Beyond Black-Box Scaling: Interpretability, Algorithms, and GPUs by Dr Jiachang Liu

Abstract

As machine learning becomes more widespread, it is increasingly used to support high-stakes decisions in domains such as healthcare and scientific discovery. However, as models become larger and more complex, we often lose control: black-box models can be difficult to troubleshoot, constrain with domain requirements, and trust in deployment. How can we extract actionable insights beyond prediction, and how can we better leverage today’s growing compute to do so?

This talk explores an alternative route: scaling optimization while shrinking model complexity, especially for problems with discrete structure—often yielding 10–100× reductions in training time relative to traditional approaches. First, it shows how scalable discrete optimization can learn interpretable medical scoring systems from real-world datasets that achieve black-box-level accuracy while remaining deployable as simple, auditable decision tools. Second, it describes GPU-compatible methods for obtaining certifiably optimal solutions to highly nonconvex, combinatorial problems, enabling new capabilities such as discovering governing differential equations from data in physical science.

About the Speaker