Adam Ibrahim
About
Welcome to my personal website ! I am a PhD student at Mila – Université de Montréal (well, currently graduating !). I am supervised by Ioannis Mitliagkas and Irina Rish, and work with many wonderful collaborators on a wide variety of topics, such as optimisation, out-of-distribution generalisation and adversarial robustness, reinforcement learning, continual learning, etc. I am part of an effort led by Irina Rish to work on large language models (LLMs) as part of the 2023 INCITE allocation program by the US Department of Energy, which has allowed me to gain experience training LLMs from scratch. Generally, the common theme of my research during my PhD has been addressing distributional shifts, which is important for efficient, fair, safe and trustworthy machine learning algorithms.
I have also worked with Apple, Microsoft, AMD, Staples, Optina Diagnostics and Bosch as a consultant to supervise and mentor research projects in multimodal machine learning, speech recognition, reinforcement learning, image generation, computer vision, large language models and foundation models for scene understanding. Additionally, I have consulted for Blackbox AI, a coding assistant company, and most recently, have been consulting for a stealth startup on pretraining and finetuning foundation models (mixture-of-experts, LLMs, and image generation).
Before my PhD, I have studied human-computer interaction in the Four Eyes lab at the University of California, Santa Barbara. Working with a multidisciplinary team, I led a research project that highlighted how beneficial Augmented Reality (AR) can be for language learning. I also worked on projects aiming to measure how AR affects the perception of user recommendations. Before that, during my BSc, I was enrolled in the Joint Honours Mathematics and Physics program at McGill University, where I did research in theoretical physics (cosmology and string theory) with Robert Brandenberger and Keshav Dasgupta.
As you might have guessed, I really enjoy all kinds of problem solving, and am always happy to tackle problems in new areas of research.
Recent work
- Simple and Scalable Strategies to Continually Pre-train Large Language Models Adam Ibrahim*, Benjamin Thérien*, Kshitij Gupta*, Mats L. Richter, Quentin Anthony, Timothée Lesort, Eugene Belilovsky, Irina Rish. Under review, TMLR.
* denotes equal contribution - Continual Pre-Training of Large Language Models: How to re-warm your model? Kshitij Gupta*, Benjamin Thérien*, Adam Ibrahim*, Mats Leon Richter, Quentin Gregory Anthony, Eugene Belilovsky, Timothée Lesort, Irina Rish. Workshop on Efficient Systems for Foundation Models, ICML 2023.
* denotes equal contribution - Towards Out-of-Distribution Adversarial Robustness. Adam Ibrahim, Charles Guille-Escuret, Ioannis Mitliagkas, Irina Rish, David Krueger, Pouya Bashivan. arXiv preprint – New Frontiers in Adversarial Machine Learning Workshop, ICML 2023.
- Learning Robust Kernel Ensembles with Kernel Average Pooling. Pouya Bashivan, Adam Ibrahim, Amirozhan Dehghani, Yifei Ren. arXiv preprint.
- Gradient Descent Is Optimal Under Lower Restricted Secant Inequality And Upper Error Bound. Charles Guille-Escuret, Adam Ibrahim, Baptiste Goujaud, Ioannis Mitliagkas. NeurIPS 2022.