Paper surveys are still a preferred in-class data collection method due to the captive audience, ease of creation (e.g. MS word), and administration. Data entry is time consuming and expensive. This research explored using convolutional neural networks (CNN, a deep learning process) to automate the data entry of thousands of surveys administered as part of a college program assessment. CNN are used in image detection (e.g. is there a cute cat in the picture?), where digital images are stored as arrays with entries corresponding to pixel color. CNN’s are constructed by creating many layers that each apply specific matrix operations to extract the most salient characteristics in the image. Our work differs from image detection; our goal was to identify where a mark exists on a page (e.g. circles, x’s), and classify it as the corresponding response. Like other machine learning models, CNN depend on quality and quantity of training data, and can be computationally intensive but feasible on a moderate powered server. Preliminary results achieved 92% validation accuracy, demonstrating it’s feasible to use paper surveys without the burden of manual data entry or costly software licenses.
Kris presented his work in the Student Poster Session at the 2020 Joint Math Meetings, January 15-18, 2020, Denver, CO.