Hossein Goli

Certified Adversarial Robustness via Partition-based Randomized Smoothing

In collaboration with Prof. Farzan Farnia, We proposed a new method for the Certified Robustness Problem that leverages the spatial correlation of image pixels using SuperPixel Algorithms from Computer Vision literature, such as SLIC SuperPixels. Our method greatly improves the visibility of the image and reduces the effective noise by a factor proportional to the square-root of the average super pixel size. The paper is currently available on arXiv and we plan to complete some of our proofs and submit it soon.

Off Policy Evalution using Generative Models

Under the Supervision of Prof. Florian Shkurti, We proposed a novel method for Off-Policy Evaluation (OPE) using generative models as a substitute for dynamic models. This approach focuses on Energy-Based Models (EBMs), particularly Diffusion Models, to generate realistic trajectories. I am currently testing and evaluating the method in high-dimensional environments, such as Mujoco.

Detecting Viral Social Events with Deep Survival Analysis in Complex Networks

In this Project We Developed a framework for early detection of viral events in a complex network using statistical methods such as survival analysis. To learn a good survival function we proposed using RNN's on the time series data. We evaluted our work on multiple datasets including Twitter, Weibo and Digg.

Computer Vision Course Project

Developed a graph-based image segmentation algorithm using superpixels to group similar pixels by spatial and color features. After constructing the graph, Disjoint Set Union (DSU) iteratively merges vertices until final segmentation. Leveraging SLIC superpixels, a k-means variant, this method enhances robustness without requiring adversarial training. In the second part of the project, we implemented Tiny-NeRF, a simplified version of the famous NeRF paper.

Image Captioning Using LSTM

In this mini project, I used an LSTM-based encoder-decoder architecture for image captioning on the Flickr8k dataset. Keras was used as the main framework for designing the architecture, while PyTorch’s "bert-base-uncased" model was utilized to measure the similarity between true and predicted captions.

System Analysis Design Project - Kafka from Scratch

In this project, we designed a fully dockerized, distributed message queue system from scratch, similar to Kafka. The project involved building and configuring a fault-tolerant, scalable messaging system using Docker containers, allowing message passing across distributed nodes.

Advanced Programming Project - YuGiOh Game

In this project, we designed the famous trading card game, YuGiOh, from scratch, utilizing various design patterns such as Singleton, Factory, Observer, and more. The programming language used was Java. The project was divided into three phases: implementing the core game logic, developing the graphical user interface, and adding network capabilities to allow peer-to-peer gameplay.

Teaching Assistant at Sharif University of Technology

• Machine Learning (Graduate), Spring 2024, Prof. Fatemeh SeyedSalehi — Head Teaching Assistant
• Machine Learning (Graduate), Fall 2023, Prof. Abolfazl Motahari — Theoretical Head Teaching Assistant
• Probability and Statistics, Spring 2023, Prof. Ali Sharifi Zarchi — Head of Midterm Exam
• Linear Algebra, Fall 2022, Prof. Hamid R. Rabiee — Recitation Class and Homeworks
• Advanced Programming, Spring 2022, Prof. Amin Fazli — Project Design Team

Physics Olympiad Teacher

Prepared students to participate in the Physics Olympiad, teaching multiple subjects such as Multivariate Calculus, Electromagnetics, and Newtonian Mechanics.

Rasta Summer School - Game Theory Mentor

We organized and led a workshop for high school students, introducing fundamental concepts of game theory in an engaging manner. The workshop was divided into three sections: Market Sharing, Braess’s Paradox, and the Development of Trust. My primary responsibility was the Development of Trust Section.

Scientific Staff of Sharif AI-Challenge

As a scientific staff member, I was involved with the design and implementation of the scientific part of the challenge.