Mehrshad Mirmohammadi

I am a last-year Computer Engineering undergraduate at Sharif University of Technology, Iran’s top-ranked institution according to QS ranking 2024. Back in high school, I won a silver medal at the International Olympiad in Informatics. In my undergraduate years, I engaged in a variety of deep-learning research projects, some of which resulted in publications at top conferences. I am interested in continuing research in this area, particularly in ML applications in Computer Vision, enhancing and leveraging LLMs’ reasoning capabilities, and applying AI and LLMs in domains with structured languages, such as programming code and mathematical proofs.

Research directions

Generally, I am interested in ways that AI can be used to solve real-world problems. I am particularly interested in the following topics:

Computer Vision: This is the area that I have the most experience in. I truly love how I can put my creativity and problem-solving capabilities in use in this area. Also there are many important problems in this area that are still unsolved.
Reasoning in AI: I am interested in how we can enhance and leverage LLMs’ reasoning capabilities. It is truelly amazing what capabilities emerge from learning from the natural language. Specially those that enable LLMs to solve complex problems. I think on the most important capabilities of large language models is their ability to reason. And there is still a lot of room for improvement in this area. More importantly this ability can be leveraged in many other areas and solve many bottlenecks like halucination.
AI in Code: There are many benefits from using the more structured languages for LLM like programming code and mathematical proofs. I am interested in how we can apply AI and LLMs in these domains, and also how we can exploit the structure of these languages to improve the performance of LLMs.

Research experiance

Since I began university, AI and its applications have fascinated me. Eager to explore, I initiated self-study and hands-on projects. In my first year, I eagerly absorbed most of the ML-related content typically covered in undergraduate courses, unable to wait for formal classes. However, recognizing that university coursework was not deep enough and couldn’t satisfy my hunger for knowledge, I turned to academia. Over the past two years, I’ve dedicated myself to in-depth Deep Learning research:

My first project was a a few‑shot system to detect and verify provided shapes in documents using YOLO for object detection and a combination of Siamese Networks, Homography, RANSAC, and K Nearest Neighbors for object verification.
While in VITA lab at EPFL, I got the chance to work on the task of Human Pose Forecasting. My main contributions to this project were showing our approach’s theoretical aspects and empirically proving its effectiveness. Additionally, our team developed an open-source library with several models, datasets, and metrics. This position allowed me to improve my communication and teamwork skills while gaining extensive research experience. A paper resulting from this project is currently under review at IEEE RA-L 2023.
In the AIT lab at ETH Zürich, I collaborated with five other researchers on the task of Human Object Interaction. In this project, I designed and implemented the core method. This opportunity expanded my applied deep learning research experience and my collaboration, team management, and planning skills. This work is published at ICCV 2023’s R6D Workshop!
During my time at INSAIT, I worked on AI Robustness by combining Randomized Smoothing with Contrastive Learning. I was able to achieved state‑of‑the‑art results by designing different Contrastive Loss Functions for Randomized Smoothing and using ensembling methods. This work is being published and submitted to ICML 2024.

Olympiad in Informatics

Thanks to the Informatics Olympiad, high school was a turning point for me. During this time, I developed competitive programming skills and learned to solve complex problems and think critically. I reached the Grandmaster level at CodeForces (a renowned platform for competitive programming) and the semi-finals at the Google Code Jam competitions. Later, at the International Olympiad of Informatics (IOI) 2019, I finished 36th out of 327 competitors, earning a silver medal. Also, during my undergraduate studies, I participated twice in the ACM-ICPC competitions, winning gold medals at regional contests.

Extra-curricular Activities

teaching
Scientific Committee
Charity Works

news

Aug 19, 2023	My Paper Accepted at ICCV 2023's R6D Workshop!
Jul 1, 2023	I have started a Research Internship @ INSAIT

selected publications

Toward Reliable Human Pose Forecasting with Uncertainty

Saeed Saadatnejad, Mehrshad Mirmohammadi, Matin Daghyani, Parham Saremi, and 5 more authors

arXiv preprint arXiv:2304.06707, 2023

Abs arXiv

Recently, there has been an arms race of pose forecasting methods aimed at solving the spatio-temporal task of predicting a sequence of future 3D poses of a person given a sequence of past observed ones. However, the lack of unified benchmarks and limited uncertainty analysis have hindered progress in the field. To address this, we first develop an open-source library for human pose forecasting, featuring multiple models, datasets, and standardized evaluation metrics, with the aim of promoting research and moving toward a unified and fair evaluation. Second, we devise two types of uncertainty in the problem to increase performance and convey better trust: 1) we propose a method for modeling aleatoric uncertainty by using uncertainty priors to inject knowledge about the behavior of uncertainty. This focuses the capacity of the model in the direction of more meaningful supervision while reducing the number of learned parameters and improving stability; 2) we introduce a novel approach for quantifying the epistemic uncertainty of any model through clustering and measuring the entropy of its assignments. Our experiments demonstrate up to 25% improvements in accuracy and better performance in uncertainty estimation.
Reconstruction of 3D Interaction Models from Images using Shape Prior

Mehrshad Mirmohammadi, Parham Saremi, Yen-Ling Kuo, and Xi Wang

ICCV R6D Workshop, 2023

Abs

We investigate the reconstruction of 3D human-object interactions from images, encompassing 3D human shape and pose estimation as well as object shape and pose estimation. To address this task, we introduce an autoregressive transformer-based variational autoencoder capable of learning a robust shape prior from extensive 3D shape datasets. Additionally, we leverage the reconstructed 3D human body as supplementary features for object shape and pose estimation. In contrast, prior methods only predict object pose and rely on shape templates for shape prediction. Experimental findings on the BEHAVE dataset underscore the effectiveness of our proposed approach, achieving a 40.7cm Chamfer distance and demonstrating the advantages of learning a shape prior.