Safa Messaoud

Safa Messaoud

AI Scientist
MenaML Co-founder & Co-director

Posts by Collection

portfolio

projects

publications

Reaction–Diffusion Modelling for Microphysiometry on Cellular Specimens

Published in B.S. Thesis, Technical University of Munich, 2011

B.S. thesis: extended a 3D finite-element simulation of diffusion and metabolic reaction in cellular specimens with a sensor-effect model, validated via an electro-chemical experiment.

Recommended citation: S. Messaoud. (2011). "Reaction–Diffusion Modelling for Microphysiometry on Cellular Specimens." B.S. Thesis, Technical University of Munich.
Download Paper

Reaction–Diffusion Modelling for Microphysiometry on Cellular Specimens

Published in Medical & Biological Engineering & Computing, 2013

Extends a 3D finite-element simulation of diffusion and metabolic reaction in cellular specimens with a model of the sensor effect, validated via an electro-chemical experiment.

Recommended citation: D. Grundl, X. Zhang, S. Messaoud, C. Pfister, F. Demmel, M.S. Mommer, B. Wolf, M. Brischwein. (2013). "Reaction–Diffusion Modelling for Microphysiometry on Cellular Specimens." Medical & Biological Engineering & Computing.

Optimal Architecture Synthesis for Aircraft Electrical Power Systems

Published in M.S. Thesis, TU Munich / UC Berkeley, 2013

M.S. thesis: optimization-oriented methodologies (Mixed Integer-Linear Programming modulo reliability; ILP with approximate reliability algebra) for synthesizing cost-effective and reliable aircraft electrical power-system topologies.

Recommended citation: S. Messaoud. (2013). "Optimal Architecture Synthesis for Aircraft Electrical Power Systems." M.S. Thesis, TU Munich / UC Berkeley.
Download Paper

Translating Discrete Time SIMULINK to SIGNAL

Published in M.S. Thesis, Virginia Tech, 2014

M.S. thesis: a semantic translator from discrete-time SIMULINK models to SIGNAL programs, enabling correct-by-design and multi-threaded code generation.

Recommended citation: S. Messaoud. (2014). "Translating Discrete Time SIMULINK to SIGNAL." M.S. Thesis, Virginia Tech.
Download Paper

Translating Discrete Time SIMULINK to SIGNAL

Published in e-TI (electronic journal), 2015

A semantic translator that transforms discrete-time SIMULINK models into SIGNAL programs, enabling correct-by-design and multi-thread code generation.

Recommended citation: S. Messaoud, N. Saeedloei, S. Shukla. (2015). "Translating Discrete Time SIMULINK to SIGNAL." e-TI.

Unsupervised Analysis of Transcriptomic Data for Demystifying the Cause of Alzheimer Disease

Published in Alzheimer's Association International Conference (AAIC), 2017

An unsupervised technique to identify genes that discriminate temporal cortex expression data of Alzheimer-affected patients from control subjects.

Recommended citation: Y. Varatharajah, M. Younkin, X. Wang, A. Athreya, S. Messaoud, R. Iyer, N. Ertekin-Taner. (2017). "Unsupervised Analysis of Transcriptomic Data for Demystifying the Cause of Alzheimer Disease." AAIC.

Structural Consistency and Controllability for Diverse Colorization

Published in European Conference on Computer Vision (ECCV), 2018

Extends Gaussian conditional random fields to model multi-modal distributions with high-order dependencies, enabling exact inference and runtime constraints for diverse image colorization.

Recommended citation: S. Messaoud, D.A. Forsyth, A.G. Schwing. (2018). "Structural Consistency and Controllability for Diverse Colorization." ECCV.

Accelerating Genomic Data Parsing on Field Programmable Gate Arrays

Published in U.S. Patent, 2019

An FPGA-based accelerator for genomic file parsing (SAM to BAM conversion) achieving a 10× speedup over a single-threaded software implementation.

Recommended citation: S. Messaoud, T. Ogasawara. "Accelerating Genomic Data Parsing on Field Programmable Gate Arrays." U.S. Patent.

Can We Learn Heuristics for Graphical Model Inference Using Reinforcement Learning?

Published in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020

A reinforcement-learning engine for solving inference in energy-based models with traditionally intractable higher-order potentials, applied to semantic segmentation. (Oral presentation.)

Recommended citation: S. Messaoud, M. Kumar, A.G. Schwing. (2020). "Can We Learn Heuristics for Graphical Model Inference Using Reinforcement Learning?" CVPR (Oral).

Medical Record Problem List Generation

Published in U.S. Patent, 2021

A novel algorithm for disease-category-based problem list generation from electronic medical records using variants of autoencoders to learn customized features per disease category.

Recommended citation: M.V. Devarakonda, S. Messaoud, C.-H. Tsou. "Medical Record Problem List Generation." U.S. Patent.

Toward More Scalable Structured Models

Published in Ph.D. Dissertation, University of Illinois at Urbana-Champaign, 2021

Doctoral dissertation: combining structured prediction, reinforcement learning, and energy-based models for vision and decision-making problems.

Recommended citation: S. Messaoud. (2021). "Toward More Scalable Structured Models." Ph.D. Dissertation, University of Illinois at Urbana-Champaign.

DeepQAMVS: Query-Aware Hierarchical Pointer Networks for Multi-Video Summarization

Published in ACM SIGIR Conference on Research and Development in Information Retrieval, 2021

A reinforcement-learning method that trains a pointer network with hierarchical attention, achieving state-of-the-art results on query-aware multi-video summarization.

Recommended citation: S. Messaoud, I. Lourentzou, A. Boughoula, M. Zehni, C. Zhai, Z. Zhao, A.G. Schwing. (2021). "DeepQAMVS: Query-Aware Hierarchical Pointer Networks for Multi-Video Summarization." ACM SIGIR.

Impact of Adversarial Training on Robustness and Generalizability of Language Models

Published in Findings of the Association for Computational Linguistics (ACL Findings), 2023

A study of how adversarial training affects robustness and generalization in pretrained language models.

Recommended citation: E. Altinisik, H. Sajjad, H.T. Sencar, S. Messaoud, S. Chawla. (2022). "Impact of Adversarial Training on Robustness and Generalizability of Language Models." Findings of ACL.

A3T: Accuracy-Aware Adversarial Training

Published in Machine Learning (Springer), 2023

A3T improves the generalization/robustness tradeoff in adversarial training by leveraging misclassification accuracy. Selected among the top 3 papers of Machine Learning (2023).

Recommended citation: E. Altinisik, S. Messaoud, H.T. Sencar, S. Chawla. (2023). "A3T: Accuracy-Aware Adversarial Training." Machine Learning (top 3 papers).

S2AC: Energy-Based Reinforcement Learning with Stein Soft Actor Critic

Published in International Conference on Learning Representations (ICLR), 2024

We propose a new variational distribution leveraging Stein Variational Gradient Descent dynamics, enabling learning of multi-modal policies in the context of Max-Entropy Reinforcement Learning.

Recommended citation: S. Messaoud, B. Mokeddem*, Z. Xue*, L. Pang, B. An, H. Chen, S. Chawla. (2024). "S2AC: Energy-Based Reinforcement Learning with Stein Soft Actor Critic." ICLR.

Fanar: An Arabic-Centric Multimodal Generative AI Platform

Published in arXiv preprint arXiv:2501.13944, 2025

Fanar is an Arabic-centric multimodal generative AI platform developed at QCRI, covering language, speech, and vision capabilities tailored to Arabic-speaking users.

Recommended citation: Fanar Team. (2025). "Fanar: An Arabic-Centric Multimodal Generative AI Platform." arXiv preprint arXiv:2501.13944.
Download Paper

Explaining the role of Intrinsic Dimensionality in Adversarial Training

Published in International Conference on Machine Learning (ICML), 2025

We study how the intrinsic dimensionality of the data manifold governs the robustness–generalization tradeoff in adversarially trained models.

Recommended citation: E. Altinisik, S. Messaoud, T. Sencar, H. Sajjad, S. Chawla. (2025). "Explaining the role of Intrinsic Dimensionality in Adversarial Training." ICML.

Particles Don’t Care About Z: Towards Scaling Entropy Estimation of Unnormalized Densities

Published in International Conference on Machine Learning (ICML), 2026

Scaling entropy estimation for unnormalized densities through particle-based methods inspired by Stein Variational Gradient Descent — the normalization constant Z is irrelevant.

Recommended citation: S. Messaoud, S. Charni, E. Bouazza, A. Pourghasemi, H. Bensmail. (2026). "Particles Don't Care About Z: Towards Scaling Entropy Estimation of Unnormalized Densities." ICML.
Download Paper

talks

Generative AI

Published:

A two-hour lecture on generative models, covering likelihood-based and energy-based approaches and their applications.

Diffusion Models

Published:

A one-hour tutorial on diffusion models — covering score-based generative modeling, denoising diffusion probabilistic models, and applications in vision and beyond.

teaching