PANAMA

Probabilistic Automated Numerical Analysis in Machine Learning and Artificial Intelligence

01/05/2018 to 28/02/2023

Numerical tasks—integration, linear algebra, optimization, solving differential equations—form the computational basis of machine intelligence. Currently, human designers pick methods for these tasks from toolboxes. The generic algorithms assembled in such collections tend to be inefficient on any specific task, and can be unsafe when used incorrectly on problems they were not designed for. Research in numerical methods thus carries the potential for groundbreaking advancements in the performance and quality of AI.

Project PANAMA will develop a framework within which numerical methods can be constructed in an increasingly automated fashion, by a parser analyzing the source code of an AI model; and within which numerical methods can assess their own suitability, and adapt both model and computations to the task, at runtime. The key tenet is that numerical methods, since they perform tractable computations to estimate a latent quantity, can themselves be interpreted explicitly as active inference agents; thus concepts from machine learning can be translated to the numerical domain. Groundwork for this paradigm—probabilistic numerics—has recently been developed into a rigorous mathematical framework by the PI and others.

The central guiding problem will be marginal inference in probabilistic models—an area of machine learning where success is currently particularly hampered by computational challenges. The proposed research, structured into four projects, will thus simultaneously deliver new general theory for the computations of learning machines, and concrete new algorithms for a core area of machine learning. Extensions to other numerical areas across machine learning and AI are part of all threads, so that Project PANAMA will significantly improve the efficiency and safety of artificial intelligence, addressing scientific, technological and societal challenges affecting Europeans today.