Depth First Learning: Learning to Understand Machine Learning

Avital Oliver, Surya Bhupatiraju, Cinjon Resnick, Kumar Krishna Agrawal

Abstract:

An experimental science is only as strong as the veracity of its experiments -- when experiments are not reproducible, researchers struggle to place their work on solid footing and progress is hampered. Many fields have had reproducibility crises. Some have recently recognized this as a problem in machine learning as well.

One way to make research more reproducible is to make it easier to develop a deep understanding of the fundamentals underlying machine learning research papers. In this work, we present a new pedagogy, Depth First Learning, that addresses the challenges in understanding these fundamentals. In Depth First Learning, papers are studied down to their core conceptual dependencies, with additional material such as exercises and reproductions in code. We have used this approach in classes and codified our curricula as an open-source website available at \url{http://www.depthfirstlearning.com}.

Keywords: machine learning, pedagogy, reproducability