I'm Saeed Razavi a final-year undergraduate student majoring in electrical engineering at Sharif University of Technology. My research interests span the intersection of ML and healthcare problems, with a focus on self-supervised pipelines, trustworthy AI, and optimization. My ultimate goal is to use ML to create a reliable system for early disease diagnosis. I'm also passionate about developing robust models to tackle manipulation methods in deepfake scenarios.
ECCV 2024
Accepted
we introduce an innovative Sparse Transformer architecture and theoretically prove its universal approximability, featuring a new upper bound for the layer count. We additionally evaluate our method on both pathology and MIL datasets, showcasing its superiority on image and patch-level accuracies compared to the previous methods.
Apr 2024 - Present
Prof. Nassir Navab
We present a novel Concept-aware Contrastive Prototypical approach to utilize the predefined concepts in an image as a prior. To this end, we propose three levels of concepts based on the granularity of the available data. Additionally, we propose a concept contrastive prototypical objective function to better align the extracted features of different concepts within each class and across other classes
Jan 2023 - Nov 2023
Prof. M. H. Rohban
The Image retrieval problem is being worked on with a specific emphasis on feature extraction using self-supervised methods. The main concentration lies in developing and implementing novel pretext tasks and meaningful augmentations tailored to the unique challenges posed by pathological images. In this project, I reviewed the efficiency of different SSL models on pathological images using both ViT and CNN as backbones.
Duration: Summer 2022
Prof. S. Amini
Conducting extensive research on the design of the training process for the deepfake detector resulted in a robust model capable of distinguishing inconsistencies between image patches. We have implemented the Word2vec concept, which was utilized for the NLP task for measuring the correlation between image patches.
Dec 2021 - Dec 2022
Prof. S. A. Motahari
The main task involved the development of a speaker verification system, in which the authentication is performed by the user’s voice. For this purpose, the state-of-the-art AutoSpeech model was trained and evaluated on the Common Voice Persian dataset.
Browse My Recent
Get in Touch
