Vittorio Palladino
Hi, I am Vittorio! I am a dual Master's student at the University of Illinois Chicago and Politecnico di Milano, working on machine learning systems, computer vision, and autonomous systems.
In summer 2026 I joined UC Berkeley as a visiting research student in the HiPeRLab, advised by Mark W. Mueller, working on motion planning and LiDAR-based perception for autonomous aircraft ground navigation.
News
- 2026 DynGhost was accepted at MMSP 2026 — preprint.
- 2026 EnergyLens submitted to SC 2026 — preprint.
- 2026 Started visiting research at UC Berkeley — motion planning and LiDAR-based object detection for autonomous aircraft ground navigation.
- 2026 Paper submitted to ECCV: WTHaar-Net: a Hybrid Quantum-Classical Approach — hybrid quantum-classical CNN using Haar Wavelet Transform, validated on IBM Quantum hardware.
- 2026 Research Assistant at UIC — developing energy-aware inference framework for Multimodal LLMs, profiling power and throughput across heterogeneous GPU hardware.
- 2025 Computer Vision Engineer at EssilorLuxottica Smart Eyewear Lab — real-time defect detection with YOLO, model optimization and on-device deployment.
- 2025 Machine Learning Engineer at AIDA Autonomous Driving Team (Politecnico di Milano) — camera-based 3D object detection for autonomous vehicles.
Research
My research interests include machine learning systems, computer vision, autonomous systems, and efficient neural architectures.
Publications
-
DynGhost: Temporally-Modelled Transformer for Dynamic Ghost Imaging with Quantum Detectors
V. Palladino, A. E. Cetin
arXiv:2605.10185 · 2026 · Accepted at MMSP 2026
A transformer with joint spatial and temporal attention for ghost imaging of dynamic scenes, trained with a quantum-aware framework that models Poisson single-photon detector statistics. arXiv -
EnergyLens: Interpretable Closed-Form Energy Models for Multimodal LLM Inference Serving
V. Palladino, G. Palermo, M. E. Papka, Z. Lan
arXiv:2605.10556 · 2026 · Submitted to SC 2026
Symbolic regression over profiling data yields a twelve-parameter closed-form energy model that picks the optimal serving configuration 88.2% of the time using ~50 profiling runs. arXiv -
WTHaar-Net: a Hybrid Quantum-Classical Approach
V. Palladino, T. Idden, A. E. Cetin
arXiv:2603.02497 · 2026 · Submitted to ECCV
Hybrid quantum-classical CNN using Haar Wavelet Transform, achieving parameter reduction with competitive accuracy on CIFAR-10 and Tiny-ImageNet; validated on IBM Quantum hardware.
Experience
-
UC Berkeley — Incoming Researcher (Summer 2026)
Motion planning and state estimation for commercial aircraft on airport grounds. LiDAR-based object detection and classification. -
University of Illinois Chicago — Research Assistant (Jan 2026 – Present)
Energy-aware inference framework for Multimodal LLMs. GPU/CPU telemetry, vLLM, NVML-based power profiling. -
EssilorLuxottica Smart Eyewear Lab — Computer Vision Engineer (Jul–Sep 2025)
Real-time defect detection with YOLO models, model quantization and pruning, on-device deployment. -
AIDA Autonomous Driving Team — ML Engineer (Jan–Sep 2025)
Camera-based 3D object detection for autonomous vehicles with PyTorch, tested on real driving data and simulation.