The Good Old Trees 🌲 in Our Memories
Guest Lecture in Embedded Computer Architectures I at UTwente
On October 8th, 2025, I was invited by Dr. Nikolas Alachiotis to give a guest lecture in the master course Embedded Computer Architectures I (part of the EMSYS program) at the University of Twente.
My talk, titled “The Good Old Trees in Our Memories,” revisited one of the most enduring models in machine learning — the decision tree — and explored how it continues to evolve on modern computing architectures.
During the lecture, we discussed:
- 🌿 Why trees are still good: explainability, efficiency, and robustness to noise;
- ⚙️ System-level insights: cache, memory, and instruction-level implications of tree inference;
- 💡 Architectural innovations: from cache-aware layout optimizations to integer-only inference (InTreeger), and further to tree execution on emerging non-volatile racetrack memories.
I also highlighted results from our recent works, including:
- Efficient Realization of Decision Trees for Real-Time Inference (TECS 2022)
- Immediate Split Trees (ECML PKDD 2022)
- TreeHouse: An MLIR-based Compilation Flow for Real-Time Tree-based Inference (TECS 2024)
- TrackScorer: Skyrmion Logic-in-Memory Accelerator for Tree-based Ranking Models (DATE 2025)
The key message: Deep learning is not everything.
With architectural awareness, decision-tree ensembles can still deliver high-performance, explainable, and energy-efficient AI on embedded and emerging hardware.
Many thanks to Dr. Nikolas Alachiotis for the kind invitation and to the EMSYS students for the lively discussion and questions during the session.
(Photos taken by Nikolas during the lecture at University of Twente.)
