ability to interpret and reason about abstract and illusory patterns. - Perceptual Studies: Understanding how AI models perceive and classify pareidolia, a phenomenon deeply tied to human visual perception. - Synthetic Data Research: Exploring the use of generated datasets to introduce unconventional challenges to machine learning models.

MNIST Multiview Datasets ======================== MNIST is a publicly available dataset consisting of 70, 000 images of handwritten digits distributed over ten classes. We generated 2 four-view datasets where each view is a vector of R^{14 x 14}: * MNIST₁: It is generated by considering 4 quarters of image as 4 views. * MNIST₂: It is generated by considering 4 overlapping views around the centre of images: this dataset brings redundancy between the views. Related Papers: Goyal, Anil, Emilie Morvant, Pascal Germain, and Massih-Reza Amini. "Multiview Boosting by Controlling the Diversity and the Accuracy of View-specific Voters." Neurocomputing, 358, 2019, pp. 81-92. Link to the ArXiv version: https://arxiv.org/abs/1808.05784 Published Version: https://doi.org/10.1016/j.neucom.2019.04.072 Goyal, Anil, Emilie Morvant, and Massih-Reza Amini. "Multiview Learning of Weighted Majority Vote by Bregman Divergence Minimization." In International Symposium on Intelligent Data Analysis, pp. 124-136. Springer, Cham, 2018. Link to the ArXiv version: https://arxiv.org/abs/1805.10212 Published Version: https://doi.org/10.1007/978-3-030-01768-2_11

MNIST dataset with included uncertainty. For details see the corresponding paper.

MNIST-MIX is a multi-language handwritten digit recognition dataset. It contains digits from 10 different languages. Source: https://github.com/jwwthu/MNIST-MIX

The Mechanical MNIST – Distribution Shift dataset contains the results of finite element simulation of heterogeneous material subject to large deformation due to equibiaxial extension at a fixed boundary displacement of d = 7.0. The result provided in this dataset is the change in strain energy after this equibiaxial extension. The Mechanical MNIST dataset is generated by converting the MNIST bitmap images (28x28 pixels) with range 0 - 255 to 2D heterogeneous blocks of material (28x28 unit square) with varying modulus in range 1- s. The original bitmap images are sourced from the MNIST Digits dataset, (http://www.pymvpa.org/datadb/mnist.html) which corresponds to Mechanical MNIST – MNIST, and the EMNIST Letters dataset (https://www.nist.gov/itl/products-and-services/emnist-dataset) which correspond to Mechanical MNIST – EMNIST Letters. The Mechanical MNIST – Distribution Shift dataset is specifically designed to demonstrate three types of data distribution shift: (1) covariate shift, (2) mechanism shift, and (3) sampling bias, for all of which the training and testing environments are drawn from different distributions. For each type of data distribution shift, we have one dataset generated from the Mechanical MNIST bitmaps and one from the Mechanical MNIST – EMNIST Letters bitmaps. For the covariate shift dataset, the training dataset is collected from two environments (2500 samples from s = 100, and 2500 samples from s = 90), and the test data is collected from two additional environments (2000 samples from s = 75, and 2000 samples from s = 50). For the mechanism shift dataset, the training data is identical to the training data in the covariate shift dataset (i.e., 2500 samples from s = 100, and 2500 samples from s = 90), and the test datasets are from two additional environments (2000 samples from s = 25, and 2000 samples from s = 10). For the sampling bias dataset, datasets are collected such that each datapoint is selected from the broader MNIST and EMNIST inputs bitmap selection by a probability which is controlled by a parameter r. The training data is collected from two environments (9800 from r = 15, and 200 from r = -2), and the test data is collected from three different environments (2000 from r = -5, 2000 from r = -10, and 2000 from r = 1). Thus, in the end we have 6 benchmark datasets with multiple training and testing environments in each. The enclosed document “folder_description.pdf'” shows the organization of each zipped folder provided on this page. The code to reproduce these simulations is available on GitHub (https://github.com/elejeune11/Mechanical-MNIST/blob/master/generate_dataset/Equibiaxial_Extension_FEA_test_FEniCS.py).

Revealing latent structure in data is an active field of research, having introduced exciting technologies such as variational autoencoders and adversarial networks, and is essential to push machine learning towards unsupervised knowledge discovery. However, a major challenge is the lack of suitable benchmarks for an objective and quantitative evaluation of learned representations. To address this issue we introduce Morpho-MNIST, a framework that aims to answer: "to what extent has my model learned to represent specific factors of variation in the data?" We extend the popular MNIST dataset by adding a morphometric analysis enabling quantitative comparison of trained models, identification of the roles of latent variables, and characterisation of sample diversity. We further propose a set of quantifiable perturbations to assess the performance of unsupervised and supervised methods on challenging tasks such as outlier detection and domain adaptation. Data and code are available at https://github.com/dccastro/Morpho-MNIST.

Typography-MNIST is a dataset comprising of 565,292 MNIST-style grayscale images representing 1,812 unique glyphs in varied styles of 1,355 Google-fonts. The glyph-list contains common characters from over 150 of the modern and historical language scripts with symbol sets, and each font-style represents varying subsets of the total unique glyphs. The dataset has been developed as part of the Cognitive Type project which aims to develop eye-tracking tools for real-time mapping of type to cognition and to create computational tools that allow for the easy design of typefaces with cognitive properties such as readability.