CITADL

Project Overview

Modern AI systems are powerful but can be extremely resource-intensive. Training and running large models often consumes vast amounts of electricity, much of it spent on parameters and data that do not meaningfully improve results. AI developers can find ways to cut these costs, but today these fixes are mostly trial and error. CITADL project takes a different route. We will study learning through the lens of algorithmic complexity, a principled way to ask how much information is really needed to represent a model or a dataset. The project will develop new theory at the intersection of AI and algorithmic information theory, and also provide practical methods that can reach the comparable accuracy with less computation, memory, and energy.

Team Members

• Raghavendra Selvan (PI)
• Project is expected to begin in Spring 2027