Abdulla Jasem Almansoori
My name is Abdulla Jasem Almansoori (عبدالله جاسم المنصوري). I started my PhD in machine learning at Mohamed bin Zayed University of Artificial Intelligence in 2021, where I am fortunate to be advised by Martin Takáč and Samuel Horváth. Previously, I got my master degree in computer science at the University of Southern California in 2018, and before that I got my bachelor in industrial engineering at Purdue University in 2016. You can find my cv here.
Research
My interests lie at the intersection of machine learning and optimization. I am generally interested in:
- collaborative learning (e.g., personalization in federated learning).
- deep learning theory and optimization (e.g., parameterization, adaptation).
- generative models (GANs, LLMs).
My long-term vision is to build collaborative AI systems that learn efficiently together and coordinate to solve complex real-world problems.
News
- 2025 Sep: “Faster Than SVD, Smarter Than SGD: The OPLoRA Alternating Update” got accepted to OPT 2025 workshop.
- 2025 Aug: Non-trivially happy that I’ll be a research intern at RIKEN-AIP in the heart of Tokyo! I’ll be hosted by Emtiyaz Khan in his Adaptive Bayesian Intelligence team.
- 2025 May: Significantly excited to be joining Meta in the Bay Area during the summer as a ML software engineering intern! (I came back quickly but the bay turned out to be really nice.)
- 2025 Feb: “Collaborative and Efficient Personalization with Mixtures of Adaptors” got accepted to CPAL 2025. (I’m going to visit Stanford!)
- 2025 Jan: Started a collaboration with Sheikh Shakhbout Medical City (SSMC), advised by Dr. Siddiq Anwar, to explore federated AI applications in healthcare.
- 2024 Nov: Presented a poster at The ADIA Lab Symposium 2024. (Nobel prize winners and Turing laureates speaking!)
- 2024 July: Really proud to have received the Best Reviewer Award at ICML 2024. (Fun fact: I’ve never been to ICML.)
- 2024 May: “PaDPaF: Partial Disentanglement with Partially-Federated GANs” got accepted to TMLR.
- 2023 Dec: Presented a poster in the 2nd collaborative learning workshop at MBZUAI on federated personalization with mixtures of LoRAs. (Great speakers list!)
- 2023 Sep: “Byzantine-Tolerant Methods for Distributed Variational Inequalities” got accepted to NeurIPS 2023. (I’ll be going to present the poster.)
- 2023 May: “PaDPaF: Partial Disentanglement with Partially-Federated GANs” got accepted to MLSys 2023 Workshop on Federated Learning Systems (oral).
Publications
See my Google Scholar for an up to date list.
Preprint
- Abdulla Jasem Almansoori, Maria Ivanova, Andrey Veprikov, Aleksandr Beznosikov, Samuel Horváth, and Martin Takáč. “Beyond SGD, Without SVD: Proximal Subspace Iteration LoRA with Diagonal Fractional K-FAC”. Arxiv. [paper] [code]
Conference
- Abdulla Jasem Almansoori, Samuel Horváth, and Martin Takáč. “Collaborative and Efficient Personalization with Mixtures of Adaptors”. CPAL 2025. [paper] [code]
- Nazarii Tupitsa, Abdulla Jasem Almansoori, Yanlin Wu, Martin Takáč, Karthik Nandakumar, Samuel Horváth, Eduard Gorbunov. “Byzantine-Tolerant Methods for Distributed Variational Inequalities”. NeurIPS 2023. [paper] [code]
Journal
- Abdulla Jasem Almansoori, Samuel Horváth, and Martin Takáč. “PaDPaF: Partial Disentanglement with Partially-Federated GANs”. Transactions on Machine Learning Research (TMLR) 05/2024. [paper] [code]
- Abdurakhmon Sadiev, Aleksandr Beznosikov, Abdulla Jasem Almansoori, Dmitry Kamzolov, Rachael Tappenden, and Martin Takáč. “Stochastic Gradient Methods with Preconditioned Updates”. Journal of Optimization Theory and Applications 03/2024. [paper] [code]
Workshop
-
Abdulla Jasem Almansoori, Maria Ivanova, Andrey Veprikov, Aleksandr Beznosikov, Samuel Horváth, and Martin Takáč. “Faster Than SVD, Smarter Than SGD: The OPLoRA Alternating Update”. OPT 2025. [paper] [code]
-
Abdulla Jasem Almansoori, Samuel Horváth, and Martin Takáč. “PaDPaF: Partial Disentanglement with Partially-Federated GANs” (Extended for TMLR). FLSys 2023 (oral). [paper] [code]
Professional Experience
-
Aug 2025 - Feb 2026: Working as a research intern in the Adaptive Bayesian Intelligence team at RIKEN-AIP in Tokyo, mainly with Thomas Möellenhoff and Emtiyaz Khan. I worked initially on modeling and designing random learning rates methods as a Bayesian Learning Rule method with a Normal-Wishart variational distribution. Shifted towards accelerated methods under Gaussians (Variational Newton Methods), then changed it to “schedule-free”. Paper coming soon. Overall, I enjoyed working with bright and passionate colleagues, had many interesting conversations, learned more Japanse, explored Tokyo, and made great friends. [show more]
-
May 2025 - Aug 2025: Joined the Ads org at Meta in Sunnyvale as an intern. During this internship, I understood many things about how machine learning is done and deployed in practice (hint: data matters). I’m proud that I was able to write code that my team found useful and will build on for future projects, and I was able to train models that will be used in production. Gaining a hands-on ML engineering experience can be very valuable for ML researchers. The work culture also gave me a lot of insight on what I’d call “frictionless productivity”. [show more]
-
Feb 2021 - Jun 2022: Started national service 6 months before my PhD, and finished after my first year. I finished military boot camp, and then got recruited by Dubai Police as a part of the Expo 2020 batch. I received airport security training in Dubai International Airport (DXB), worked at Expo 2020 as a vehicle inspector (from dusty trucks to shiny rolls royces), and then worked as a security guard again in Al Maktoum International Airport (DWC). It had been a great, diverse, mind-opening, humbling experience overall with a slightly hectic shift schedule and long commutes balanced with PhD studies on the side. [show more]
-
Jan 2020 - Feb 2021: Worked as a research assistant in the Center of Genomics and Systems Biology at New York University - Abu Dhabi. I was there for about a year before my PhD, where I was sponsored by the Kawader research assistantship. I was part of the core bioinformatics team, working on next-generation sequencing pipelines, mostly submitting slurm jobs of AI-based bioinformatics tools. The team was absolutely great and the work environment was chill (we used to work from home before it was a thing). [show more]
Hobbies
I do enjoy my research, but I also have a lot of other interests. I love cooking (especially Italian cuisine) and travelling (especially for food). Other hobbies include: reading, writing, coding (used to be fun before LLMs), video games (would love to have more time for this), learning languages (Japanese and Italian, slowly but surely), and opening parentheses (like this one). Recently, I’ve been reading some history and politics, which I find quite illuminating (they say history repeats itself for a reason).
Contact Me
You can contact me at abdu.[middle name]@gmail.com.
Fun fact regarding my middle name
It is customary to use the naming format [first name + father’s name (+ desired number of ancestors) + last name] in the arab world because many people share the same first + last name (e.g. imagine being called John Smith). We use the shortest length that can sufficiently distinguish us in the given setting. In the US, I would use the standard two-name format (first + last) as it is slightly unlikely to see another person sharing the same name abroad at the same institution, for example (though it is still possible). In research, I would use three names (first + father + last) because there could be another researcher in marine biology called Abdulla Almansoori, for example. During my national service, however, I had to use four or even five names! You would be surprised at how many times people sharing the same 3-name have been confused for each other (either first + father + grandfather or first + father + last), especially when the assembly list is sorted alphabetically. In fact, I once saw an Emirati guy on Instagram who shares my first + father + grandfather + last name! [show more]