Good Old Trees, New Lessons
Abstract
I am happy to share that I will give a keynote at the “Systems Meet AI” Summer School.
My talk is titled “Good Old Trees, New Lessons.” Despite the dominance of deep learning, decision tree ensembles remain highly relevant for structured data, explainability, low-latency inference, and resource-constrained deployment. In this keynote, I will revisit these “good old trees” from a systems perspective.
The central question of the talk is simple: what happens after training?
Using tree-based inference as a concrete example, I will discuss how classic systems problems reappear in modern AI deployment: memory locality, code layout, data representation, instruction encoding, arithmetic cost, and predictability. These are not merely implementation details. They shape how models behave once they become executable artifacts on real hardware.
The talk will cover system-level techniques such as cache-aware layout, hardware-conscious code generation, immediate encoding of split values, and integer-only execution on CPUs and RISC-V platforms. Along the way, I will show how seemingly simple model properties can become useful handles for compiler and architecture optimization.
The broader message is that efficient AI systems are not only about new accelerators or larger models. They also require understanding how models become executable artifacts.
Good old trees remind us that AI systems are still systems.
More information about the summer school and the list of speakers can be found here: https://www.systems-ai.nl/speakers
