Laure Thompson, PhD candidate in Computer Science at Cornell University, will give the talk, "Understanding and Directing What Models Learn."
Abstract: Machine learning and statistical methods, such as unsupervised semantic models, make massive cultural heritage collections more explorable and analyzable. These models capture many underlying patterns of the raw textual and visual materials, but how can we better understand which specific patterns are learned by these models? While interpretability is helpful for improving models, it is crucial for building user trust. I will open my talk by showing how image analysis tools can be reframed to aid scholars in studying visual materials. I will demonstrate the use of convolutional neural networks as an exploratory tool for questioning “What is Dada?” Through this example, I will highlight how using machine learning tools in new, unexpected ways can expand our understanding of these tools and what we can gain from them.
Our goal is to use models to uncover new, interesting patterns, but given that cultural heritage collections often contain individually well-studied artifacts, models tend to identify patterns we already know. While it might be useful to organize novels by authors, learning this structure is seldom useful when already known and can be problematic if it is mischaracterized as a cross-cutting pattern. So, how might we direct what models learn so that they are useful to a wider range of scholarly inquiry? I will discuss my recent work on directing topics models to identify more cross-cutting topics. I focus on building simple, transparent interventions for existing, well-established algorithms in order to build intuitive and trusted tools useful to scholars. I will show how topic models of speculative fiction novels can be directed away from learning author and series-oriented topics to more topical and cross-cutting ones. I will close with future work that applies model-direction techniques to images, with an example in images of engraved gemstones from the Greco-Roman world.
Laure Thompson is a PhD candidate in Computer Science at Cornell University where she is advised by David Mimno. Her research bridges machine learning and natural language processing with humanistic scholarship. More specifically, Thompson’s work focuses on understanding what computational models learn and how we can intentionally change what they learn. Since her work is centered on humanities applications, she works with a wide range of cultural heritage collections: from texts of science fiction novels and medieval manuscripts to images of avant-garde journals and magical gems. She is a recipient of an NSF Graduate Research Fellowship and a Cornell University Fellowship and her work has received a best paper award at COLING as well as an honorable mention at EMNLP.
Questions? Contact Lori Kelso