We formulate the world in optimization terms (i.e., objectives, constraints, and tradeoffs), then hand it to learning to make decisions. Too often, learning chases benchmarks while overlooking that underlying formulation. My work aligns learning with optimization algorithms, so that neural netwroks inherit the defining properties of iterative algorithms and generalize more reliably.

This perspective goes beyond traditional neural architectures and extends to transformers and diffusion models. I am particularly interested in diffusion models as optimization tools. However, mainstream diffusion models are not yet equipped for the challenges of optimization domains, especially when constraints, tradeoffs, and feasibility matter. I therefore develop architectures and dynamics that enable diffusion models to operate as principled tools for these settings.

My name is Samar (pronounced as Summer), a PhD candidate at the University of Pennsylvania. I work under the supervision of Prof. Alejandro Ribeiro.

I am on the academic job market.

__

News
Apr. 2026: A constrained optimization perspective of unrolled transformers has been accepted to ICML as a Spotlight!
Apr. 2026: A new preprint Graph signal diffusion models for wireless resource allocation is out and will appear as an invited paper to a special session at SPAWC 2026.
Jan. 2026: A new preprint is out: Unrolled neural networks for constrained optimization.
Jan. 2026: Our paper, Stochastic unrolled neural networks, has been accepted to Conference on Parsimony and Learning (CPAL) in March 2026.
Jan. 2026: Two papers, Unrolled graph neural networks for constrained optimization and Graph signal generative diffusion models, got accepted to ICASSP 2026.
Dec. 2025: Presented our papers Unrolled neural networks for constrained optimization and A constrained optimization perspective of unrolled transformers at NeurIPS workoshop on Constrained Optimization for Machine Learning.
Dec. 2025: Presented our paper GNN-parametrized diffusion policies for wireless resource allocation at NeurIPS workoshop on New Perspectives in Advancing Graph Machine Learning.
Sep. 2025: Three papers got accepted to NeurIPS workshops. Details to follow.
Sep. 2025: Our paper Generative diffusion models for resource allocation in wireless networks got accepted to IEEE Int. Workshop on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP). Oral.
Apr. 2025: Gave a talk about our work Robust stochastically-descending unrolled networks at ICASSP 2025.