The computer science major studied how machine learning models work by using a novel approach to determine which features of a data set are the most useful to the model.

Machine learning is all around us. It’s how self-driving cars, Google ads, and Netflix suggestions work. But exactly how do they work? That’s exactly what computer science major Casey Falk wanted to explore in senior thesis.

“These models often perform quite well, as exemplified by the excellent voice-recognition in Amazon’s Echo, the accuracy of Google’s search recommendations, and the uncanny ability of Facebook to recognize our faces in photos, however, we’re often left in the dark about how these computational models are arriving at a decision—thus turning them into ‘black-boxes'” says Falk. “In my thesis, I explored a novel approach that we called Gradient Feature Auditing, which allows us to determine which features of a data-set are the most useful to a model and thus peer inside these black-boxes.”

Falk worked on his capstone project with Assistant Professor of Computer Science Sorelle Friedler, with whom he already worked on the Dark Reactions Project, which uses failed chemical reactions as the basis for a database that can predict future reactions. (Falk was co-author of the paper—along with Friedler, two Haverford chemistry professors, and four other Haverford students, among others—on the subject that appeared on the cover of the international journal Nature.)

“She not only proposed the initial thesis idea, which is strongly related to her other work on ‘fairness’ of data, but offered insight and guidance throughout my experiments and analysis,” he says of Friedler.

Falk is now living in Seattle, where he is a software engineer for Amazon, but where he hopes to eventually join the machine-learning community.

“I am greatly indebted to Haverford for allowing me the flexibility to pursue research before commencing my thesis,” says Falk. “I worked on machine learning on the Dark Reactions Project for a few years, and without that work my thesis would have taken quite a different avenue.”

What did you learn from working on your thesis?

The key point that it drove home for me was that different types of computational models learn different things, and that we don’t need to treat these models as magical black-boxes. In other words, it is possible to gain insight into what information a model is using to come to a decision. I also got to explore the exciting area of deep learning, which has long been a personal interest of mine and has recently seen enthusiasm in the news for the possibilities it may offer in the area of computational decision making.

What are the implications for your research?

The implications of this work are multifaceted. On one end, users of external machine learning systems–such as those available via websites or bundled into software programs–have a way to better understand what information is important for a machine learning model. On another hand, creators of machine-learning models have another approach to understand what it is that their models are actually learning. Similarly, since our algorithm is agnostic of the type of computational model, many hard-to-dissect models–such as deep neural networks–will benefit greatly from this technique. We also published a similar paper to arXiv so that this work is more accessible to other researchers.


“What They Learned” is a blog series exploring the thesis work of recent graduates.

Photo by Patrick Montero