Mohammad Saadati

Mohammad Saadati

Computer Engineering B.Sc. at the University of Tehran

Research

Transfer Learning in Reinforcement Learning

My Bachelor’s Thesis, University of Tehran, ECE Department, 2024

In my bachelor’s thesis, I explored the application of Transfer Learning to improve the efficiency of Reinforcement Learning, with a particular focus on utilizing Large Language Models. Reinforcement Learning often faces challenges, such as extended learning times and difficulty adapting to new problems. My research aimed to address these challenges by using Transfer Learning to facilitate the transfer of knowledge from past experiences to new tasks. By leveraging Large Language Models, I investigated how identifying similarities and differences between previous knowledge and new problems can enhance learning efficiency. The experiments demonstrated that Large Language Models can significantly accelerate learning and improve decision-making, presenting a more efficient approach than traditional Reinforcement Learning techniques.

Educational Technology

Cognitive Systems Lab, University of Tehran, ECE Department, 2023

Driven by a passion for transforming education using cutting-edge technology, we’ve spearheaded research into Artificial Intelligence, particularly Reinforcement Learning, to create tailored adaptive learning systems for students. Complementing this, we harness the potential of Large Language Models (LLMs), employing techniques such as fine-tuning and Reinforcement Learning from Human Feedback (RLHF), to produce bespoke educational content. Further enhancing our endeavor, we’ve introduced a novel pedagogy model for Deep Learning and designed a state-of-the-art Learning Management System (LMS). This LMS, developed with Python, Django, Bootstrap, and JavaScript, integrates an optimized ChatGPT with Google Classroom, leveraging AI tools like Discord and AI-based robots to offer a rich, personalized learning landscape. Our combined efforts aim for an engaging, accessible, and truly inclusive educational future, details of which can be found on our website.

Hate Speech Detection

Intelligent Information Systems Lab, University of Tehran, ECE Department, 2022

In our research, we aimed to detect hate speech in texts using a predefined dataset. We jointly trained the GCN (Graph Convolutional Networks) and BERT to model this data. Our model formed a heterogeneous graph with documents represented as nodes using BERT, allowing mutual interactions between local and global information. This interaction created a comprehensive classification representation. The multilayer graph contained connections between words, tweets, and hashtags, with inter-layer edges denoting semantic equivalence across languages. We also incorporated predefined words and tweets to enhance the graph’s relationships. Throughout this process, we reviewed numerous articles on GCN types, their synergy with BERT, and heterogeneous graphs. After extensive experimentation with different structures and state-of-the-art models, our research is now concluded.

Crystalline

Data Analytics Lab, University of Tehran, ECE Department, 2021

At the University of Tehran, my team and I developed a cryptocurrency as a proof of concept using pure Python, introducing a unique Proof of Activity as its main consensus protocol. This innovative protocol was crafted to amalgamate the advantages of both the PoS and PoA protocols. Our approach was informed by extensive research, including numerous papers and courses, and the study has been successfully concluded. For more information, please visit our repo.