Title: Performance Guarantees for LLMs

Abstract: Performance guarantees for LLMs are essential to myriad critical applications — enabling for any application for which quantifiable confidence in LLM performance is nonnegotiable. We develop theory and methods for providing performance guarantees for LLMs in the context of a map between vector embeddings. The Data Kernel Perspective Space (DKPS) provides a framework to generate empirically validated theoretical predictions allowing guaranteed performance quantification for these transformative models. I will present probability/statistics/linear algebra for DKPS, and illustrative examples.

Bio: Carey E. Priebe is Professor in the Department of Applied Mathematics & Statistics and Director of the Mathematical Institute for Data Science (MINDS) at Johns Hopkins University. He is an Elected Member of the International Statistical Institute, Senior Member of the IEEE, Fellow of the Institute of Mathematical Statistics, and Fellow of the American Statistical Association. His current work on foundations of AI is funded by DARPA, ONR, AFOSR, Microsoft, Amazon, etc.

Zoom: https://wse.zoom.us/j/91340867859?pwd=3sJNjYuDVX49ANoLFFFX7VOw6ZTFjQ.1
Meeting ID: 913 4086 7859
Passcode: 322920