
Alireza Olama
Machine Learning and Optimization
What I work on
Machine learning engineer with a background in mathematical optimization and high-performance computing. Focused on building scalable algorithms and high-performance systems for AI workloads, numerical optimization, and distributed computing. Interests include LLM training and inference, GPU computing, and performance engineering.
News
- Jul 2026Joined QuTwo in Helsinki as a Member of Technical Staff.
- Jul 2026Presented PsFiT at TKTP2026, the 43rd Annual Symposium of Computer Science, in Turku.
- Dec 2025PruneX preprint online: hierarchical communication-efficient distributed CNN training.
- Autumn 2025Launched the master's-level GPU Programming course at Åbo Akademi University.
- Spring 2025Ran an HPC workshop on the Mahti supercomputer for graduate students and researchers.
Selected work
PruneX
Hierarchical communication-efficient distributed CNN training. Co-designs a structured-sparsity algorithm (H-SADMM) with cluster topology, so the pruning decision and the Allreduce pattern reinforce each other.
PsFiT
Distributed sparse machine learning training under a hard ℓ₀ constraint via Bi-cADMM, with no relaxation. Data is sharded across nodes while features are partitioned across GPUs.
SCOT
C++17 distributed solver that reformulates sparse convex problems exactly as mixed-integer programs (DiPOA, DiHOA). Benchmarked on 140 problems against state-of-the-art MINLP solvers.
SHOT(contributor)
Gave this COIN-OR convex MINLP solver a thread-pool execution model, so primal heuristics run concurrently with the dual search instead of blocking it.
Selected publications
- PsFiT: A GPU-Accelerated Distributed Framework for Scalable Sparse Machine Learning Training
- PruneX: A Hierarchical Communication-Efficient System for Distributed CNN Training with Structured Pruning
- A GPU-Accelerated Bi-linear ADMM Algorithm for Distributed Sparse Machine Learning
- Sparse Convex Optimization Toolkit: A Mixed-Integer Framework
- A Distributed Primal Outer Approximation Algorithm for Sparse Convex Programming with Separable Structures
- A Tracking Augmented Lagrangian Method for ℓ₀ Sparse Consensus Optimization