A notebook for a weak supervision project.
The U.S generates 292.4 million tons of waste annually, with only 32.1% of that waste recycled. Poor waste management contributes to the release of methane, exacerbating greenhouse gases contributing to climate change. This study aims to use weakly supervised machine learning [1] to sort waste at materials recovery facilities, which would reduce the frequency of sorting mistakes. In comparison to previous approaches [2, 3], this method obviates the need for manual labeling of images, which reduces the time necessary to label a large image dataset. Weakly supervised ML also allows for greater adaptability for region-specific regulations and the ability to use a greater number of images, thereby increasing the accuracy. Using images with identified objects, programmatic labels can classify these images into different waste categories. This approach can result in a 10% increase in classification accuracy, which would result in 9.4 million tons less landfill each year.
[1] Snorkel: Rapid Training Data Creation with Weak Supervision http://www.vldb.org/pvldb/vol11/p269-ratner.pdf
[2] Classification of Trash for Recyclability Status http://cs229.stanford.edu/proj2016/report/ThungYang-ClassificationOfTrashForRecyclabilityStatus-report.pdf
[3] Final Report: Smart Trash Net: Waste Localization and Trash Collection http://cs229.stanford.edu/proj2017/final-reports/5226723.pdf