Machado

Why care about interpretable models when “black-box” networks usually get the answer right?

This is a common misconception and the heart of my research. In the Interpretable Machine Learning (IML) lab, I design interpretable computer vision models and apply them to high-stakes domains like breast cancer risk prediction. A core part of my research is distilling high-dimensional data into insights stakeholders can understand and act on.

So why does interpretability matter? It tells you three things black-box models can’t: whether your model cheated on confounders, what genuine patterns it actually found (often clinically meaningful), and opportunities for humans to directly influence model behavior.

My goal: build models that don’t just predict accurately, but uncover the mechanisms behind the data — earning the trust to deploy where every prediction shapes a patient outcome.

selected publications

Cosine Similarity is Almost All You Need (for Prototypical-Part Models)

Luke Moffett, Frank Willard, Maximillian Machado, and 8 more authors

In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2026

Bib Code

@inproceedings{moffett2026cosine,
  title = {Cosine Similarity is Almost All You Need (for Prototypical-Part Models)},
  author = {Moffett, Luke and Willard, Frank and Machado, Maximillian and Mokel, Emmanuel and Donnelly, Jon and Guo, Zhicheng and Costarino, Adam and Yang, Julia and Kim, Giyoung and Barnett, Alina Jade and others},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages = {1586--1596},
  year = {2026},
}