ImageNets for Math Handwriting Recognition: Aida Calculus Dataset

Authors: Zac Hancock, Michael Chifala, Callie Federer, Jiamin He, & Quinn N Lathrop

One of the best ways to learn and practice math is by hand on paper. Digital math applications can take advantage of this natural interaction by including a handwriting recognition capability. We introduce a dataset that can be used to create such models to bridge math learners and digital applications. Given the importance of mathematical expressions across all scientific branches, including physics, engineering, and economics, this dataset can become an important resource for advancing the use of machine learning for the benefit of education.

Our dataset (available at www.kaggle.com/aidapearson/ocr-data) consists of 100,000 images of handwritten math expressions within calculus. The images are synthetically generated which affords 100% correct pixel-level tagging and results in realistic images capable of training models whose performance generalize to real images. It has a very permissive license, the full collaboration tools of Kaggle, and standard data formats that increase generalizability and usability. The dataset offers something to all levels, from beginners building simple character recognition models to experts who wish to predict pixel-by-pixel masks with object detection models and decode the complex structure of math expressions.

The most similar dataset is CROHME, which provides digital ink with stroke data. Our dataset differs in that it focuses on images of math and covers a targeted scope of limits expressions. Also, because our dataset is generated, this scope of math could be changed as needed and the size of the dataset is limited only by practicalities.

The ease of use and richness of the dataset will hopefully excite ML researchers within education and draw new ML researchers to the field. Applications beyond handwriting recognition include translating students’ math to pdfs and automated grading for instructors. ML capabilities built using this dataset would benefit many educational institutions by helping to connect the natural mode of math learners and digital educational applications